Commit
·
32e89fe
1
Parent(s):
4bd749c
Initial commit
Browse files- .github/FUNDING.yml +3 -0
- .gitignore +11 -0
- LICENSE +201 -0
- README.md +4 -4
- app.py +389 -0
- examples/1.jpg +0 -0
- examples/2.jpg +0 -0
- examples/3.jpg +0 -0
- examples/4.jpg +0 -0
- examples/5.jpg +0 -0
- pipeline/mod_controlnet_tile_sr_sdxl.py +1845 -0
- pipeline/util.py +328 -0
- requirements.txt +11 -0
.github/FUNDING.yml
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# These are supported funding model platforms
|
| 2 |
+
|
| 3 |
+
ko_fi: elismasilva
|
.gitignore
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
__pycache__/
|
| 2 |
+
*.py[cod]
|
| 3 |
+
mod_tests/
|
| 4 |
+
/.vs
|
| 5 |
+
.vscode/
|
| 6 |
+
.idea/
|
| 7 |
+
venv/
|
| 8 |
+
.venv/
|
| 9 |
+
*.log
|
| 10 |
+
.DS_Store
|
| 11 |
+
.gradio
|
LICENSE
ADDED
|
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Apache License
|
| 2 |
+
Version 2.0, January 2004
|
| 3 |
+
http://www.apache.org/licenses/
|
| 4 |
+
|
| 5 |
+
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
| 6 |
+
|
| 7 |
+
1. Definitions.
|
| 8 |
+
|
| 9 |
+
"License" shall mean the terms and conditions for use, reproduction,
|
| 10 |
+
and distribution as defined by Sections 1 through 9 of this document.
|
| 11 |
+
|
| 12 |
+
"Licensor" shall mean the copyright owner or entity authorized by
|
| 13 |
+
the copyright owner that is granting the License.
|
| 14 |
+
|
| 15 |
+
"Legal Entity" shall mean the union of the acting entity and all
|
| 16 |
+
other entities that control, are controlled by, or are under common
|
| 17 |
+
control with that entity. For the purposes of this definition,
|
| 18 |
+
"control" means (i) the power, direct or indirect, to cause the
|
| 19 |
+
direction or management of such entity, whether by contract or
|
| 20 |
+
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
| 21 |
+
outstanding shares, or (iii) beneficial ownership of such entity.
|
| 22 |
+
|
| 23 |
+
"You" (or "Your") shall mean an individual or Legal Entity
|
| 24 |
+
exercising permissions granted by this License.
|
| 25 |
+
|
| 26 |
+
"Source" form shall mean the preferred form for making modifications,
|
| 27 |
+
including but not limited to software source code, documentation
|
| 28 |
+
source, and configuration files.
|
| 29 |
+
|
| 30 |
+
"Object" form shall mean any form resulting from mechanical
|
| 31 |
+
transformation or translation of a Source form, including but
|
| 32 |
+
not limited to compiled object code, generated documentation,
|
| 33 |
+
and conversions to other media types.
|
| 34 |
+
|
| 35 |
+
"Work" shall mean the work of authorship, whether in Source or
|
| 36 |
+
Object form, made available under the License, as indicated by a
|
| 37 |
+
copyright notice that is included in or attached to the work
|
| 38 |
+
(an example is provided in the Appendix below).
|
| 39 |
+
|
| 40 |
+
"Derivative Works" shall mean any work, whether in Source or Object
|
| 41 |
+
form, that is based on (or derived from) the Work and for which the
|
| 42 |
+
editorial revisions, annotations, elaborations, or other modifications
|
| 43 |
+
represent, as a whole, an original work of authorship. For the purposes
|
| 44 |
+
of this License, Derivative Works shall not include works that remain
|
| 45 |
+
separable from, or merely link (or bind by name) to the interfaces of,
|
| 46 |
+
the Work and Derivative Works thereof.
|
| 47 |
+
|
| 48 |
+
"Contribution" shall mean any work of authorship, including
|
| 49 |
+
the original version of the Work and any modifications or additions
|
| 50 |
+
to that Work or Derivative Works thereof, that is intentionally
|
| 51 |
+
submitted to Licensor for inclusion in the Work by the copyright owner
|
| 52 |
+
or by an individual or Legal Entity authorized to submit on behalf of
|
| 53 |
+
the copyright owner. For the purposes of this definition, "submitted"
|
| 54 |
+
means any form of electronic, verbal, or written communication sent
|
| 55 |
+
to the Licensor or its representatives, including but not limited to
|
| 56 |
+
communication on electronic mailing lists, source code control systems,
|
| 57 |
+
and issue tracking systems that are managed by, or on behalf of, the
|
| 58 |
+
Licensor for the purpose of discussing and improving the Work, but
|
| 59 |
+
excluding communication that is conspicuously marked or otherwise
|
| 60 |
+
designated in writing by the copyright owner as "Not a Contribution."
|
| 61 |
+
|
| 62 |
+
"Contributor" shall mean Licensor and any individual or Legal Entity
|
| 63 |
+
on behalf of whom a Contribution has been received by Licensor and
|
| 64 |
+
subsequently incorporated within the Work.
|
| 65 |
+
|
| 66 |
+
2. Grant of Copyright License. Subject to the terms and conditions of
|
| 67 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 68 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 69 |
+
copyright license to reproduce, prepare Derivative Works of,
|
| 70 |
+
publicly display, publicly perform, sublicense, and distribute the
|
| 71 |
+
Work and such Derivative Works in Source or Object form.
|
| 72 |
+
|
| 73 |
+
3. Grant of Patent License. Subject to the terms and conditions of
|
| 74 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 75 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 76 |
+
(except as stated in this section) patent license to make, have made,
|
| 77 |
+
use, offer to sell, sell, import, and otherwise transfer the Work,
|
| 78 |
+
where such license applies only to those patent claims licensable
|
| 79 |
+
by such Contributor that are necessarily infringed by their
|
| 80 |
+
Contribution(s) alone or by combination of their Contribution(s)
|
| 81 |
+
with the Work to which such Contribution(s) was submitted. If You
|
| 82 |
+
institute patent litigation against any entity (including a
|
| 83 |
+
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
| 84 |
+
or a Contribution incorporated within the Work constitutes direct
|
| 85 |
+
or contributory patent infringement, then any patent licenses
|
| 86 |
+
granted to You under this License for that Work shall terminate
|
| 87 |
+
as of the date such litigation is filed.
|
| 88 |
+
|
| 89 |
+
4. Redistribution. You may reproduce and distribute copies of the
|
| 90 |
+
Work or Derivative Works thereof in any medium, with or without
|
| 91 |
+
modifications, and in Source or Object form, provided that You
|
| 92 |
+
meet the following conditions:
|
| 93 |
+
|
| 94 |
+
(a) You must give any other recipients of the Work or
|
| 95 |
+
Derivative Works a copy of this License; and
|
| 96 |
+
|
| 97 |
+
(b) You must cause any modified files to carry prominent notices
|
| 98 |
+
stating that You changed the files; and
|
| 99 |
+
|
| 100 |
+
(c) You must retain, in the Source form of any Derivative Works
|
| 101 |
+
that You distribute, all copyright, patent, trademark, and
|
| 102 |
+
attribution notices from the Source form of the Work,
|
| 103 |
+
excluding those notices that do not pertain to any part of
|
| 104 |
+
the Derivative Works; and
|
| 105 |
+
|
| 106 |
+
(d) If the Work includes a "NOTICE" text file as part of its
|
| 107 |
+
distribution, then any Derivative Works that You distribute must
|
| 108 |
+
include a readable copy of the attribution notices contained
|
| 109 |
+
within such NOTICE file, excluding those notices that do not
|
| 110 |
+
pertain to any part of the Derivative Works, in at least one
|
| 111 |
+
of the following places: within a NOTICE text file distributed
|
| 112 |
+
as part of the Derivative Works; within the Source form or
|
| 113 |
+
documentation, if provided along with the Derivative Works; or,
|
| 114 |
+
within a display generated by the Derivative Works, if and
|
| 115 |
+
wherever such third-party notices normally appear. The contents
|
| 116 |
+
of the NOTICE file are for informational purposes only and
|
| 117 |
+
do not modify the License. You may add Your own attribution
|
| 118 |
+
notices within Derivative Works that You distribute, alongside
|
| 119 |
+
or as an addendum to the NOTICE text from the Work, provided
|
| 120 |
+
that such additional attribution notices cannot be construed
|
| 121 |
+
as modifying the License.
|
| 122 |
+
|
| 123 |
+
You may add Your own copyright statement to Your modifications and
|
| 124 |
+
may provide additional or different license terms and conditions
|
| 125 |
+
for use, reproduction, or distribution of Your modifications, or
|
| 126 |
+
for any such Derivative Works as a whole, provided Your use,
|
| 127 |
+
reproduction, and distribution of the Work otherwise complies with
|
| 128 |
+
the conditions stated in this License.
|
| 129 |
+
|
| 130 |
+
5. Submission of Contributions. Unless You explicitly state otherwise,
|
| 131 |
+
any Contribution intentionally submitted for inclusion in the Work
|
| 132 |
+
by You to the Licensor shall be under the terms and conditions of
|
| 133 |
+
this License, without any additional terms or conditions.
|
| 134 |
+
Notwithstanding the above, nothing herein shall supersede or modify
|
| 135 |
+
the terms of any separate license agreement you may have executed
|
| 136 |
+
with Licensor regarding such Contributions.
|
| 137 |
+
|
| 138 |
+
6. Trademarks. This License does not grant permission to use the trade
|
| 139 |
+
names, trademarks, service marks, or product names of the Licensor,
|
| 140 |
+
except as required for reasonable and customary use in describing the
|
| 141 |
+
origin of the Work and reproducing the content of the NOTICE file.
|
| 142 |
+
|
| 143 |
+
7. Disclaimer of Warranty. Unless required by applicable law or
|
| 144 |
+
agreed to in writing, Licensor provides the Work (and each
|
| 145 |
+
Contributor provides its Contributions) on an "AS IS" BASIS,
|
| 146 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
| 147 |
+
implied, including, without limitation, any warranties or conditions
|
| 148 |
+
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
| 149 |
+
PARTICULAR PURPOSE. You are solely responsible for determining the
|
| 150 |
+
appropriateness of using or redistributing the Work and assume any
|
| 151 |
+
risks associated with Your exercise of permissions under this License.
|
| 152 |
+
|
| 153 |
+
8. Limitation of Liability. In no event and under no legal theory,
|
| 154 |
+
whether in tort (including negligence), contract, or otherwise,
|
| 155 |
+
unless required by applicable law (such as deliberate and grossly
|
| 156 |
+
negligent acts) or agreed to in writing, shall any Contributor be
|
| 157 |
+
liable to You for damages, including any direct, indirect, special,
|
| 158 |
+
incidental, or consequential damages of any character arising as a
|
| 159 |
+
result of this License or out of the use or inability to use the
|
| 160 |
+
Work (including but not limited to damages for loss of goodwill,
|
| 161 |
+
work stoppage, computer failure or malfunction, or any and all
|
| 162 |
+
other commercial damages or losses), even if such Contributor
|
| 163 |
+
has been advised of the possibility of such damages.
|
| 164 |
+
|
| 165 |
+
9. Accepting Warranty or Additional Liability. While redistributing
|
| 166 |
+
the Work or Derivative Works thereof, You may choose to offer,
|
| 167 |
+
and charge a fee for, acceptance of support, warranty, indemnity,
|
| 168 |
+
or other liability obligations and/or rights consistent with this
|
| 169 |
+
License. However, in accepting such obligations, You may act only
|
| 170 |
+
on Your own behalf and on Your sole responsibility, not on behalf
|
| 171 |
+
of any other Contributor, and only if You agree to indemnify,
|
| 172 |
+
defend, and hold each Contributor harmless for any liability
|
| 173 |
+
incurred by, or claims asserted against, such Contributor by reason
|
| 174 |
+
of your accepting any such warranty or additional liability.
|
| 175 |
+
|
| 176 |
+
END OF TERMS AND CONDITIONS
|
| 177 |
+
|
| 178 |
+
APPENDIX: How to apply the Apache License to your work.
|
| 179 |
+
|
| 180 |
+
To apply the Apache License to your work, attach the following
|
| 181 |
+
boilerplate notice, with the fields enclosed by brackets "[]"
|
| 182 |
+
replaced with your own identifying information. (Don't include
|
| 183 |
+
the brackets!) The text should be enclosed in the appropriate
|
| 184 |
+
comment syntax for the file format. We also recommend that a
|
| 185 |
+
file or class name and description of purpose be included on the
|
| 186 |
+
same "printed page" as the copyright notice for easier
|
| 187 |
+
identification within third-party archives.
|
| 188 |
+
|
| 189 |
+
Copyright [yyyy] [name of copyright owner]
|
| 190 |
+
|
| 191 |
+
Licensed under the Apache License, Version 2.0 (the "License");
|
| 192 |
+
you may not use this file except in compliance with the License.
|
| 193 |
+
You may obtain a copy of the License at
|
| 194 |
+
|
| 195 |
+
http://www.apache.org/licenses/LICENSE-2.0
|
| 196 |
+
|
| 197 |
+
Unless required by applicable law or agreed to in writing, software
|
| 198 |
+
distributed under the License is distributed on an "AS IS" BASIS,
|
| 199 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 200 |
+
See the License for the specific language governing permissions and
|
| 201 |
+
limitations under the License.
|
README.md
CHANGED
|
@@ -1,14 +1,14 @@
|
|
| 1 |
---
|
| 2 |
-
title: Mod
|
| 3 |
-
emoji:
|
| 4 |
colorFrom: purple
|
| 5 |
colorTo: yellow
|
| 6 |
sdk: gradio
|
| 7 |
-
sdk_version: 5.
|
| 8 |
app_file: app.py
|
| 9 |
pinned: false
|
| 10 |
license: apache-2.0
|
| 11 |
-
short_description:
|
| 12 |
---
|
| 13 |
|
| 14 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Mod ControlNet Tile Upscaler SDXL
|
| 3 |
+
emoji: 🚀
|
| 4 |
colorFrom: purple
|
| 5 |
colorTo: yellow
|
| 6 |
sdk: gradio
|
| 7 |
+
sdk_version: 5.15.0
|
| 8 |
app_file: app.py
|
| 9 |
pinned: false
|
| 10 |
license: apache-2.0
|
| 11 |
+
short_description: Mixture of Diffusers and ControlNet Tile Upscaler for SDXL
|
| 12 |
---
|
| 13 |
|
| 14 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
app.py
ADDED
|
@@ -0,0 +1,389 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import spaces
|
| 3 |
+
from diffusers import ControlNetUnionModel, AutoencoderKL, UNet2DConditionModel
|
| 4 |
+
import gradio as gr
|
| 5 |
+
|
| 6 |
+
from pipeline.mod_controlnet_tile_sr_sdxl import StableDiffusionXLControlNetTileSRPipeline
|
| 7 |
+
from pipeline.util import (
|
| 8 |
+
SAMPLERS,
|
| 9 |
+
Platinum,
|
| 10 |
+
calculate_overlap,
|
| 11 |
+
create_hdr_effect,
|
| 12 |
+
progressive_upscale,
|
| 13 |
+
quantize_8bit,
|
| 14 |
+
select_scheduler,
|
| 15 |
+
)
|
| 16 |
+
|
| 17 |
+
device = "cuda"
|
| 18 |
+
|
| 19 |
+
# Initialize the models and pipeline
|
| 20 |
+
controlnet = ControlNetUnionModel.from_pretrained(
|
| 21 |
+
"brad-twinkl/controlnet-union-sdxl-1.0-promax", torch_dtype=torch.float16
|
| 22 |
+
).to(device=device)
|
| 23 |
+
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to(device=device)
|
| 24 |
+
|
| 25 |
+
model_id = "SG161222/RealVisXL_V5.0"
|
| 26 |
+
pipe = StableDiffusionXLControlNetTileSRPipeline.from_pretrained(
|
| 27 |
+
model_id, controlnet=controlnet, vae=vae, torch_dtype=torch.float16, use_safetensors=True, variant="fp16"
|
| 28 |
+
).to(device)
|
| 29 |
+
|
| 30 |
+
unet = UNet2DConditionModel.from_pretrained(model_id, subfolder="unet", variant="fp16", use_safetensors=True)
|
| 31 |
+
quantize_8bit(unet) # << Enable this if you have limited VRAM
|
| 32 |
+
pipe.unet = unet
|
| 33 |
+
|
| 34 |
+
pipe.enable_model_cpu_offload() # << Enable this if you have limited VRAM
|
| 35 |
+
pipe.enable_vae_tiling() # << Enable this if you have limited VRAM
|
| 36 |
+
pipe.enable_vae_slicing() # << Enable this if you have limited VRAM
|
| 37 |
+
|
| 38 |
+
# region functions
|
| 39 |
+
@spaces.GPU
|
| 40 |
+
def predict(
|
| 41 |
+
image,
|
| 42 |
+
prompt,
|
| 43 |
+
negative_prompt,
|
| 44 |
+
resolution,
|
| 45 |
+
hdr,
|
| 46 |
+
num_inference_steps,
|
| 47 |
+
denoising_strenght,
|
| 48 |
+
controlnet_strength,
|
| 49 |
+
tile_gaussian_sigma,
|
| 50 |
+
scheduler,
|
| 51 |
+
guidance_scale,
|
| 52 |
+
max_tile_size,
|
| 53 |
+
tile_weighting_method,
|
| 54 |
+
progress=gr.Progress(track_tqdm=True),
|
| 55 |
+
):
|
| 56 |
+
global pipe
|
| 57 |
+
|
| 58 |
+
# Set selected scheduler
|
| 59 |
+
print(f"Using scheduler: {scheduler}...")
|
| 60 |
+
pipe.scheduler = select_scheduler(pipe, scheduler)
|
| 61 |
+
|
| 62 |
+
# Get current image size
|
| 63 |
+
original_height = image.height
|
| 64 |
+
original_width = image.width
|
| 65 |
+
print(f"Current resolution: H:{original_height} x W:{original_width}")
|
| 66 |
+
|
| 67 |
+
# Pre-upscale image for tiling
|
| 68 |
+
control_image = progressive_upscale(image, resolution)
|
| 69 |
+
control_image = create_hdr_effect(control_image, hdr)
|
| 70 |
+
|
| 71 |
+
# Update target height and width
|
| 72 |
+
target_height = control_image.height
|
| 73 |
+
target_width = control_image.width
|
| 74 |
+
print(f"Target resolution: H:{target_height} x W:{target_width}")
|
| 75 |
+
print(f"Applied HDR effect: {True if hdr > 0 else False}")
|
| 76 |
+
|
| 77 |
+
# Calculate overlap size
|
| 78 |
+
normal_tile_overlap, border_tile_overlap = calculate_overlap(target_width, target_height)
|
| 79 |
+
|
| 80 |
+
# Image generation
|
| 81 |
+
print("Diffusion kicking in... almost done, coffee's on you!")
|
| 82 |
+
image = pipe(
|
| 83 |
+
image=control_image,
|
| 84 |
+
control_image=image,
|
| 85 |
+
control_mode=[6],
|
| 86 |
+
controlnet_conditioning_scale=float(controlnet_strength),
|
| 87 |
+
prompt=prompt,
|
| 88 |
+
negative_prompt=negative_prompt,
|
| 89 |
+
normal_tile_overlap=normal_tile_overlap,
|
| 90 |
+
border_tile_overlap=border_tile_overlap,
|
| 91 |
+
height=target_height,
|
| 92 |
+
width=target_width,
|
| 93 |
+
original_size=(original_width, original_height),
|
| 94 |
+
target_size=(target_width, target_height),
|
| 95 |
+
guidance_scale=guidance_scale,
|
| 96 |
+
strength=float(denoising_strenght),
|
| 97 |
+
tile_weighting_method=tile_weighting_method,
|
| 98 |
+
max_tile_size=max_tile_size,
|
| 99 |
+
tile_gaussian_sigma=float(tile_gaussian_sigma),
|
| 100 |
+
num_inference_steps=num_inference_steps,
|
| 101 |
+
)["images"][0]
|
| 102 |
+
image.save("result.png")
|
| 103 |
+
return image
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
def clear_result():
|
| 107 |
+
return gr.update(value=None)
|
| 108 |
+
|
| 109 |
+
def set_maximum_resolution(max_tile_size, current_value):
|
| 110 |
+
max_scale = 8 # <- you can try increase it to 12x, 16x if you wish!
|
| 111 |
+
maximum_value = max_tile_size * max_scale
|
| 112 |
+
if current_value > maximum_value:
|
| 113 |
+
return gr.update(maximum=maximum_value, value=maximum_value)
|
| 114 |
+
return gr.update(maximum=maximum_value)
|
| 115 |
+
|
| 116 |
+
def select_tile_weighting_method(tile_weighting_method):
|
| 117 |
+
return gr.update(visible=True if tile_weighting_method=="Gaussian" else False)
|
| 118 |
+
|
| 119 |
+
# endregion
|
| 120 |
+
|
| 121 |
+
css = """
|
| 122 |
+
body {
|
| 123 |
+
background: linear-gradient(135deg, #667eea, #764ba2);
|
| 124 |
+
font-family: 'Helvetica Neue', Helvetica, Arial, sans-serif;
|
| 125 |
+
color: #333;
|
| 126 |
+
margin: 0;
|
| 127 |
+
padding: 0;
|
| 128 |
+
}
|
| 129 |
+
.gradio-container {
|
| 130 |
+
background: rgba(255, 255, 255, 0.95);
|
| 131 |
+
border-radius: 15px;
|
| 132 |
+
padding: 30px 40px;
|
| 133 |
+
box-shadow: 0 8px 30px rgba(0, 0, 0, 0.3);
|
| 134 |
+
margin: 40px 340px;
|
| 135 |
+
/*max-width: 1200px;*/
|
| 136 |
+
}
|
| 137 |
+
.gradio-container h1 {
|
| 138 |
+
color: #333;
|
| 139 |
+
text-shadow: 1px 1px 2px rgba(0, 0, 0, 0.2);
|
| 140 |
+
}
|
| 141 |
+
.fillable {
|
| 142 |
+
width: 95% !important;
|
| 143 |
+
max-width: unset !important;
|
| 144 |
+
}
|
| 145 |
+
#examples_container {
|
| 146 |
+
margin: auto;
|
| 147 |
+
width: 90%;
|
| 148 |
+
}
|
| 149 |
+
#examples_row {
|
| 150 |
+
justify-content: center;
|
| 151 |
+
}
|
| 152 |
+
#tips_row{
|
| 153 |
+
padding-left: 20px;
|
| 154 |
+
}
|
| 155 |
+
.sidebar {
|
| 156 |
+
background: rgba(255, 255, 255, 0.98);
|
| 157 |
+
border-radius: 10px;
|
| 158 |
+
padding: 10px;
|
| 159 |
+
box-shadow: 0 4px 15px rgba(0, 0, 0, 0.2);
|
| 160 |
+
}
|
| 161 |
+
.sidebar .toggle-button {
|
| 162 |
+
background: linear-gradient(90deg, #7367f0, #9c93f4);
|
| 163 |
+
border: none;
|
| 164 |
+
color: #fff;
|
| 165 |
+
padding: 12px 24px;
|
| 166 |
+
text-transform: uppercase;
|
| 167 |
+
font-weight: bold;
|
| 168 |
+
letter-spacing: 1px;
|
| 169 |
+
border-radius: 5px;
|
| 170 |
+
cursor: pointer;
|
| 171 |
+
transition: transform 0.2s ease-in-out;
|
| 172 |
+
}
|
| 173 |
+
.toggle-button:hover {
|
| 174 |
+
transform: scale(1.05);
|
| 175 |
+
}
|
| 176 |
+
"""
|
| 177 |
+
title = """<h1 align="center">MoD ControlNet Tile Upscaler for SDXL🤗</h1>
|
| 178 |
+
<div style="display: flex; flex-direction: column; justify-content: center; align-items: center; text-align: center; overflow:hidden;">
|
| 179 |
+
<span>This project implements the <a href="https://arxiv.org/pdf/2408.06072">📜 MoD (Mixture-of-Diffusers)</a> tiled diffusion technique and combines it with SDXL's ControlNet Tile process.</span>
|
| 180 |
+
<span>💻 <b><a href="https://github.com/DEVAIEXP/mod-control-tile-upscaler-sdxl">GitHub Code</a></b>
|
| 181 |
+
<span>🚀 <b>Controlnet Union Power!</b> Check out the model: <a href="https://huggingface.co/xinsir/controlnet-union-sdxl-1.0">Controlnet Union</a></span>
|
| 182 |
+
<span>🎨 <b>RealVisXL V5.0 for Stunning Visuals!</b> Explore it here: <a href="https://huggingface.co/SG161222/RealVisXL_V5.0">RealVisXL</a></span>
|
| 183 |
+
</div>
|
| 184 |
+
"""
|
| 185 |
+
|
| 186 |
+
tips = """
|
| 187 |
+
### Method
|
| 188 |
+
This project proposes an enhanced image upscaling method that leverages ControlNet Tile and Mixture-of-Diffusers techniques, integrating tile diffusion directly into the denoising process within the latent space.
|
| 189 |
+
|
| 190 |
+
Let's compare our method with conventional ControlNet Tile upscaling:
|
| 191 |
+
|
| 192 |
+
**Conventional ControlNet Tile:**
|
| 193 |
+
* Processes tiles in pixel space, potentially leading to edge artifacts during fusion.
|
| 194 |
+
* Processes each tile sequentially, increasing overall execution time (e.g., 16 tiles x 3 min = 48 min).
|
| 195 |
+
* Pixel space fusion using masks (e.g., Gaussian) can result in visible seams.
|
| 196 |
+
* Fixed or adaptively sized tiles and overlap can vary, causing inconsistencies.
|
| 197 |
+
|
| 198 |
+
**Proposed Method (MoD ControlNet Tile Upscaler):**
|
| 199 |
+
* Processes tiles in latent space, enabling smoother fusion and mitigating edge artifacts.
|
| 200 |
+
* Processes all tiles in parallel during denoising, drastically reducing execution time.
|
| 201 |
+
* Latent space fusion with dynamically calculated weights ensures seamless transitions between tiles.
|
| 202 |
+
* Tile size and overlap are dynamically adjusted based on the upscaling scale. For scales below 4x, fixed overlap maintains consistency.
|
| 203 |
+
|
| 204 |
+
"""
|
| 205 |
+
|
| 206 |
+
about = """
|
| 207 |
+
📧 **Contact**
|
| 208 |
+
<br>
|
| 209 |
+
If you have any questions or suggestions, feel free to send your question to <b>contact@devaiexp.com</b>.
|
| 210 |
+
"""
|
| 211 |
+
|
| 212 |
+
with gr.Blocks(css=css, theme=Platinum(), title="MoD ControlNet Tile Upscaler") as app:
|
| 213 |
+
gr.Markdown(title)
|
| 214 |
+
with gr.Row():
|
| 215 |
+
with gr.Column(scale=3):
|
| 216 |
+
with gr.Row():
|
| 217 |
+
with gr.Column():
|
| 218 |
+
input_image = gr.Image(type="pil", label="Input Image",sources=["upload"], height=500)
|
| 219 |
+
with gr.Column():
|
| 220 |
+
result = gr.Image(
|
| 221 |
+
label="Generated Image", show_label=True, format="png", interactive=False, scale=1, height=500, min_width=670
|
| 222 |
+
)
|
| 223 |
+
with gr.Row():
|
| 224 |
+
with gr.Accordion("Input Prompt", open=False):
|
| 225 |
+
with gr.Column():
|
| 226 |
+
prompt = gr.Textbox(
|
| 227 |
+
lines=2,
|
| 228 |
+
label="Prompt",
|
| 229 |
+
placeholder="Default prompt for image",
|
| 230 |
+
value="high-quality, noise-free edges, high quality, 4k, hd, 8k",
|
| 231 |
+
)
|
| 232 |
+
with gr.Column():
|
| 233 |
+
negative_prompt = gr.Textbox(
|
| 234 |
+
lines=2,
|
| 235 |
+
label="Negative Prompt (Optional)",
|
| 236 |
+
placeholder="e.g., blurry, low resolution, artifacts, poor details",
|
| 237 |
+
value="blurry, pixelated, noisy, low resolution, artifacts, poor details",
|
| 238 |
+
)
|
| 239 |
+
with gr.Row():
|
| 240 |
+
generate_button = gr.Button("Generate", variant="primary")
|
| 241 |
+
with gr.Column(scale=1):
|
| 242 |
+
with gr.Row(elem_id="tips_row"):
|
| 243 |
+
gr.Markdown(tips)
|
| 244 |
+
with gr.Sidebar(label="Parameters", open=True):
|
| 245 |
+
with gr.Row(elem_id="parameters_row"):
|
| 246 |
+
gr.Markdown("### General parameters")
|
| 247 |
+
tile_weighting_method = gr.Dropdown(
|
| 248 |
+
label="Tile Weighting Meethod", choices=["Cosine", "Gaussian"], value="Cosine"
|
| 249 |
+
)
|
| 250 |
+
tile_gaussian_sigma = gr.Slider(label="Gaussian Sigma", minimum=0.05, maximum=1.0, step=0.01, value=0.3, visible=False)
|
| 251 |
+
max_tile_size = gr.Dropdown(label="Max. Tile Size", choices=[1024, 1280], value=1024)
|
| 252 |
+
resolution = gr.Slider(minimum=128, maximum=8192, value=2048, step=128, label="Resolution")
|
| 253 |
+
num_inference_steps = gr.Slider(minimum=2, maximum=100, value=30, step=1, label="Inference Steps")
|
| 254 |
+
guidance_scale = gr.Slider(minimum=1, maximum=20, value=6, step=0.1, label="Guidance Scale")
|
| 255 |
+
denoising_strength = gr.Slider(minimum=0.1, maximum=1, value=0.6, step=0.01, label="Denoising Strength")
|
| 256 |
+
controlnet_strength = gr.Slider(
|
| 257 |
+
minimum=0.1, maximum=2.0, value=1.0, step=0.05, label="ControlNet Strength"
|
| 258 |
+
)
|
| 259 |
+
hdr = gr.Slider(minimum=0, maximum=1, value=0, step=0.1, label="HDR Effect")
|
| 260 |
+
with gr.Row():
|
| 261 |
+
scheduler = gr.Dropdown(
|
| 262 |
+
label="Sampler",
|
| 263 |
+
choices=list(SAMPLERS.keys()),
|
| 264 |
+
value="UniPC",
|
| 265 |
+
)
|
| 266 |
+
with gr.Accordion(label="Example Images", open=True):
|
| 267 |
+
with gr.Row(elem_id="examples_row"):
|
| 268 |
+
with gr.Column(scale=12, elem_id="examples_container"):
|
| 269 |
+
gr.Examples(
|
| 270 |
+
examples=[
|
| 271 |
+
[ "./examples/1.jpg",
|
| 272 |
+
prompt.value,
|
| 273 |
+
negative_prompt.value,
|
| 274 |
+
4096,
|
| 275 |
+
0.0,
|
| 276 |
+
35,
|
| 277 |
+
0.65,
|
| 278 |
+
1.0,
|
| 279 |
+
0.3,
|
| 280 |
+
"UniPC",
|
| 281 |
+
4,
|
| 282 |
+
1024,
|
| 283 |
+
"Cosine"
|
| 284 |
+
],
|
| 285 |
+
[ "./examples/2.jpg",
|
| 286 |
+
prompt.value,
|
| 287 |
+
negative_prompt.value,
|
| 288 |
+
4096,
|
| 289 |
+
0.5,
|
| 290 |
+
35,
|
| 291 |
+
0.65,
|
| 292 |
+
1.0,
|
| 293 |
+
0.3,
|
| 294 |
+
"UniPC",
|
| 295 |
+
4,
|
| 296 |
+
1024,
|
| 297 |
+
"Cosine"
|
| 298 |
+
],
|
| 299 |
+
[ "./examples/3.jpg",
|
| 300 |
+
prompt.value,
|
| 301 |
+
negative_prompt.value,
|
| 302 |
+
5120,
|
| 303 |
+
0.5,
|
| 304 |
+
50,
|
| 305 |
+
0.65,
|
| 306 |
+
1.0,
|
| 307 |
+
0.3,
|
| 308 |
+
"UniPC",
|
| 309 |
+
4,
|
| 310 |
+
1280,
|
| 311 |
+
"Gaussian"
|
| 312 |
+
],
|
| 313 |
+
[ "./examples/4.jpg",
|
| 314 |
+
prompt.value,
|
| 315 |
+
negative_prompt.value,
|
| 316 |
+
8192,
|
| 317 |
+
0.1,
|
| 318 |
+
50,
|
| 319 |
+
0.5,
|
| 320 |
+
1.0,
|
| 321 |
+
0.3,
|
| 322 |
+
"UniPC",
|
| 323 |
+
4,
|
| 324 |
+
1024,
|
| 325 |
+
"Gaussian"
|
| 326 |
+
],
|
| 327 |
+
[ "./examples/5.jpg",
|
| 328 |
+
prompt.value,
|
| 329 |
+
negative_prompt.value,
|
| 330 |
+
8192,
|
| 331 |
+
0.3,
|
| 332 |
+
50,
|
| 333 |
+
0.5,
|
| 334 |
+
1.0,
|
| 335 |
+
0.3,
|
| 336 |
+
"UniPC",
|
| 337 |
+
4,
|
| 338 |
+
1024,
|
| 339 |
+
"Cosine"
|
| 340 |
+
],
|
| 341 |
+
],
|
| 342 |
+
inputs=[
|
| 343 |
+
input_image,
|
| 344 |
+
prompt,
|
| 345 |
+
negative_prompt,
|
| 346 |
+
resolution,
|
| 347 |
+
hdr,
|
| 348 |
+
num_inference_steps,
|
| 349 |
+
denoising_strength,
|
| 350 |
+
controlnet_strength,
|
| 351 |
+
tile_gaussian_sigma,
|
| 352 |
+
scheduler,
|
| 353 |
+
guidance_scale,
|
| 354 |
+
max_tile_size,
|
| 355 |
+
tile_weighting_method,
|
| 356 |
+
],
|
| 357 |
+
fn=predict,
|
| 358 |
+
outputs=result,
|
| 359 |
+
cache_examples=False,
|
| 360 |
+
)
|
| 361 |
+
|
| 362 |
+
max_tile_size.select(fn=set_maximum_resolution, inputs=[max_tile_size, resolution], outputs=resolution)
|
| 363 |
+
tile_weighting_method.select(fn=select_tile_weighting_method, inputs=tile_weighting_method, outputs=tile_gaussian_sigma)
|
| 364 |
+
generate_button.click(
|
| 365 |
+
fn=clear_result,
|
| 366 |
+
inputs=None,
|
| 367 |
+
outputs=result,
|
| 368 |
+
).then(
|
| 369 |
+
fn=predict,
|
| 370 |
+
inputs=[
|
| 371 |
+
input_image,
|
| 372 |
+
prompt,
|
| 373 |
+
negative_prompt,
|
| 374 |
+
resolution,
|
| 375 |
+
hdr,
|
| 376 |
+
num_inference_steps,
|
| 377 |
+
denoising_strength,
|
| 378 |
+
controlnet_strength,
|
| 379 |
+
tile_gaussian_sigma,
|
| 380 |
+
scheduler,
|
| 381 |
+
guidance_scale,
|
| 382 |
+
max_tile_size,
|
| 383 |
+
tile_weighting_method,
|
| 384 |
+
],
|
| 385 |
+
outputs=result,
|
| 386 |
+
show_progress="full"
|
| 387 |
+
)
|
| 388 |
+
gr.Markdown(about)
|
| 389 |
+
app.launch(share=False)
|
examples/1.jpg
ADDED
|
examples/2.jpg
ADDED
|
examples/3.jpg
ADDED
|
examples/4.jpg
ADDED
|
examples/5.jpg
ADDED
|
pipeline/mod_controlnet_tile_sr_sdxl.py
ADDED
|
@@ -0,0 +1,1845 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2025 DEVAIEXP and The HuggingFace Team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
|
| 15 |
+
import inspect
|
| 16 |
+
from enum import Enum
|
| 17 |
+
from typing import Any, Dict, List, Optional, Tuple, Union
|
| 18 |
+
|
| 19 |
+
import numpy as np
|
| 20 |
+
import torch
|
| 21 |
+
import torch.nn.functional as F
|
| 22 |
+
from PIL import Image
|
| 23 |
+
from transformers import (
|
| 24 |
+
CLIPTextModel,
|
| 25 |
+
CLIPTextModelWithProjection,
|
| 26 |
+
CLIPTokenizer,
|
| 27 |
+
)
|
| 28 |
+
|
| 29 |
+
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
|
| 30 |
+
from diffusers.loaders import (
|
| 31 |
+
FromSingleFileMixin,
|
| 32 |
+
StableDiffusionXLLoraLoaderMixin,
|
| 33 |
+
TextualInversionLoaderMixin,
|
| 34 |
+
)
|
| 35 |
+
from diffusers.models import (
|
| 36 |
+
AutoencoderKL,
|
| 37 |
+
ControlNetModel,
|
| 38 |
+
ControlNetUnionModel,
|
| 39 |
+
MultiControlNetModel,
|
| 40 |
+
UNet2DConditionModel,
|
| 41 |
+
)
|
| 42 |
+
from diffusers.models.attention_processor import (
|
| 43 |
+
AttnProcessor2_0,
|
| 44 |
+
XFormersAttnProcessor,
|
| 45 |
+
)
|
| 46 |
+
from diffusers.models.lora import adjust_lora_scale_text_encoder
|
| 47 |
+
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
|
| 48 |
+
from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
|
| 49 |
+
from diffusers.schedulers import KarrasDiffusionSchedulers, LMSDiscreteScheduler
|
| 50 |
+
from diffusers.utils import (
|
| 51 |
+
USE_PEFT_BACKEND,
|
| 52 |
+
logging,
|
| 53 |
+
replace_example_docstring,
|
| 54 |
+
scale_lora_layers,
|
| 55 |
+
unscale_lora_layers,
|
| 56 |
+
)
|
| 57 |
+
from diffusers.utils.import_utils import is_invisible_watermark_available
|
| 58 |
+
from diffusers.utils.torch_utils import is_compiled_module, randn_tensor
|
| 59 |
+
|
| 60 |
+
if is_invisible_watermark_available():
|
| 61 |
+
from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
|
| 62 |
+
|
| 63 |
+
from diffusers.utils import is_torch_xla_available
|
| 64 |
+
|
| 65 |
+
if is_torch_xla_available():
|
| 66 |
+
import torch_xla.core.xla_model as xm
|
| 67 |
+
|
| 68 |
+
XLA_AVAILABLE = True
|
| 69 |
+
else:
|
| 70 |
+
XLA_AVAILABLE = False
|
| 71 |
+
|
| 72 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
EXAMPLE_DOC_STRING = """
|
| 76 |
+
Examples:
|
| 77 |
+
```py
|
| 78 |
+
# !pip install controlnet_aux
|
| 79 |
+
from diffusers import (
|
| 80 |
+
StableDiffusionXLControlNetUnionImg2ImgPipeline,
|
| 81 |
+
ControlNetUnionModel,
|
| 82 |
+
AutoencoderKL,
|
| 83 |
+
)
|
| 84 |
+
from diffusers.utils import load_image
|
| 85 |
+
import torch
|
| 86 |
+
from PIL import Image
|
| 87 |
+
import numpy as np
|
| 88 |
+
|
| 89 |
+
prompt = "A cat"
|
| 90 |
+
# download an image
|
| 91 |
+
image = load_image(
|
| 92 |
+
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/cat.png"
|
| 93 |
+
)
|
| 94 |
+
# initialize the models and pipeline
|
| 95 |
+
controlnet = ControlNetUnionModel.from_pretrained(
|
| 96 |
+
"brad-twinkl/controlnet-union-sdxl-1.0-promax", torch_dtype=torch.float16
|
| 97 |
+
)
|
| 98 |
+
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
|
| 99 |
+
pipe = StableDiffusionXLControlNetUnionImg2ImgPipeline.from_pretrained(
|
| 100 |
+
"stabilityai/stable-diffusion-xl-base-1.0",
|
| 101 |
+
controlnet=controlnet,
|
| 102 |
+
vae=vae,
|
| 103 |
+
torch_dtype=torch.float16,
|
| 104 |
+
variant="fp16",
|
| 105 |
+
).to("cuda")
|
| 106 |
+
# `enable_model_cpu_offload` is not recommended due to multiple generations
|
| 107 |
+
height = image.height
|
| 108 |
+
width = image.width
|
| 109 |
+
ratio = np.sqrt(1024.0 * 1024.0 / (width * height))
|
| 110 |
+
# 3 * 3 upscale correspond to 16 * 3 multiply, 2 * 2 correspond to 16 * 2 multiply and so on.
|
| 111 |
+
scale_image_factor = 3
|
| 112 |
+
base_factor = 16
|
| 113 |
+
factor = scale_image_factor * base_factor
|
| 114 |
+
W, H = int(width * ratio) // factor * factor, int(height * ratio) // factor * factor
|
| 115 |
+
image = image.resize((W, H))
|
| 116 |
+
target_width = W // scale_image_factor
|
| 117 |
+
target_height = H // scale_image_factor
|
| 118 |
+
images = []
|
| 119 |
+
crops_coords_list = [
|
| 120 |
+
(0, 0),
|
| 121 |
+
(0, width // 2),
|
| 122 |
+
(height // 2, 0),
|
| 123 |
+
(width // 2, height // 2),
|
| 124 |
+
0,
|
| 125 |
+
0,
|
| 126 |
+
0,
|
| 127 |
+
0,
|
| 128 |
+
0,
|
| 129 |
+
]
|
| 130 |
+
for i in range(scale_image_factor):
|
| 131 |
+
for j in range(scale_image_factor):
|
| 132 |
+
left = j * target_width
|
| 133 |
+
top = i * target_height
|
| 134 |
+
right = left + target_width
|
| 135 |
+
bottom = top + target_height
|
| 136 |
+
cropped_image = image.crop((left, top, right, bottom))
|
| 137 |
+
cropped_image = cropped_image.resize((W, H))
|
| 138 |
+
images.append(cropped_image)
|
| 139 |
+
# set ControlNetUnion input
|
| 140 |
+
result_images = []
|
| 141 |
+
for sub_img, crops_coords in zip(images, crops_coords_list):
|
| 142 |
+
new_width, new_height = W, H
|
| 143 |
+
out = pipe(
|
| 144 |
+
prompt=[prompt] * 1,
|
| 145 |
+
image=sub_img,
|
| 146 |
+
control_image=[sub_img],
|
| 147 |
+
control_mode=[6],
|
| 148 |
+
width=new_width,
|
| 149 |
+
height=new_height,
|
| 150 |
+
num_inference_steps=30,
|
| 151 |
+
crops_coords_top_left=(W, H),
|
| 152 |
+
target_size=(W, H),
|
| 153 |
+
original_size=(W * 2, H * 2),
|
| 154 |
+
)
|
| 155 |
+
result_images.append(out.images[0])
|
| 156 |
+
new_im = Image.new("RGB", (new_width * scale_image_factor, new_height * scale_image_factor))
|
| 157 |
+
new_im.paste(result_images[0], (0, 0))
|
| 158 |
+
new_im.paste(result_images[1], (new_width, 0))
|
| 159 |
+
new_im.paste(result_images[2], (new_width * 2, 0))
|
| 160 |
+
new_im.paste(result_images[3], (0, new_height))
|
| 161 |
+
new_im.paste(result_images[4], (new_width, new_height))
|
| 162 |
+
new_im.paste(result_images[5], (new_width * 2, new_height))
|
| 163 |
+
new_im.paste(result_images[6], (0, new_height * 2))
|
| 164 |
+
new_im.paste(result_images[7], (new_width, new_height * 2))
|
| 165 |
+
new_im.paste(result_images[8], (new_width * 2, new_height * 2))
|
| 166 |
+
```
|
| 167 |
+
"""
|
| 168 |
+
|
| 169 |
+
|
| 170 |
+
# This function was copied and adapted from https://huggingface.co/spaces/gokaygokay/TileUpscalerV2, licensed under Apache 2.0.
|
| 171 |
+
def _adaptive_tile_size(image_size, base_tile_size=512, max_tile_size=1280):
|
| 172 |
+
"""
|
| 173 |
+
Calculate the adaptive tile size based on the image dimensions, ensuring the tile
|
| 174 |
+
respects the aspect ratio and stays within the specified size limits.
|
| 175 |
+
"""
|
| 176 |
+
width, height = image_size
|
| 177 |
+
aspect_ratio = width / height
|
| 178 |
+
|
| 179 |
+
if aspect_ratio > 1:
|
| 180 |
+
# Landscape orientation
|
| 181 |
+
tile_width = min(width, max_tile_size)
|
| 182 |
+
tile_height = min(int(tile_width / aspect_ratio), max_tile_size)
|
| 183 |
+
else:
|
| 184 |
+
# Portrait or square orientation
|
| 185 |
+
tile_height = min(height, max_tile_size)
|
| 186 |
+
tile_width = min(int(tile_height * aspect_ratio), max_tile_size)
|
| 187 |
+
|
| 188 |
+
# Ensure the tile size is not smaller than the base_tile_size
|
| 189 |
+
tile_width = max(tile_width, base_tile_size)
|
| 190 |
+
tile_height = max(tile_height, base_tile_size)
|
| 191 |
+
|
| 192 |
+
return tile_width, tile_height
|
| 193 |
+
|
| 194 |
+
|
| 195 |
+
# Copied and adapted from https://github.com/huggingface/diffusers/blob/main/examples/community/mixture_tiling.py
|
| 196 |
+
def _tile2pixel_indices(
|
| 197 |
+
tile_row, tile_col, tile_width, tile_height, tile_row_overlap, tile_col_overlap, image_width, image_height
|
| 198 |
+
):
|
| 199 |
+
"""Given a tile row and column numbers returns the range of pixels affected by that tiles in the overall image
|
| 200 |
+
|
| 201 |
+
Returns a tuple with:
|
| 202 |
+
- Starting coordinates of rows in pixel space
|
| 203 |
+
- Ending coordinates of rows in pixel space
|
| 204 |
+
- Starting coordinates of columns in pixel space
|
| 205 |
+
- Ending coordinates of columns in pixel space
|
| 206 |
+
"""
|
| 207 |
+
# Calculate initial indices
|
| 208 |
+
px_row_init = 0 if tile_row == 0 else tile_row * (tile_height - tile_row_overlap)
|
| 209 |
+
px_col_init = 0 if tile_col == 0 else tile_col * (tile_width - tile_col_overlap)
|
| 210 |
+
|
| 211 |
+
# Calculate end indices
|
| 212 |
+
px_row_end = px_row_init + tile_height
|
| 213 |
+
px_col_end = px_col_init + tile_width
|
| 214 |
+
|
| 215 |
+
# Ensure the last tile does not exceed the image dimensions
|
| 216 |
+
px_row_end = min(px_row_end, image_height)
|
| 217 |
+
px_col_end = min(px_col_end, image_width)
|
| 218 |
+
|
| 219 |
+
return px_row_init, px_row_end, px_col_init, px_col_end
|
| 220 |
+
|
| 221 |
+
|
| 222 |
+
# Copied and adapted from https://github.com/huggingface/diffusers/blob/main/examples/community/mixture_tiling.py
|
| 223 |
+
def _tile2latent_indices(
|
| 224 |
+
tile_row, tile_col, tile_width, tile_height, tile_row_overlap, tile_col_overlap, image_width, image_height
|
| 225 |
+
):
|
| 226 |
+
"""Given a tile row and column numbers returns the range of latents affected by that tiles in the overall image
|
| 227 |
+
|
| 228 |
+
Returns a tuple with:
|
| 229 |
+
- Starting coordinates of rows in latent space
|
| 230 |
+
- Ending coordinates of rows in latent space
|
| 231 |
+
- Starting coordinates of columns in latent space
|
| 232 |
+
- Ending coordinates of columns in latent space
|
| 233 |
+
"""
|
| 234 |
+
# Get pixel indices
|
| 235 |
+
px_row_init, px_row_end, px_col_init, px_col_end = _tile2pixel_indices(
|
| 236 |
+
tile_row, tile_col, tile_width, tile_height, tile_row_overlap, tile_col_overlap, image_width, image_height
|
| 237 |
+
)
|
| 238 |
+
|
| 239 |
+
# Convert to latent space
|
| 240 |
+
latent_row_init = px_row_init // 8
|
| 241 |
+
latent_row_end = px_row_end // 8
|
| 242 |
+
latent_col_init = px_col_init // 8
|
| 243 |
+
latent_col_end = px_col_end // 8
|
| 244 |
+
latent_height = image_height // 8
|
| 245 |
+
latent_width = image_width // 8
|
| 246 |
+
|
| 247 |
+
# Ensure the last tile does not exceed the latent dimensions
|
| 248 |
+
latent_row_end = min(latent_row_end, latent_height)
|
| 249 |
+
latent_col_end = min(latent_col_end, latent_width)
|
| 250 |
+
|
| 251 |
+
return latent_row_init, latent_row_end, latent_col_init, latent_col_end
|
| 252 |
+
|
| 253 |
+
|
| 254 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
|
| 255 |
+
def retrieve_latents(
|
| 256 |
+
encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
|
| 257 |
+
):
|
| 258 |
+
if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
|
| 259 |
+
return encoder_output.latent_dist.sample(generator)
|
| 260 |
+
elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
|
| 261 |
+
return encoder_output.latent_dist.mode()
|
| 262 |
+
elif hasattr(encoder_output, "latents"):
|
| 263 |
+
return encoder_output.latents
|
| 264 |
+
else:
|
| 265 |
+
raise AttributeError("Could not access latents of provided encoder_output")
|
| 266 |
+
|
| 267 |
+
class TileWeightingMethod(Enum):
|
| 268 |
+
"""Mode in which the tile weights will be generated"""
|
| 269 |
+
|
| 270 |
+
COSINE = "Cosine"
|
| 271 |
+
GAUSSIAN = "Gaussian"
|
| 272 |
+
|
| 273 |
+
class StableDiffusionXLControlNetTileSRPipeline(
|
| 274 |
+
DiffusionPipeline,
|
| 275 |
+
StableDiffusionMixin,
|
| 276 |
+
TextualInversionLoaderMixin,
|
| 277 |
+
StableDiffusionXLLoraLoaderMixin,
|
| 278 |
+
FromSingleFileMixin,
|
| 279 |
+
):
|
| 280 |
+
r"""
|
| 281 |
+
Pipeline for image-to-image generation using Stable Diffusion XL with ControlNet guidance.
|
| 282 |
+
|
| 283 |
+
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
|
| 284 |
+
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
|
| 285 |
+
|
| 286 |
+
The pipeline also inherits the following loading methods:
|
| 287 |
+
- [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
|
| 288 |
+
- [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
|
| 289 |
+
- [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
|
| 290 |
+
|
| 291 |
+
Args:
|
| 292 |
+
vae ([`AutoencoderKL`]):
|
| 293 |
+
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
|
| 294 |
+
text_encoder ([`CLIPTextModel`]):
|
| 295 |
+
Frozen text-encoder. Stable Diffusion uses the text portion of
|
| 296 |
+
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
|
| 297 |
+
the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
|
| 298 |
+
text_encoder_2 ([` CLIPTextModelWithProjection`]):
|
| 299 |
+
Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
|
| 300 |
+
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
|
| 301 |
+
specifically the
|
| 302 |
+
[laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
|
| 303 |
+
variant.
|
| 304 |
+
tokenizer (`CLIPTokenizer`):
|
| 305 |
+
Tokenizer of class
|
| 306 |
+
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
|
| 307 |
+
tokenizer_2 (`CLIPTokenizer`):
|
| 308 |
+
Second Tokenizer of class
|
| 309 |
+
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
|
| 310 |
+
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
|
| 311 |
+
controlnet ([`ControlNetUnionModel`]):
|
| 312 |
+
Provides additional conditioning to the unet during the denoising process.
|
| 313 |
+
scheduler ([`SchedulerMixin`]):
|
| 314 |
+
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
|
| 315 |
+
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
|
| 316 |
+
requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`):
|
| 317 |
+
Whether the `unet` requires an `aesthetic_score` condition to be passed during inference. Also see the
|
| 318 |
+
config of `stabilityai/stable-diffusion-xl-refiner-1-0`.
|
| 319 |
+
force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
|
| 320 |
+
Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
|
| 321 |
+
`stabilityai/stable-diffusion-xl-base-1-0`.
|
| 322 |
+
add_watermarker (`bool`, *optional*):
|
| 323 |
+
Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
|
| 324 |
+
watermark output images. If not defined, it will default to True if the package is installed, otherwise no
|
| 325 |
+
watermarker will be used.
|
| 326 |
+
"""
|
| 327 |
+
|
| 328 |
+
model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
|
| 329 |
+
_optional_components = [
|
| 330 |
+
"tokenizer",
|
| 331 |
+
"tokenizer_2",
|
| 332 |
+
"text_encoder",
|
| 333 |
+
"text_encoder_2",
|
| 334 |
+
]
|
| 335 |
+
|
| 336 |
+
def __init__(
|
| 337 |
+
self,
|
| 338 |
+
vae: AutoencoderKL,
|
| 339 |
+
text_encoder: CLIPTextModel,
|
| 340 |
+
text_encoder_2: CLIPTextModelWithProjection,
|
| 341 |
+
tokenizer: CLIPTokenizer,
|
| 342 |
+
tokenizer_2: CLIPTokenizer,
|
| 343 |
+
unet: UNet2DConditionModel,
|
| 344 |
+
controlnet: ControlNetUnionModel,
|
| 345 |
+
scheduler: KarrasDiffusionSchedulers,
|
| 346 |
+
requires_aesthetics_score: bool = False,
|
| 347 |
+
force_zeros_for_empty_prompt: bool = True,
|
| 348 |
+
add_watermarker: Optional[bool] = None,
|
| 349 |
+
):
|
| 350 |
+
super().__init__()
|
| 351 |
+
|
| 352 |
+
if not isinstance(controlnet, ControlNetUnionModel):
|
| 353 |
+
raise ValueError("Expected `controlnet` to be of type `ControlNetUnionModel`.")
|
| 354 |
+
|
| 355 |
+
self.register_modules(
|
| 356 |
+
vae=vae,
|
| 357 |
+
text_encoder=text_encoder,
|
| 358 |
+
text_encoder_2=text_encoder_2,
|
| 359 |
+
tokenizer=tokenizer,
|
| 360 |
+
tokenizer_2=tokenizer_2,
|
| 361 |
+
unet=unet,
|
| 362 |
+
controlnet=controlnet,
|
| 363 |
+
scheduler=scheduler,
|
| 364 |
+
)
|
| 365 |
+
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
|
| 366 |
+
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
|
| 367 |
+
self.control_image_processor = VaeImageProcessor(
|
| 368 |
+
vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
|
| 369 |
+
)
|
| 370 |
+
self.mask_processor = VaeImageProcessor(
|
| 371 |
+
vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
|
| 372 |
+
)
|
| 373 |
+
add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
|
| 374 |
+
|
| 375 |
+
if add_watermarker:
|
| 376 |
+
self.watermark = StableDiffusionXLWatermarker()
|
| 377 |
+
else:
|
| 378 |
+
self.watermark = None
|
| 379 |
+
|
| 380 |
+
self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
|
| 381 |
+
self.register_to_config(requires_aesthetics_score=requires_aesthetics_score)
|
| 382 |
+
|
| 383 |
+
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
|
| 384 |
+
def encode_prompt(
|
| 385 |
+
self,
|
| 386 |
+
prompt: str,
|
| 387 |
+
prompt_2: Optional[str] = None,
|
| 388 |
+
device: Optional[torch.device] = None,
|
| 389 |
+
num_images_per_prompt: int = 1,
|
| 390 |
+
do_classifier_free_guidance: bool = True,
|
| 391 |
+
negative_prompt: Optional[str] = None,
|
| 392 |
+
negative_prompt_2: Optional[str] = None,
|
| 393 |
+
prompt_embeds: Optional[torch.Tensor] = None,
|
| 394 |
+
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
| 395 |
+
pooled_prompt_embeds: Optional[torch.Tensor] = None,
|
| 396 |
+
negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
|
| 397 |
+
lora_scale: Optional[float] = None,
|
| 398 |
+
clip_skip: Optional[int] = None,
|
| 399 |
+
):
|
| 400 |
+
r"""
|
| 401 |
+
Encodes the prompt into text encoder hidden states.
|
| 402 |
+
|
| 403 |
+
Args:
|
| 404 |
+
prompt (`str` or `List[str]`, *optional*):
|
| 405 |
+
prompt to be encoded
|
| 406 |
+
prompt_2 (`str` or `List[str]`, *optional*):
|
| 407 |
+
The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
|
| 408 |
+
used in both text-encoders
|
| 409 |
+
device: (`torch.device`):
|
| 410 |
+
torch device
|
| 411 |
+
num_images_per_prompt (`int`):
|
| 412 |
+
number of images that should be generated per prompt
|
| 413 |
+
do_classifier_free_guidance (`bool`):
|
| 414 |
+
whether to use classifier free guidance or not
|
| 415 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
| 416 |
+
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
| 417 |
+
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
| 418 |
+
less than `1`).
|
| 419 |
+
negative_prompt_2 (`str` or `List[str]`, *optional*):
|
| 420 |
+
The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
|
| 421 |
+
`text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
|
| 422 |
+
prompt_embeds (`torch.Tensor`, *optional*):
|
| 423 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
| 424 |
+
provided, text embeddings will be generated from `prompt` input argument.
|
| 425 |
+
negative_prompt_embeds (`torch.Tensor`, *optional*):
|
| 426 |
+
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
| 427 |
+
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
| 428 |
+
argument.
|
| 429 |
+
pooled_prompt_embeds (`torch.Tensor`, *optional*):
|
| 430 |
+
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
|
| 431 |
+
If not provided, pooled text embeddings will be generated from `prompt` input argument.
|
| 432 |
+
negative_pooled_prompt_embeds (`torch.Tensor`, *optional*):
|
| 433 |
+
Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
| 434 |
+
weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
|
| 435 |
+
input argument.
|
| 436 |
+
lora_scale (`float`, *optional*):
|
| 437 |
+
A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
|
| 438 |
+
clip_skip (`int`, *optional*):
|
| 439 |
+
Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
|
| 440 |
+
the output of the pre-final layer will be used for computing the prompt embeddings.
|
| 441 |
+
"""
|
| 442 |
+
device = device or self._execution_device
|
| 443 |
+
|
| 444 |
+
# set lora scale so that monkey patched LoRA
|
| 445 |
+
# function of text encoder can correctly access it
|
| 446 |
+
if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
|
| 447 |
+
self._lora_scale = lora_scale
|
| 448 |
+
|
| 449 |
+
# dynamically adjust the LoRA scale
|
| 450 |
+
if self.text_encoder is not None:
|
| 451 |
+
if not USE_PEFT_BACKEND:
|
| 452 |
+
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
|
| 453 |
+
else:
|
| 454 |
+
scale_lora_layers(self.text_encoder, lora_scale)
|
| 455 |
+
|
| 456 |
+
if self.text_encoder_2 is not None:
|
| 457 |
+
if not USE_PEFT_BACKEND:
|
| 458 |
+
adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
|
| 459 |
+
else:
|
| 460 |
+
scale_lora_layers(self.text_encoder_2, lora_scale)
|
| 461 |
+
|
| 462 |
+
prompt = [prompt] if isinstance(prompt, str) else prompt
|
| 463 |
+
|
| 464 |
+
if prompt is not None:
|
| 465 |
+
batch_size = len(prompt)
|
| 466 |
+
else:
|
| 467 |
+
batch_size = prompt_embeds.shape[0]
|
| 468 |
+
|
| 469 |
+
# Define tokenizers and text encoders
|
| 470 |
+
tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
|
| 471 |
+
text_encoders = (
|
| 472 |
+
[self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
|
| 473 |
+
)
|
| 474 |
+
dtype = text_encoders[0].dtype
|
| 475 |
+
if prompt_embeds is None:
|
| 476 |
+
prompt_2 = prompt_2 or prompt
|
| 477 |
+
prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
|
| 478 |
+
|
| 479 |
+
# textual inversion: process multi-vector tokens if necessary
|
| 480 |
+
prompt_embeds_list = []
|
| 481 |
+
prompts = [prompt, prompt_2]
|
| 482 |
+
for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
|
| 483 |
+
if isinstance(self, TextualInversionLoaderMixin):
|
| 484 |
+
prompt = self.maybe_convert_prompt(prompt, tokenizer)
|
| 485 |
+
|
| 486 |
+
text_inputs = tokenizer(
|
| 487 |
+
prompt,
|
| 488 |
+
padding="max_length",
|
| 489 |
+
max_length=tokenizer.model_max_length,
|
| 490 |
+
truncation=True,
|
| 491 |
+
return_tensors="pt",
|
| 492 |
+
)
|
| 493 |
+
|
| 494 |
+
text_input_ids = text_inputs.input_ids
|
| 495 |
+
untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
|
| 496 |
+
|
| 497 |
+
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
|
| 498 |
+
text_input_ids, untruncated_ids
|
| 499 |
+
):
|
| 500 |
+
removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
|
| 501 |
+
logger.warning(
|
| 502 |
+
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
| 503 |
+
f" {tokenizer.model_max_length} tokens: {removed_text}"
|
| 504 |
+
)
|
| 505 |
+
text_encoder.to(dtype)
|
| 506 |
+
prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
|
| 507 |
+
|
| 508 |
+
# We are only ALWAYS interested in the pooled output of the final text encoder
|
| 509 |
+
if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2:
|
| 510 |
+
pooled_prompt_embeds = prompt_embeds[0]
|
| 511 |
+
|
| 512 |
+
if clip_skip is None:
|
| 513 |
+
prompt_embeds = prompt_embeds.hidden_states[-2]
|
| 514 |
+
else:
|
| 515 |
+
# "2" because SDXL always indexes from the penultimate layer.
|
| 516 |
+
prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
|
| 517 |
+
|
| 518 |
+
prompt_embeds_list.append(prompt_embeds)
|
| 519 |
+
|
| 520 |
+
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
|
| 521 |
+
|
| 522 |
+
# get unconditional embeddings for classifier free guidance
|
| 523 |
+
zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
|
| 524 |
+
if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
|
| 525 |
+
negative_prompt_embeds = torch.zeros_like(prompt_embeds)
|
| 526 |
+
negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
|
| 527 |
+
elif do_classifier_free_guidance and negative_prompt_embeds is None:
|
| 528 |
+
negative_prompt = negative_prompt or ""
|
| 529 |
+
negative_prompt_2 = negative_prompt_2 or negative_prompt
|
| 530 |
+
|
| 531 |
+
# normalize str to list
|
| 532 |
+
negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
|
| 533 |
+
negative_prompt_2 = (
|
| 534 |
+
batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
|
| 535 |
+
)
|
| 536 |
+
|
| 537 |
+
uncond_tokens: List[str]
|
| 538 |
+
if prompt is not None and type(prompt) is not type(negative_prompt):
|
| 539 |
+
raise TypeError(
|
| 540 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
| 541 |
+
f" {type(prompt)}."
|
| 542 |
+
)
|
| 543 |
+
elif batch_size != len(negative_prompt):
|
| 544 |
+
raise ValueError(
|
| 545 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
| 546 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
| 547 |
+
" the batch size of `prompt`."
|
| 548 |
+
)
|
| 549 |
+
else:
|
| 550 |
+
uncond_tokens = [negative_prompt, negative_prompt_2]
|
| 551 |
+
|
| 552 |
+
negative_prompt_embeds_list = []
|
| 553 |
+
for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
|
| 554 |
+
if isinstance(self, TextualInversionLoaderMixin):
|
| 555 |
+
negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
|
| 556 |
+
|
| 557 |
+
max_length = prompt_embeds.shape[1]
|
| 558 |
+
uncond_input = tokenizer(
|
| 559 |
+
negative_prompt,
|
| 560 |
+
padding="max_length",
|
| 561 |
+
max_length=max_length,
|
| 562 |
+
truncation=True,
|
| 563 |
+
return_tensors="pt",
|
| 564 |
+
)
|
| 565 |
+
|
| 566 |
+
negative_prompt_embeds = text_encoder(
|
| 567 |
+
uncond_input.input_ids.to(device),
|
| 568 |
+
output_hidden_states=True,
|
| 569 |
+
)
|
| 570 |
+
|
| 571 |
+
# We are only ALWAYS interested in the pooled output of the final text encoder
|
| 572 |
+
if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2:
|
| 573 |
+
negative_pooled_prompt_embeds = negative_prompt_embeds[0]
|
| 574 |
+
negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
|
| 575 |
+
|
| 576 |
+
negative_prompt_embeds_list.append(negative_prompt_embeds)
|
| 577 |
+
|
| 578 |
+
negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
|
| 579 |
+
|
| 580 |
+
if self.text_encoder_2 is not None:
|
| 581 |
+
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
|
| 582 |
+
else:
|
| 583 |
+
prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
|
| 584 |
+
|
| 585 |
+
bs_embed, seq_len, _ = prompt_embeds.shape
|
| 586 |
+
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
| 587 |
+
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
| 588 |
+
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
| 589 |
+
|
| 590 |
+
if do_classifier_free_guidance:
|
| 591 |
+
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
|
| 592 |
+
seq_len = negative_prompt_embeds.shape[1]
|
| 593 |
+
|
| 594 |
+
if self.text_encoder_2 is not None:
|
| 595 |
+
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
|
| 596 |
+
else:
|
| 597 |
+
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
|
| 598 |
+
|
| 599 |
+
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
| 600 |
+
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
| 601 |
+
|
| 602 |
+
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
|
| 603 |
+
bs_embed * num_images_per_prompt, -1
|
| 604 |
+
)
|
| 605 |
+
if do_classifier_free_guidance:
|
| 606 |
+
negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
|
| 607 |
+
bs_embed * num_images_per_prompt, -1
|
| 608 |
+
)
|
| 609 |
+
|
| 610 |
+
if self.text_encoder is not None:
|
| 611 |
+
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
|
| 612 |
+
# Retrieve the original scale by scaling back the LoRA layers
|
| 613 |
+
unscale_lora_layers(self.text_encoder, lora_scale)
|
| 614 |
+
|
| 615 |
+
if self.text_encoder_2 is not None:
|
| 616 |
+
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
|
| 617 |
+
# Retrieve the original scale by scaling back the LoRA layers
|
| 618 |
+
unscale_lora_layers(self.text_encoder_2, lora_scale)
|
| 619 |
+
|
| 620 |
+
return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
|
| 621 |
+
|
| 622 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
|
| 623 |
+
def prepare_extra_step_kwargs(self, generator, eta):
|
| 624 |
+
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
| 625 |
+
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
| 626 |
+
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
| 627 |
+
# and should be between [0, 1]
|
| 628 |
+
|
| 629 |
+
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
| 630 |
+
extra_step_kwargs = {}
|
| 631 |
+
if accepts_eta:
|
| 632 |
+
extra_step_kwargs["eta"] = eta
|
| 633 |
+
|
| 634 |
+
# check if the scheduler accepts generator
|
| 635 |
+
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
| 636 |
+
if accepts_generator:
|
| 637 |
+
extra_step_kwargs["generator"] = generator
|
| 638 |
+
return extra_step_kwargs
|
| 639 |
+
|
| 640 |
+
def check_inputs(
|
| 641 |
+
self,
|
| 642 |
+
prompt,
|
| 643 |
+
height,
|
| 644 |
+
width,
|
| 645 |
+
image,
|
| 646 |
+
strength,
|
| 647 |
+
num_inference_steps,
|
| 648 |
+
normal_tile_overlap,
|
| 649 |
+
border_tile_overlap,
|
| 650 |
+
max_tile_size,
|
| 651 |
+
tile_gaussian_sigma,
|
| 652 |
+
tile_weighting_method,
|
| 653 |
+
controlnet_conditioning_scale=1.0,
|
| 654 |
+
control_guidance_start=0.0,
|
| 655 |
+
control_guidance_end=1.0,
|
| 656 |
+
):
|
| 657 |
+
if height % 8 != 0 or width % 8 != 0:
|
| 658 |
+
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
|
| 659 |
+
|
| 660 |
+
if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
|
| 661 |
+
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
| 662 |
+
|
| 663 |
+
if strength < 0 or strength > 1:
|
| 664 |
+
raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
|
| 665 |
+
if num_inference_steps is None:
|
| 666 |
+
raise ValueError("`num_inference_steps` cannot be None.")
|
| 667 |
+
elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0:
|
| 668 |
+
raise ValueError(
|
| 669 |
+
f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type"
|
| 670 |
+
f" {type(num_inference_steps)}."
|
| 671 |
+
)
|
| 672 |
+
if normal_tile_overlap is None:
|
| 673 |
+
raise ValueError("`normal_tile_overlap` cannot be None.")
|
| 674 |
+
elif not isinstance(normal_tile_overlap, int) or normal_tile_overlap < 64:
|
| 675 |
+
raise ValueError(
|
| 676 |
+
f"`normal_tile_overlap` has to be greater than 64 but is {normal_tile_overlap} of type"
|
| 677 |
+
f" {type(normal_tile_overlap)}."
|
| 678 |
+
)
|
| 679 |
+
if border_tile_overlap is None:
|
| 680 |
+
raise ValueError("`border_tile_overlap` cannot be None.")
|
| 681 |
+
elif not isinstance(border_tile_overlap, int) or border_tile_overlap < 128:
|
| 682 |
+
raise ValueError(
|
| 683 |
+
f"`border_tile_overlap` has to be greater than 128 but is {border_tile_overlap} of type"
|
| 684 |
+
f" {type(border_tile_overlap)}."
|
| 685 |
+
)
|
| 686 |
+
if max_tile_size is None:
|
| 687 |
+
raise ValueError("`max_tile_size` cannot be None.")
|
| 688 |
+
elif not isinstance(max_tile_size, int) or max_tile_size not in(1024, 1280):
|
| 689 |
+
raise ValueError(
|
| 690 |
+
f"`max_tile_size` has to be in 1024 or 1280 but is {max_tile_size} of type"
|
| 691 |
+
f" {type(max_tile_size)}."
|
| 692 |
+
)
|
| 693 |
+
if tile_gaussian_sigma is None:
|
| 694 |
+
raise ValueError("`tile_gaussian_sigma` cannot be None.")
|
| 695 |
+
elif not isinstance(tile_gaussian_sigma, float) or tile_gaussian_sigma <= 0:
|
| 696 |
+
raise ValueError(
|
| 697 |
+
f"`tile_gaussian_sigma` has to be a positive float but is {tile_gaussian_sigma} of type"
|
| 698 |
+
f" {type(tile_gaussian_sigma)}."
|
| 699 |
+
)
|
| 700 |
+
if tile_weighting_method is None:
|
| 701 |
+
raise ValueError("`tile_weighting_method` cannot be None.")
|
| 702 |
+
elif not isinstance(tile_weighting_method, str) or tile_weighting_method not in [t.value for t in TileWeightingMethod]:
|
| 703 |
+
raise ValueError(
|
| 704 |
+
f"`tile_weighting_method` has to be a string in ({[t.value for t in TileWeightingMethod]}) but is {tile_weighting_method} of type"
|
| 705 |
+
f" {type(tile_weighting_method)}."
|
| 706 |
+
)
|
| 707 |
+
|
| 708 |
+
# Check `image`
|
| 709 |
+
is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
|
| 710 |
+
self.controlnet, torch._dynamo.eval_frame.OptimizedModule
|
| 711 |
+
)
|
| 712 |
+
if (
|
| 713 |
+
isinstance(self.controlnet, ControlNetModel)
|
| 714 |
+
or is_compiled
|
| 715 |
+
and isinstance(self.controlnet._orig_mod, ControlNetModel)
|
| 716 |
+
):
|
| 717 |
+
self.check_image(image, prompt)
|
| 718 |
+
elif (
|
| 719 |
+
isinstance(self.controlnet, ControlNetUnionModel)
|
| 720 |
+
or is_compiled
|
| 721 |
+
and isinstance(self.controlnet._orig_mod, ControlNetUnionModel)
|
| 722 |
+
):
|
| 723 |
+
self.check_image(image, prompt)
|
| 724 |
+
else:
|
| 725 |
+
assert False
|
| 726 |
+
|
| 727 |
+
# Check `controlnet_conditioning_scale`
|
| 728 |
+
if (
|
| 729 |
+
isinstance(self.controlnet, ControlNetUnionModel)
|
| 730 |
+
or is_compiled
|
| 731 |
+
and isinstance(self.controlnet._orig_mod, ControlNetUnionModel)
|
| 732 |
+
) or (
|
| 733 |
+
isinstance(self.controlnet, MultiControlNetModel)
|
| 734 |
+
or is_compiled
|
| 735 |
+
and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
|
| 736 |
+
):
|
| 737 |
+
if not isinstance(controlnet_conditioning_scale, float):
|
| 738 |
+
raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
|
| 739 |
+
elif (
|
| 740 |
+
isinstance(self.controlnet, MultiControlNetModel)
|
| 741 |
+
or is_compiled
|
| 742 |
+
and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
|
| 743 |
+
):
|
| 744 |
+
if isinstance(controlnet_conditioning_scale, list):
|
| 745 |
+
if any(isinstance(i, list) for i in controlnet_conditioning_scale):
|
| 746 |
+
raise ValueError("A single batch of multiple conditionings are supported at the moment.")
|
| 747 |
+
elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
|
| 748 |
+
self.controlnet.nets
|
| 749 |
+
):
|
| 750 |
+
raise ValueError(
|
| 751 |
+
"For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
|
| 752 |
+
" the same length as the number of controlnets"
|
| 753 |
+
)
|
| 754 |
+
else:
|
| 755 |
+
assert False
|
| 756 |
+
|
| 757 |
+
if not isinstance(control_guidance_start, (tuple, list)):
|
| 758 |
+
control_guidance_start = [control_guidance_start]
|
| 759 |
+
|
| 760 |
+
if not isinstance(control_guidance_end, (tuple, list)):
|
| 761 |
+
control_guidance_end = [control_guidance_end]
|
| 762 |
+
|
| 763 |
+
if len(control_guidance_start) != len(control_guidance_end):
|
| 764 |
+
raise ValueError(
|
| 765 |
+
f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
|
| 766 |
+
)
|
| 767 |
+
|
| 768 |
+
for start, end in zip(control_guidance_start, control_guidance_end):
|
| 769 |
+
if start >= end:
|
| 770 |
+
raise ValueError(
|
| 771 |
+
f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
|
| 772 |
+
)
|
| 773 |
+
if start < 0.0:
|
| 774 |
+
raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
|
| 775 |
+
if end > 1.0:
|
| 776 |
+
raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
|
| 777 |
+
|
| 778 |
+
# Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl.StableDiffusionXLControlNetPipeline.check_image
|
| 779 |
+
def check_image(self, image, prompt):
|
| 780 |
+
image_is_pil = isinstance(image, Image.Image)
|
| 781 |
+
image_is_tensor = isinstance(image, torch.Tensor)
|
| 782 |
+
image_is_np = isinstance(image, np.ndarray)
|
| 783 |
+
image_is_pil_list = isinstance(image, list) and isinstance(image[0], Image.Image)
|
| 784 |
+
image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
|
| 785 |
+
image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
|
| 786 |
+
|
| 787 |
+
if (
|
| 788 |
+
not image_is_pil
|
| 789 |
+
and not image_is_tensor
|
| 790 |
+
and not image_is_np
|
| 791 |
+
and not image_is_pil_list
|
| 792 |
+
and not image_is_tensor_list
|
| 793 |
+
and not image_is_np_list
|
| 794 |
+
):
|
| 795 |
+
raise TypeError(
|
| 796 |
+
f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
|
| 797 |
+
)
|
| 798 |
+
|
| 799 |
+
if image_is_pil:
|
| 800 |
+
image_batch_size = 1
|
| 801 |
+
else:
|
| 802 |
+
image_batch_size = len(image)
|
| 803 |
+
|
| 804 |
+
if prompt is not None and isinstance(prompt, str):
|
| 805 |
+
prompt_batch_size = 1
|
| 806 |
+
elif prompt is not None and isinstance(prompt, list):
|
| 807 |
+
prompt_batch_size = len(prompt)
|
| 808 |
+
|
| 809 |
+
if image_batch_size != 1 and image_batch_size != prompt_batch_size:
|
| 810 |
+
raise ValueError(
|
| 811 |
+
f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
|
| 812 |
+
)
|
| 813 |
+
|
| 814 |
+
# Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl.StableDiffusionXLControlNetPipeline.prepare_image
|
| 815 |
+
def prepare_control_image(
|
| 816 |
+
self,
|
| 817 |
+
image,
|
| 818 |
+
width,
|
| 819 |
+
height,
|
| 820 |
+
batch_size,
|
| 821 |
+
num_images_per_prompt,
|
| 822 |
+
device,
|
| 823 |
+
dtype,
|
| 824 |
+
do_classifier_free_guidance=False,
|
| 825 |
+
guess_mode=False,
|
| 826 |
+
):
|
| 827 |
+
image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
|
| 828 |
+
image_batch_size = image.shape[0]
|
| 829 |
+
|
| 830 |
+
if image_batch_size == 1:
|
| 831 |
+
repeat_by = batch_size
|
| 832 |
+
else:
|
| 833 |
+
# image batch size is the same as prompt batch size
|
| 834 |
+
repeat_by = num_images_per_prompt
|
| 835 |
+
|
| 836 |
+
image = image.repeat_interleave(repeat_by, dim=0)
|
| 837 |
+
|
| 838 |
+
image = image.to(device=device, dtype=dtype)
|
| 839 |
+
|
| 840 |
+
if do_classifier_free_guidance and not guess_mode:
|
| 841 |
+
image = torch.cat([image] * 2)
|
| 842 |
+
|
| 843 |
+
return image
|
| 844 |
+
|
| 845 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
|
| 846 |
+
def get_timesteps(self, num_inference_steps, strength):
|
| 847 |
+
# get the original timestep using init_timestep
|
| 848 |
+
init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
|
| 849 |
+
|
| 850 |
+
t_start = max(num_inference_steps - init_timestep, 0)
|
| 851 |
+
timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
|
| 852 |
+
if hasattr(self.scheduler, "set_begin_index"):
|
| 853 |
+
self.scheduler.set_begin_index(t_start * self.scheduler.order)
|
| 854 |
+
|
| 855 |
+
return timesteps, num_inference_steps - t_start
|
| 856 |
+
|
| 857 |
+
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.prepare_latents
|
| 858 |
+
def prepare_latents(
|
| 859 |
+
self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None, add_noise=True
|
| 860 |
+
):
|
| 861 |
+
if not isinstance(image, (torch.Tensor, Image.Image, list)):
|
| 862 |
+
raise ValueError(
|
| 863 |
+
f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
|
| 864 |
+
)
|
| 865 |
+
|
| 866 |
+
latents_mean = latents_std = None
|
| 867 |
+
if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
|
| 868 |
+
latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
|
| 869 |
+
if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
|
| 870 |
+
latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
|
| 871 |
+
|
| 872 |
+
# Offload text encoder if `enable_model_cpu_offload` was enabled
|
| 873 |
+
if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
|
| 874 |
+
self.text_encoder_2.to("cpu")
|
| 875 |
+
torch.cuda.empty_cache()
|
| 876 |
+
|
| 877 |
+
image = image.to(device=device, dtype=dtype)
|
| 878 |
+
|
| 879 |
+
batch_size = batch_size * num_images_per_prompt
|
| 880 |
+
|
| 881 |
+
if image.shape[1] == 4:
|
| 882 |
+
init_latents = image
|
| 883 |
+
|
| 884 |
+
else:
|
| 885 |
+
# make sure the VAE is in float32 mode, as it overflows in float16
|
| 886 |
+
if self.vae.config.force_upcast:
|
| 887 |
+
image = image.float()
|
| 888 |
+
self.vae.to(dtype=torch.float32)
|
| 889 |
+
|
| 890 |
+
if isinstance(generator, list) and len(generator) != batch_size:
|
| 891 |
+
raise ValueError(
|
| 892 |
+
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
|
| 893 |
+
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
|
| 894 |
+
)
|
| 895 |
+
|
| 896 |
+
elif isinstance(generator, list):
|
| 897 |
+
if image.shape[0] < batch_size and batch_size % image.shape[0] == 0:
|
| 898 |
+
image = torch.cat([image] * (batch_size // image.shape[0]), dim=0)
|
| 899 |
+
elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0:
|
| 900 |
+
raise ValueError(
|
| 901 |
+
f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} "
|
| 902 |
+
)
|
| 903 |
+
|
| 904 |
+
init_latents = [
|
| 905 |
+
retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
|
| 906 |
+
for i in range(batch_size)
|
| 907 |
+
]
|
| 908 |
+
init_latents = torch.cat(init_latents, dim=0)
|
| 909 |
+
else:
|
| 910 |
+
init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
|
| 911 |
+
|
| 912 |
+
if self.vae.config.force_upcast:
|
| 913 |
+
self.vae.to(dtype)
|
| 914 |
+
|
| 915 |
+
init_latents = init_latents.to(dtype)
|
| 916 |
+
if latents_mean is not None and latents_std is not None:
|
| 917 |
+
latents_mean = latents_mean.to(device=device, dtype=dtype)
|
| 918 |
+
latents_std = latents_std.to(device=device, dtype=dtype)
|
| 919 |
+
init_latents = (init_latents - latents_mean) * self.vae.config.scaling_factor / latents_std
|
| 920 |
+
else:
|
| 921 |
+
init_latents = self.vae.config.scaling_factor * init_latents
|
| 922 |
+
|
| 923 |
+
if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
|
| 924 |
+
# expand init_latents for batch_size
|
| 925 |
+
additional_image_per_prompt = batch_size // init_latents.shape[0]
|
| 926 |
+
init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
|
| 927 |
+
elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
|
| 928 |
+
raise ValueError(
|
| 929 |
+
f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
|
| 930 |
+
)
|
| 931 |
+
else:
|
| 932 |
+
init_latents = torch.cat([init_latents], dim=0)
|
| 933 |
+
|
| 934 |
+
if add_noise:
|
| 935 |
+
shape = init_latents.shape
|
| 936 |
+
noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
| 937 |
+
# get latents
|
| 938 |
+
init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
|
| 939 |
+
|
| 940 |
+
latents = init_latents
|
| 941 |
+
|
| 942 |
+
return latents
|
| 943 |
+
|
| 944 |
+
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids
|
| 945 |
+
def _get_add_time_ids(
|
| 946 |
+
self,
|
| 947 |
+
original_size,
|
| 948 |
+
crops_coords_top_left,
|
| 949 |
+
target_size,
|
| 950 |
+
aesthetic_score,
|
| 951 |
+
negative_aesthetic_score,
|
| 952 |
+
negative_original_size,
|
| 953 |
+
negative_crops_coords_top_left,
|
| 954 |
+
negative_target_size,
|
| 955 |
+
dtype,
|
| 956 |
+
text_encoder_projection_dim=None,
|
| 957 |
+
):
|
| 958 |
+
if self.config.requires_aesthetics_score:
|
| 959 |
+
add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,))
|
| 960 |
+
add_neg_time_ids = list(
|
| 961 |
+
negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,)
|
| 962 |
+
)
|
| 963 |
+
else:
|
| 964 |
+
add_time_ids = list(original_size + crops_coords_top_left + target_size)
|
| 965 |
+
add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size)
|
| 966 |
+
|
| 967 |
+
passed_add_embed_dim = (
|
| 968 |
+
self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
|
| 969 |
+
)
|
| 970 |
+
expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
|
| 971 |
+
|
| 972 |
+
if (
|
| 973 |
+
expected_add_embed_dim > passed_add_embed_dim
|
| 974 |
+
and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim
|
| 975 |
+
):
|
| 976 |
+
raise ValueError(
|
| 977 |
+
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model."
|
| 978 |
+
)
|
| 979 |
+
elif (
|
| 980 |
+
expected_add_embed_dim < passed_add_embed_dim
|
| 981 |
+
and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim
|
| 982 |
+
):
|
| 983 |
+
raise ValueError(
|
| 984 |
+
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model."
|
| 985 |
+
)
|
| 986 |
+
elif expected_add_embed_dim != passed_add_embed_dim:
|
| 987 |
+
raise ValueError(
|
| 988 |
+
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
|
| 989 |
+
)
|
| 990 |
+
|
| 991 |
+
add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
|
| 992 |
+
add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype)
|
| 993 |
+
|
| 994 |
+
return add_time_ids, add_neg_time_ids
|
| 995 |
+
|
| 996 |
+
def _generate_cosine_weights(self, tile_width, tile_height, nbatches, device, dtype):
|
| 997 |
+
"""
|
| 998 |
+
Generates cosine weights as a PyTorch tensor for blending tiles.
|
| 999 |
+
|
| 1000 |
+
Args:
|
| 1001 |
+
tile_width (int): Width of the tile in pixels.
|
| 1002 |
+
tile_height (int): Height of the tile in pixels.
|
| 1003 |
+
nbatches (int): Number of batches.
|
| 1004 |
+
device (torch.device): Device where the tensor will be allocated (e.g., 'cuda' or 'cpu').
|
| 1005 |
+
dtype (torch.dtype): Data type of the tensor (e.g., torch.float32).
|
| 1006 |
+
|
| 1007 |
+
Returns:
|
| 1008 |
+
torch.Tensor: A tensor containing cosine weights for blending tiles, expanded to match batch and channel dimensions.
|
| 1009 |
+
"""
|
| 1010 |
+
# Convert tile dimensions to latent space
|
| 1011 |
+
latent_width = tile_width // 8
|
| 1012 |
+
latent_height = tile_height // 8
|
| 1013 |
+
|
| 1014 |
+
# Generate x and y coordinates in latent space
|
| 1015 |
+
x = np.arange(0, latent_width)
|
| 1016 |
+
y = np.arange(0, latent_height)
|
| 1017 |
+
|
| 1018 |
+
# Calculate midpoints
|
| 1019 |
+
midpoint_x = (latent_width - 1) / 2
|
| 1020 |
+
midpoint_y = (latent_height - 1) / 2
|
| 1021 |
+
|
| 1022 |
+
# Compute cosine probabilities for x and y
|
| 1023 |
+
x_probs = np.cos(np.pi * (x - midpoint_x) / latent_width)
|
| 1024 |
+
y_probs = np.cos(np.pi * (y - midpoint_y) / latent_height)
|
| 1025 |
+
|
| 1026 |
+
# Create a 2D weight matrix using the outer product
|
| 1027 |
+
weights_np = np.outer(y_probs, x_probs)
|
| 1028 |
+
|
| 1029 |
+
# Convert to a PyTorch tensor with the correct device and dtype
|
| 1030 |
+
weights_torch = torch.tensor(weights_np, device=device, dtype=dtype)
|
| 1031 |
+
|
| 1032 |
+
# Expand for batch and channel dimensions
|
| 1033 |
+
tile_weights_expanded = torch.tile(weights_torch, (nbatches, self.unet.config.in_channels, 1, 1))
|
| 1034 |
+
|
| 1035 |
+
return tile_weights_expanded
|
| 1036 |
+
|
| 1037 |
+
def _generate_gaussian_weights(self, tile_width, tile_height, nbatches, device, dtype, sigma=0.05):
|
| 1038 |
+
"""
|
| 1039 |
+
Generates Gaussian weights as a PyTorch tensor for blending tiles in latent space.
|
| 1040 |
+
|
| 1041 |
+
Args:
|
| 1042 |
+
tile_width (int): Width of the tile in pixels.
|
| 1043 |
+
tile_height (int): Height of the tile in pixels.
|
| 1044 |
+
nbatches (int): Number of batches.
|
| 1045 |
+
device (torch.device): Device where the tensor will be allocated (e.g., 'cuda' or 'cpu').
|
| 1046 |
+
dtype (torch.dtype): Data type of the tensor (e.g., torch.float32).
|
| 1047 |
+
sigma (float, optional): Standard deviation of the Gaussian distribution. Controls the smoothness of the weights. Defaults to 0.05.
|
| 1048 |
+
|
| 1049 |
+
Returns:
|
| 1050 |
+
torch.Tensor: A tensor containing Gaussian weights for blending tiles, expanded to match batch and channel dimensions.
|
| 1051 |
+
"""
|
| 1052 |
+
# Convert tile dimensions to latent space
|
| 1053 |
+
latent_width = tile_width // 8
|
| 1054 |
+
latent_height = tile_height // 8
|
| 1055 |
+
|
| 1056 |
+
# Generate Gaussian weights in latent space
|
| 1057 |
+
x = np.linspace(-1, 1, latent_width)
|
| 1058 |
+
y = np.linspace(-1, 1, latent_height)
|
| 1059 |
+
xx, yy = np.meshgrid(x, y)
|
| 1060 |
+
gaussian_weight = np.exp(-(xx**2 + yy**2) / (2 * sigma**2))
|
| 1061 |
+
|
| 1062 |
+
# Convert to a PyTorch tensor with the correct device and dtype
|
| 1063 |
+
weights_torch = torch.tensor(gaussian_weight, device=device, dtype=dtype)
|
| 1064 |
+
|
| 1065 |
+
# Expand for batch and channel dimensions
|
| 1066 |
+
weights_expanded = weights_torch.unsqueeze(0).unsqueeze(0) # Add batch and channel dimensions
|
| 1067 |
+
weights_expanded = weights_expanded.expand(nbatches, -1, -1, -1) # Expand to the number of batches
|
| 1068 |
+
|
| 1069 |
+
return weights_expanded
|
| 1070 |
+
|
| 1071 |
+
def _get_num_tiles(self, height, width, tile_height, tile_width, normal_tile_overlap, border_tile_overlap):
|
| 1072 |
+
"""
|
| 1073 |
+
Calculates the number of tiles needed to cover an image, choosing the appropriate formula based on the
|
| 1074 |
+
ratio between the image size and the tile size.
|
| 1075 |
+
|
| 1076 |
+
This function automatically selects between two formulas:
|
| 1077 |
+
1. A universal formula for typical cases (image-to-tile ratio <= 6:1).
|
| 1078 |
+
2. A specialized formula with border tile overlap for larger or atypical cases (image-to-tile ratio > 6:1).
|
| 1079 |
+
|
| 1080 |
+
Args:
|
| 1081 |
+
height (int): Height of the image in pixels.
|
| 1082 |
+
width (int): Width of the image in pixels.
|
| 1083 |
+
tile_height (int): Height of each tile in pixels.
|
| 1084 |
+
tile_width (int): Width of each tile in pixels.
|
| 1085 |
+
normal_tile_overlap (int): Overlap between tiles in pixels for normal (non-border) tiles.
|
| 1086 |
+
border_tile_overlap (int): Overlap between tiles in pixels for border tiles.
|
| 1087 |
+
|
| 1088 |
+
Returns:
|
| 1089 |
+
tuple: A tuple containing:
|
| 1090 |
+
- grid_rows (int): Number of rows in the tile grid.
|
| 1091 |
+
- grid_cols (int): Number of columns in the tile grid.
|
| 1092 |
+
|
| 1093 |
+
Notes:
|
| 1094 |
+
- The function uses the universal formula (without border_tile_overlap) for typical cases where the
|
| 1095 |
+
image-to-tile ratio is 6:1 or smaller.
|
| 1096 |
+
- For larger or atypical cases (image-to-tile ratio > 6:1), it uses a specialized formula that includes
|
| 1097 |
+
border_tile_overlap to ensure complete coverage of the image, especially at the edges.
|
| 1098 |
+
"""
|
| 1099 |
+
# Calculate the ratio between the image size and the tile size
|
| 1100 |
+
height_ratio = height / tile_height
|
| 1101 |
+
width_ratio = width / tile_width
|
| 1102 |
+
|
| 1103 |
+
# If the ratio is greater than 6:1, use the formula with border_tile_overlap
|
| 1104 |
+
if height_ratio > 6 or width_ratio > 6:
|
| 1105 |
+
grid_rows = int(np.ceil((height - border_tile_overlap) / (tile_height - normal_tile_overlap))) + 1
|
| 1106 |
+
grid_cols = int(np.ceil((width - border_tile_overlap) / (tile_width - normal_tile_overlap))) + 1
|
| 1107 |
+
else:
|
| 1108 |
+
# Otherwise, use the universal formula
|
| 1109 |
+
grid_rows = int(np.ceil((height - normal_tile_overlap) / (tile_height - normal_tile_overlap)))
|
| 1110 |
+
grid_cols = int(np.ceil((width - normal_tile_overlap) / (tile_width - normal_tile_overlap)))
|
| 1111 |
+
|
| 1112 |
+
return grid_rows, grid_cols
|
| 1113 |
+
|
| 1114 |
+
def prepare_tiles(
|
| 1115 |
+
self,
|
| 1116 |
+
grid_rows,
|
| 1117 |
+
grid_cols,
|
| 1118 |
+
tile_weighting_method,
|
| 1119 |
+
tile_width,
|
| 1120 |
+
tile_height,
|
| 1121 |
+
normal_tile_overlap,
|
| 1122 |
+
border_tile_overlap,
|
| 1123 |
+
width,
|
| 1124 |
+
height,
|
| 1125 |
+
tile_sigma,
|
| 1126 |
+
batch_size,
|
| 1127 |
+
device,
|
| 1128 |
+
dtype,
|
| 1129 |
+
):
|
| 1130 |
+
"""
|
| 1131 |
+
Processes image tiles by dynamically adjusting overlap and calculating Gaussian or cosine weights.
|
| 1132 |
+
|
| 1133 |
+
Args:
|
| 1134 |
+
grid_rows (int): Number of rows in the tile grid.
|
| 1135 |
+
grid_cols (int): Number of columns in the tile grid.
|
| 1136 |
+
tile_weighting_method (str): Method for weighting tiles. Options: "Gaussian" or "Cosine".
|
| 1137 |
+
tile_width (int): Width of each tile in pixels.
|
| 1138 |
+
tile_height (int): Height of each tile in pixels.
|
| 1139 |
+
normal_tile_overlap (int): Overlap between tiles in pixels for normal tiles.
|
| 1140 |
+
border_tile_overlap (int): Overlap between tiles in pixels for border tiles.
|
| 1141 |
+
width (int): Width of the image in pixels.
|
| 1142 |
+
height (int): Height of the image in pixels.
|
| 1143 |
+
tile_sigma (float): Sigma parameter for Gaussian weighting.
|
| 1144 |
+
batch_size (int): Batch size for weight tiles.
|
| 1145 |
+
device (torch.device): Device where tensors will be allocated (e.g., 'cuda' or 'cpu').
|
| 1146 |
+
dtype (torch.dtype): Data type of the tensors (e.g., torch.float32).
|
| 1147 |
+
|
| 1148 |
+
Returns:
|
| 1149 |
+
tuple: A tuple containing:
|
| 1150 |
+
- tile_weights (np.ndarray): Array of weights for each tile.
|
| 1151 |
+
- tile_row_overlaps (np.ndarray): Array of row overlaps for each tile.
|
| 1152 |
+
- tile_col_overlaps (np.ndarray): Array of column overlaps for each tile.
|
| 1153 |
+
"""
|
| 1154 |
+
|
| 1155 |
+
# Create arrays to store dynamic overlaps and weights
|
| 1156 |
+
tile_row_overlaps = np.full((grid_rows, grid_cols), normal_tile_overlap)
|
| 1157 |
+
tile_col_overlaps = np.full((grid_rows, grid_cols), normal_tile_overlap)
|
| 1158 |
+
tile_weights = np.empty((grid_rows, grid_cols), dtype=object) # Stores Gaussian or cosine weights
|
| 1159 |
+
|
| 1160 |
+
# Iterate over tiles to adjust overlap and calculate weights
|
| 1161 |
+
for row in range(grid_rows):
|
| 1162 |
+
for col in range(grid_cols):
|
| 1163 |
+
# Calculate the size of the current tile
|
| 1164 |
+
px_row_init, px_row_end, px_col_init, px_col_end = _tile2pixel_indices(
|
| 1165 |
+
row, col, tile_width, tile_height, normal_tile_overlap, normal_tile_overlap, width, height
|
| 1166 |
+
)
|
| 1167 |
+
current_tile_width = px_col_end - px_col_init
|
| 1168 |
+
current_tile_height = px_row_end - px_row_init
|
| 1169 |
+
sigma = tile_sigma
|
| 1170 |
+
|
| 1171 |
+
# Adjust overlap for smaller tiles
|
| 1172 |
+
if current_tile_width < tile_width:
|
| 1173 |
+
px_row_init, px_row_end, px_col_init, px_col_end = _tile2pixel_indices(
|
| 1174 |
+
row, col, tile_width, tile_height, border_tile_overlap, border_tile_overlap, width, height
|
| 1175 |
+
)
|
| 1176 |
+
current_tile_width = px_col_end - px_col_init
|
| 1177 |
+
tile_col_overlaps[row, col] = border_tile_overlap
|
| 1178 |
+
sigma = tile_sigma * 1.2
|
| 1179 |
+
if current_tile_height < tile_height:
|
| 1180 |
+
px_row_init, px_row_end, px_col_init, px_col_end = _tile2pixel_indices(
|
| 1181 |
+
row, col, tile_width, tile_height, border_tile_overlap, border_tile_overlap, width, height
|
| 1182 |
+
)
|
| 1183 |
+
current_tile_height = px_row_end - px_row_init
|
| 1184 |
+
tile_row_overlaps[row, col] = border_tile_overlap
|
| 1185 |
+
sigma = tile_sigma * 1.2
|
| 1186 |
+
|
| 1187 |
+
# Calculate weights for the current tile
|
| 1188 |
+
if tile_weighting_method == TileWeightingMethod.COSINE.value:
|
| 1189 |
+
tile_weights[row, col] = self._generate_cosine_weights(
|
| 1190 |
+
tile_width=current_tile_width,
|
| 1191 |
+
tile_height=current_tile_height,
|
| 1192 |
+
nbatches=batch_size,
|
| 1193 |
+
device=device,
|
| 1194 |
+
dtype=torch.float32,
|
| 1195 |
+
)
|
| 1196 |
+
else:
|
| 1197 |
+
tile_weights[row, col] = self._generate_gaussian_weights(
|
| 1198 |
+
tile_width=current_tile_width,
|
| 1199 |
+
tile_height=current_tile_height,
|
| 1200 |
+
nbatches=batch_size,
|
| 1201 |
+
device=device,
|
| 1202 |
+
dtype=dtype,
|
| 1203 |
+
sigma=sigma,
|
| 1204 |
+
)
|
| 1205 |
+
|
| 1206 |
+
return tile_weights, tile_row_overlaps, tile_col_overlaps
|
| 1207 |
+
|
| 1208 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
|
| 1209 |
+
def upcast_vae(self):
|
| 1210 |
+
dtype = self.vae.dtype
|
| 1211 |
+
self.vae.to(dtype=torch.float32)
|
| 1212 |
+
use_torch_2_0_or_xformers = isinstance(
|
| 1213 |
+
self.vae.decoder.mid_block.attentions[0].processor,
|
| 1214 |
+
(
|
| 1215 |
+
AttnProcessor2_0,
|
| 1216 |
+
XFormersAttnProcessor,
|
| 1217 |
+
),
|
| 1218 |
+
)
|
| 1219 |
+
# if xformers or torch_2_0 is used attention block does not need
|
| 1220 |
+
# to be in float32 which can save lots of memory
|
| 1221 |
+
if use_torch_2_0_or_xformers:
|
| 1222 |
+
self.vae.post_quant_conv.to(dtype)
|
| 1223 |
+
self.vae.decoder.conv_in.to(dtype)
|
| 1224 |
+
self.vae.decoder.mid_block.to(dtype)
|
| 1225 |
+
|
| 1226 |
+
@property
|
| 1227 |
+
def guidance_scale(self):
|
| 1228 |
+
return self._guidance_scale
|
| 1229 |
+
|
| 1230 |
+
@property
|
| 1231 |
+
def clip_skip(self):
|
| 1232 |
+
return self._clip_skip
|
| 1233 |
+
|
| 1234 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
| 1235 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
| 1236 |
+
# corresponds to doing no classifier free guidance.
|
| 1237 |
+
@property
|
| 1238 |
+
def do_classifier_free_guidance(self):
|
| 1239 |
+
return self._guidance_scale > 1
|
| 1240 |
+
|
| 1241 |
+
@property
|
| 1242 |
+
def cross_attention_kwargs(self):
|
| 1243 |
+
return self._cross_attention_kwargs
|
| 1244 |
+
|
| 1245 |
+
@property
|
| 1246 |
+
def num_timesteps(self):
|
| 1247 |
+
return self._num_timesteps
|
| 1248 |
+
|
| 1249 |
+
@property
|
| 1250 |
+
def interrupt(self):
|
| 1251 |
+
return self._interrupt
|
| 1252 |
+
|
| 1253 |
+
@torch.no_grad()
|
| 1254 |
+
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
| 1255 |
+
def __call__(
|
| 1256 |
+
self,
|
| 1257 |
+
prompt: Union[str, List[str]] = None,
|
| 1258 |
+
image: PipelineImageInput = None,
|
| 1259 |
+
control_image: PipelineImageInput = None,
|
| 1260 |
+
height: Optional[int] = None,
|
| 1261 |
+
width: Optional[int] = None,
|
| 1262 |
+
strength: float = 0.9999,
|
| 1263 |
+
num_inference_steps: int = 50,
|
| 1264 |
+
guidance_scale: float = 5.0,
|
| 1265 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
| 1266 |
+
num_images_per_prompt: Optional[int] = 1,
|
| 1267 |
+
eta: float = 0.0,
|
| 1268 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
| 1269 |
+
latents: Optional[torch.Tensor] = None,
|
| 1270 |
+
output_type: Optional[str] = "pil",
|
| 1271 |
+
return_dict: bool = True,
|
| 1272 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
| 1273 |
+
controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
|
| 1274 |
+
guess_mode: bool = False,
|
| 1275 |
+
control_guidance_start: Union[float, List[float]] = 0.0,
|
| 1276 |
+
control_guidance_end: Union[float, List[float]] = 1.0,
|
| 1277 |
+
control_mode: Optional[Union[int, List[int]]] = None,
|
| 1278 |
+
original_size: Tuple[int, int] = None,
|
| 1279 |
+
crops_coords_top_left: Tuple[int, int] = (0, 0),
|
| 1280 |
+
target_size: Tuple[int, int] = None,
|
| 1281 |
+
negative_original_size: Optional[Tuple[int, int]] = None,
|
| 1282 |
+
negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
|
| 1283 |
+
negative_target_size: Optional[Tuple[int, int]] = None,
|
| 1284 |
+
aesthetic_score: float = 6.0,
|
| 1285 |
+
negative_aesthetic_score: float = 2.5,
|
| 1286 |
+
clip_skip: Optional[int] = None,
|
| 1287 |
+
normal_tile_overlap: int = 64,
|
| 1288 |
+
border_tile_overlap: int = 128,
|
| 1289 |
+
max_tile_size: int = 1024,
|
| 1290 |
+
tile_gaussian_sigma: float = 0.05,
|
| 1291 |
+
tile_weighting_method: str = "Cosine",
|
| 1292 |
+
**kwargs,
|
| 1293 |
+
):
|
| 1294 |
+
r"""
|
| 1295 |
+
Function invoked when calling the pipeline for generation.
|
| 1296 |
+
|
| 1297 |
+
Args:
|
| 1298 |
+
prompt (`str` or `List[str]`, *optional*):
|
| 1299 |
+
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
|
| 1300 |
+
image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`, *optional*):
|
| 1301 |
+
The initial image to be used as the starting point for the image generation process. Can also accept
|
| 1302 |
+
image latents as `image`, if passing latents directly, they will not be encoded again.
|
| 1303 |
+
control_image (`PipelineImageInput`, *optional*):
|
| 1304 |
+
The ControlNet input condition. ControlNet uses this input condition to generate guidance for Unet.
|
| 1305 |
+
If the type is specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also
|
| 1306 |
+
be accepted as an image. The dimensions of the output image default to `image`'s dimensions. If height
|
| 1307 |
+
and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
|
| 1308 |
+
init, images must be passed as a list such that each element of the list can be correctly batched for
|
| 1309 |
+
input to a single ControlNet.
|
| 1310 |
+
height (`int`, *optional*):
|
| 1311 |
+
The height in pixels of the generated image. If not provided, defaults to the height of `control_image`.
|
| 1312 |
+
width (`int`, *optional*):
|
| 1313 |
+
The width in pixels of the generated image. If not provided, defaults to the width of `control_image`.
|
| 1314 |
+
strength (`float`, *optional*, defaults to 0.9999):
|
| 1315 |
+
Indicates the extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
|
| 1316 |
+
starting point, and more noise is added the higher the `strength`. The number of denoising steps depends
|
| 1317 |
+
on the amount of noise initially added. When `strength` is 1, added noise is maximum, and the denoising
|
| 1318 |
+
process runs for the full number of iterations specified in `num_inference_steps`.
|
| 1319 |
+
num_inference_steps (`int`, *optional*, defaults to 50):
|
| 1320 |
+
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
| 1321 |
+
expense of slower inference.
|
| 1322 |
+
guidance_scale (`float`, *optional*, defaults to 5.0):
|
| 1323 |
+
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
| 1324 |
+
`guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf).
|
| 1325 |
+
Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages generating
|
| 1326 |
+
images closely linked to the text `prompt`, usually at the expense of lower image quality.
|
| 1327 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
| 1328 |
+
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
| 1329 |
+
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
| 1330 |
+
less than `1`).
|
| 1331 |
+
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
| 1332 |
+
The number of images to generate per prompt.
|
| 1333 |
+
eta (`float`, *optional*, defaults to 0.0):
|
| 1334 |
+
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
|
| 1335 |
+
[`schedulers.DDIMScheduler`], will be ignored for others.
|
| 1336 |
+
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
| 1337 |
+
One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
|
| 1338 |
+
to make generation deterministic.
|
| 1339 |
+
latents (`torch.Tensor`, *optional*):
|
| 1340 |
+
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
|
| 1341 |
+
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
|
| 1342 |
+
tensor will be generated by sampling using the supplied random `generator`.
|
| 1343 |
+
output_type (`str`, *optional*, defaults to `"pil"`):
|
| 1344 |
+
The output format of the generated image. Choose between [PIL](https://pillow.readthedocs.io/en/stable/):
|
| 1345 |
+
`PIL.Image.Image` or `np.array`.
|
| 1346 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
| 1347 |
+
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
|
| 1348 |
+
plain tuple.
|
| 1349 |
+
cross_attention_kwargs (`dict`, *optional*):
|
| 1350 |
+
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
| 1351 |
+
`self.processor` in
|
| 1352 |
+
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
| 1353 |
+
controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
|
| 1354 |
+
The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
|
| 1355 |
+
to the residual in the original UNet. If multiple ControlNets are specified in init, you can set the
|
| 1356 |
+
corresponding scale as a list.
|
| 1357 |
+
guess_mode (`bool`, *optional*, defaults to `False`):
|
| 1358 |
+
In this mode, the ControlNet encoder will try to recognize the content of the input image even if
|
| 1359 |
+
you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended.
|
| 1360 |
+
control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
|
| 1361 |
+
The percentage of total steps at which the ControlNet starts applying.
|
| 1362 |
+
control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
|
| 1363 |
+
The percentage of total steps at which the ControlNet stops applying.
|
| 1364 |
+
control_mode (`int` or `List[int]`, *optional*):
|
| 1365 |
+
The mode of ControlNet guidance. Can be used to specify different behaviors for multiple ControlNets.
|
| 1366 |
+
original_size (`Tuple[int, int]`, *optional*):
|
| 1367 |
+
If `original_size` is not the same as `target_size`, the image will appear to be down- or upsampled.
|
| 1368 |
+
`original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning.
|
| 1369 |
+
crops_coords_top_left (`Tuple[int, int]`, *optional*, defaults to (0, 0)):
|
| 1370 |
+
`crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
|
| 1371 |
+
`crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
|
| 1372 |
+
`crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning.
|
| 1373 |
+
target_size (`Tuple[int, int]`, *optional*):
|
| 1374 |
+
For most cases, `target_size` should be set to the desired height and width of the generated image. If
|
| 1375 |
+
not specified, it will default to `(height, width)`. Part of SDXL's micro-conditioning.
|
| 1376 |
+
negative_original_size (`Tuple[int, int]`, *optional*):
|
| 1377 |
+
To negatively condition the generation process based on a specific image resolution. Part of SDXL's
|
| 1378 |
+
micro-conditioning.
|
| 1379 |
+
negative_crops_coords_top_left (`Tuple[int, int]`, *optional*, defaults to (0, 0)):
|
| 1380 |
+
To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
|
| 1381 |
+
micro-conditioning.
|
| 1382 |
+
negative_target_size (`Tuple[int, int]`, *optional*):
|
| 1383 |
+
To negatively condition the generation process based on a target image resolution. It should be the same
|
| 1384 |
+
as the `target_size` for most cases. Part of SDXL's micro-conditioning.
|
| 1385 |
+
aesthetic_score (`float`, *optional*, defaults to 6.0):
|
| 1386 |
+
Used to simulate an aesthetic score of the generated image by influencing the positive text condition.
|
| 1387 |
+
Part of SDXL's micro-conditioning.
|
| 1388 |
+
negative_aesthetic_score (`float`, *optional*, defaults to 2.5):
|
| 1389 |
+
Used to simulate an aesthetic score of the generated image by influencing the negative text condition.
|
| 1390 |
+
Part of SDXL's micro-conditioning.
|
| 1391 |
+
clip_skip (`int`, *optional*):
|
| 1392 |
+
Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
|
| 1393 |
+
the output of the pre-final layer will be used for computing the prompt embeddings.
|
| 1394 |
+
normal_tile_overlap (`int`, *optional*, defaults to 64):
|
| 1395 |
+
Number of overlapping pixels between tiles in consecutive rows.
|
| 1396 |
+
border_tile_overlap (`int`, *optional*, defaults to 128):
|
| 1397 |
+
Number of overlapping pixels between tiles at the borders.
|
| 1398 |
+
max_tile_size (`int`, *optional*, defaults to 1024):
|
| 1399 |
+
Maximum size of a tile in pixels.
|
| 1400 |
+
tile_gaussian_sigma (`float`, *optional*, defaults to 0.3):
|
| 1401 |
+
Sigma parameter for Gaussian weighting of tiles.
|
| 1402 |
+
tile_weighting_method (`str`, *optional*, defaults to "Cosine"):
|
| 1403 |
+
Method for weighting tiles. Options: "Cosine" or "Gaussian".
|
| 1404 |
+
|
| 1405 |
+
Examples:
|
| 1406 |
+
|
| 1407 |
+
Returns:
|
| 1408 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
|
| 1409 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple`
|
| 1410 |
+
containing the output images.
|
| 1411 |
+
"""
|
| 1412 |
+
|
| 1413 |
+
controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
|
| 1414 |
+
|
| 1415 |
+
# align format for control guidance
|
| 1416 |
+
if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
|
| 1417 |
+
control_guidance_start = len(control_guidance_end) * [control_guidance_start]
|
| 1418 |
+
elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
|
| 1419 |
+
control_guidance_end = len(control_guidance_start) * [control_guidance_end]
|
| 1420 |
+
|
| 1421 |
+
if not isinstance(control_image, list):
|
| 1422 |
+
control_image = [control_image]
|
| 1423 |
+
else:
|
| 1424 |
+
control_image = control_image.copy()
|
| 1425 |
+
|
| 1426 |
+
if control_mode is None or isinstance(control_mode, list) and len(control_mode) == 0:
|
| 1427 |
+
raise ValueError("The value for `control_mode` is expected!")
|
| 1428 |
+
|
| 1429 |
+
if not isinstance(control_mode, list):
|
| 1430 |
+
control_mode = [control_mode]
|
| 1431 |
+
|
| 1432 |
+
if len(control_image) != len(control_mode):
|
| 1433 |
+
raise ValueError("Expected len(control_image) == len(control_mode)")
|
| 1434 |
+
|
| 1435 |
+
num_control_type = controlnet.config.num_control_type
|
| 1436 |
+
|
| 1437 |
+
# 0. Set internal use parameters
|
| 1438 |
+
height = height or self.unet.config.sample_size * self.vae_scale_factor
|
| 1439 |
+
width = width or self.unet.config.sample_size * self.vae_scale_factor
|
| 1440 |
+
original_size = original_size or (height, width)
|
| 1441 |
+
target_size = target_size or (height, width)
|
| 1442 |
+
negative_original_size = negative_original_size or original_size
|
| 1443 |
+
negative_target_size = negative_target_size or target_size
|
| 1444 |
+
control_type = [0 for _ in range(num_control_type)]
|
| 1445 |
+
control_type = torch.Tensor(control_type)
|
| 1446 |
+
self._guidance_scale = guidance_scale
|
| 1447 |
+
self._clip_skip = clip_skip
|
| 1448 |
+
self._cross_attention_kwargs = cross_attention_kwargs
|
| 1449 |
+
self._interrupt = False
|
| 1450 |
+
batch_size = 1
|
| 1451 |
+
device = self._execution_device
|
| 1452 |
+
global_pool_conditions = controlnet.config.global_pool_conditions
|
| 1453 |
+
guess_mode = guess_mode or global_pool_conditions
|
| 1454 |
+
|
| 1455 |
+
# 1. Check inputs
|
| 1456 |
+
for _image, control_idx in zip(control_image, control_mode):
|
| 1457 |
+
control_type[control_idx] = 1
|
| 1458 |
+
self.check_inputs(
|
| 1459 |
+
prompt,
|
| 1460 |
+
height,
|
| 1461 |
+
width,
|
| 1462 |
+
_image,
|
| 1463 |
+
strength,
|
| 1464 |
+
num_inference_steps,
|
| 1465 |
+
normal_tile_overlap,
|
| 1466 |
+
border_tile_overlap,
|
| 1467 |
+
max_tile_size,
|
| 1468 |
+
tile_gaussian_sigma,
|
| 1469 |
+
tile_weighting_method,
|
| 1470 |
+
controlnet_conditioning_scale,
|
| 1471 |
+
control_guidance_start,
|
| 1472 |
+
control_guidance_end,
|
| 1473 |
+
)
|
| 1474 |
+
|
| 1475 |
+
# 2 Get tile width and tile height size
|
| 1476 |
+
tile_width, tile_height = _adaptive_tile_size((width, height), max_tile_size=max_tile_size)
|
| 1477 |
+
|
| 1478 |
+
# 2.1 Calculate the number of tiles needed
|
| 1479 |
+
grid_rows, grid_cols = self._get_num_tiles(height, width, tile_height, tile_width, normal_tile_overlap, border_tile_overlap)
|
| 1480 |
+
|
| 1481 |
+
# 2.2 Expand prompt to number of tiles
|
| 1482 |
+
if not isinstance(prompt, list):
|
| 1483 |
+
prompt = [[prompt] * grid_cols] * grid_rows
|
| 1484 |
+
|
| 1485 |
+
# 2.3 Update height and width tile size by tile size and tile overlap size
|
| 1486 |
+
width = (grid_cols - 1) * (tile_width - normal_tile_overlap) + min(
|
| 1487 |
+
tile_width, width - (grid_cols - 1) * (tile_width - normal_tile_overlap)
|
| 1488 |
+
)
|
| 1489 |
+
height = (grid_rows - 1) * (tile_height - normal_tile_overlap) + min(
|
| 1490 |
+
tile_height, height - (grid_rows - 1) * (tile_height - normal_tile_overlap)
|
| 1491 |
+
)
|
| 1492 |
+
|
| 1493 |
+
# 3. Encode input prompt
|
| 1494 |
+
text_encoder_lora_scale = (
|
| 1495 |
+
self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
|
| 1496 |
+
)
|
| 1497 |
+
text_embeddings = [
|
| 1498 |
+
[
|
| 1499 |
+
self.encode_prompt(
|
| 1500 |
+
prompt=col,
|
| 1501 |
+
device=device,
|
| 1502 |
+
num_images_per_prompt=num_images_per_prompt,
|
| 1503 |
+
do_classifier_free_guidance=self.do_classifier_free_guidance,
|
| 1504 |
+
negative_prompt=negative_prompt,
|
| 1505 |
+
prompt_embeds=None,
|
| 1506 |
+
negative_prompt_embeds=None,
|
| 1507 |
+
pooled_prompt_embeds=None,
|
| 1508 |
+
negative_pooled_prompt_embeds=None,
|
| 1509 |
+
lora_scale=text_encoder_lora_scale,
|
| 1510 |
+
clip_skip=self.clip_skip,
|
| 1511 |
+
)
|
| 1512 |
+
for col in row
|
| 1513 |
+
]
|
| 1514 |
+
for row in prompt
|
| 1515 |
+
]
|
| 1516 |
+
|
| 1517 |
+
# 4. Prepare latent image
|
| 1518 |
+
image_tensor = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
|
| 1519 |
+
|
| 1520 |
+
# 4.1 Prepare controlnet_conditioning_image
|
| 1521 |
+
control_image = self.prepare_control_image(
|
| 1522 |
+
image=image,
|
| 1523 |
+
width=width,
|
| 1524 |
+
height=height,
|
| 1525 |
+
batch_size=batch_size * num_images_per_prompt,
|
| 1526 |
+
num_images_per_prompt=num_images_per_prompt,
|
| 1527 |
+
device=device,
|
| 1528 |
+
dtype=controlnet.dtype,
|
| 1529 |
+
do_classifier_free_guidance=self.do_classifier_free_guidance,
|
| 1530 |
+
guess_mode=guess_mode,
|
| 1531 |
+
)
|
| 1532 |
+
control_type = (
|
| 1533 |
+
control_type.reshape(1, -1)
|
| 1534 |
+
.to(device, dtype=controlnet.dtype)
|
| 1535 |
+
.repeat(batch_size * num_images_per_prompt * 2, 1)
|
| 1536 |
+
)
|
| 1537 |
+
|
| 1538 |
+
# 5. Prepare timesteps
|
| 1539 |
+
accepts_offset = "offset" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys())
|
| 1540 |
+
extra_set_kwargs = {}
|
| 1541 |
+
if accepts_offset:
|
| 1542 |
+
extra_set_kwargs["offset"] = 1
|
| 1543 |
+
self.scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
|
| 1544 |
+
timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength)
|
| 1545 |
+
latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
|
| 1546 |
+
self._num_timesteps = len(timesteps)
|
| 1547 |
+
|
| 1548 |
+
# 6. Prepare latent variables
|
| 1549 |
+
dtype = text_embeddings[0][0][0].dtype
|
| 1550 |
+
if latents is None:
|
| 1551 |
+
latents = self.prepare_latents(
|
| 1552 |
+
image_tensor,
|
| 1553 |
+
latent_timestep,
|
| 1554 |
+
batch_size,
|
| 1555 |
+
num_images_per_prompt,
|
| 1556 |
+
dtype,
|
| 1557 |
+
device,
|
| 1558 |
+
generator,
|
| 1559 |
+
True,
|
| 1560 |
+
)
|
| 1561 |
+
|
| 1562 |
+
# if we use LMSDiscreteScheduler, let's make sure latents are multiplied by sigmas
|
| 1563 |
+
if isinstance(self.scheduler, LMSDiscreteScheduler):
|
| 1564 |
+
latents = latents * self.scheduler.sigmas[0]
|
| 1565 |
+
|
| 1566 |
+
# 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
| 1567 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
| 1568 |
+
|
| 1569 |
+
# 8. Create tensor stating which controlnets to keep
|
| 1570 |
+
controlnet_keep = []
|
| 1571 |
+
for i in range(len(timesteps)):
|
| 1572 |
+
controlnet_keep.append(
|
| 1573 |
+
1.0
|
| 1574 |
+
- float(i / len(timesteps) < control_guidance_start or (i + 1) / len(timesteps) > control_guidance_end)
|
| 1575 |
+
)
|
| 1576 |
+
|
| 1577 |
+
# 8.1 Prepare added time ids & embeddings
|
| 1578 |
+
# text_embeddings order: prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
|
| 1579 |
+
embeddings_and_added_time = []
|
| 1580 |
+
crops_coords_top_left = negative_crops_coords_top_left = (tile_width, tile_height)
|
| 1581 |
+
for row in range(grid_rows):
|
| 1582 |
+
addition_embed_type_row = []
|
| 1583 |
+
for col in range(grid_cols):
|
| 1584 |
+
# extract generated values
|
| 1585 |
+
prompt_embeds = text_embeddings[row][col][0]
|
| 1586 |
+
negative_prompt_embeds = text_embeddings[row][col][1]
|
| 1587 |
+
pooled_prompt_embeds = text_embeddings[row][col][2]
|
| 1588 |
+
negative_pooled_prompt_embeds = text_embeddings[row][col][3]
|
| 1589 |
+
|
| 1590 |
+
if negative_original_size is None:
|
| 1591 |
+
negative_original_size = original_size
|
| 1592 |
+
if negative_target_size is None:
|
| 1593 |
+
negative_target_size = target_size
|
| 1594 |
+
add_text_embeds = pooled_prompt_embeds
|
| 1595 |
+
|
| 1596 |
+
if self.text_encoder_2 is None:
|
| 1597 |
+
text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
|
| 1598 |
+
else:
|
| 1599 |
+
text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
|
| 1600 |
+
|
| 1601 |
+
add_time_ids, add_neg_time_ids = self._get_add_time_ids(
|
| 1602 |
+
original_size,
|
| 1603 |
+
crops_coords_top_left,
|
| 1604 |
+
target_size,
|
| 1605 |
+
aesthetic_score,
|
| 1606 |
+
negative_aesthetic_score,
|
| 1607 |
+
negative_original_size,
|
| 1608 |
+
negative_crops_coords_top_left,
|
| 1609 |
+
negative_target_size,
|
| 1610 |
+
dtype=prompt_embeds.dtype,
|
| 1611 |
+
text_encoder_projection_dim=text_encoder_projection_dim,
|
| 1612 |
+
)
|
| 1613 |
+
add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1)
|
| 1614 |
+
|
| 1615 |
+
if self.do_classifier_free_guidance:
|
| 1616 |
+
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
|
| 1617 |
+
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
|
| 1618 |
+
add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1)
|
| 1619 |
+
add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0)
|
| 1620 |
+
|
| 1621 |
+
prompt_embeds = prompt_embeds.to(device)
|
| 1622 |
+
add_text_embeds = add_text_embeds.to(device)
|
| 1623 |
+
add_time_ids = add_time_ids.to(device)
|
| 1624 |
+
addition_embed_type_row.append((prompt_embeds, add_text_embeds, add_time_ids))
|
| 1625 |
+
|
| 1626 |
+
embeddings_and_added_time.append(addition_embed_type_row)
|
| 1627 |
+
|
| 1628 |
+
# 9. Prepare tiles weights and latent overlaps size to denoising process
|
| 1629 |
+
tile_weights, tile_row_overlaps, tile_col_overlaps = self.prepare_tiles(
|
| 1630 |
+
grid_rows,
|
| 1631 |
+
grid_cols,
|
| 1632 |
+
tile_weighting_method,
|
| 1633 |
+
tile_width,
|
| 1634 |
+
tile_height,
|
| 1635 |
+
normal_tile_overlap,
|
| 1636 |
+
border_tile_overlap,
|
| 1637 |
+
width,
|
| 1638 |
+
height,
|
| 1639 |
+
tile_gaussian_sigma,
|
| 1640 |
+
batch_size,
|
| 1641 |
+
device,
|
| 1642 |
+
dtype,
|
| 1643 |
+
)
|
| 1644 |
+
|
| 1645 |
+
# 10. Denoising loop
|
| 1646 |
+
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
| 1647 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
| 1648 |
+
for i, t in enumerate(timesteps):
|
| 1649 |
+
# Diffuse each tile
|
| 1650 |
+
noise_preds = []
|
| 1651 |
+
for row in range(grid_rows):
|
| 1652 |
+
noise_preds_row = []
|
| 1653 |
+
for col in range(grid_cols):
|
| 1654 |
+
if self.interrupt:
|
| 1655 |
+
continue
|
| 1656 |
+
tile_row_overlap = tile_row_overlaps[row, col]
|
| 1657 |
+
tile_col_overlap = tile_col_overlaps[row, col]
|
| 1658 |
+
|
| 1659 |
+
px_row_init, px_row_end, px_col_init, px_col_end = _tile2latent_indices(
|
| 1660 |
+
row, col, tile_width, tile_height, tile_row_overlap, tile_col_overlap, width, height
|
| 1661 |
+
)
|
| 1662 |
+
|
| 1663 |
+
tile_latents = latents[:, :, px_row_init:px_row_end, px_col_init:px_col_end]
|
| 1664 |
+
|
| 1665 |
+
# expand the latents if we are doing classifier free guidance
|
| 1666 |
+
latent_model_input = (
|
| 1667 |
+
torch.cat([tile_latents] * 2)
|
| 1668 |
+
if self.do_classifier_free_guidance
|
| 1669 |
+
else tile_latents # 1, 4, ...
|
| 1670 |
+
)
|
| 1671 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
| 1672 |
+
|
| 1673 |
+
# predict the noise residual
|
| 1674 |
+
added_cond_kwargs = {
|
| 1675 |
+
"text_embeds": embeddings_and_added_time[row][col][1],
|
| 1676 |
+
"time_ids": embeddings_and_added_time[row][col][2],
|
| 1677 |
+
}
|
| 1678 |
+
|
| 1679 |
+
# controlnet(s) inference
|
| 1680 |
+
if guess_mode and self.do_classifier_free_guidance:
|
| 1681 |
+
# Infer ControlNet only for the conditional batch.
|
| 1682 |
+
control_model_input = tile_latents
|
| 1683 |
+
control_model_input = self.scheduler.scale_model_input(control_model_input, t)
|
| 1684 |
+
controlnet_prompt_embeds = embeddings_and_added_time[row][col][0].chunk(2)[1]
|
| 1685 |
+
controlnet_added_cond_kwargs = {
|
| 1686 |
+
"text_embeds": embeddings_and_added_time[row][col][1].chunk(2)[1],
|
| 1687 |
+
"time_ids": embeddings_and_added_time[row][col][2].chunk(2)[1],
|
| 1688 |
+
}
|
| 1689 |
+
else:
|
| 1690 |
+
control_model_input = latent_model_input
|
| 1691 |
+
controlnet_prompt_embeds = embeddings_and_added_time[row][col][0]
|
| 1692 |
+
controlnet_added_cond_kwargs = added_cond_kwargs
|
| 1693 |
+
|
| 1694 |
+
if isinstance(controlnet_keep[i], list):
|
| 1695 |
+
cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
|
| 1696 |
+
else:
|
| 1697 |
+
controlnet_cond_scale = controlnet_conditioning_scale
|
| 1698 |
+
if isinstance(controlnet_cond_scale, list):
|
| 1699 |
+
controlnet_cond_scale = controlnet_cond_scale[0]
|
| 1700 |
+
cond_scale = controlnet_cond_scale * controlnet_keep[i]
|
| 1701 |
+
|
| 1702 |
+
px_row_init_pixel, px_row_end_pixel, px_col_init_pixel, px_col_end_pixel = _tile2pixel_indices(
|
| 1703 |
+
row, col, tile_width, tile_height, tile_row_overlap, tile_col_overlap, width, height
|
| 1704 |
+
)
|
| 1705 |
+
|
| 1706 |
+
tile_control_image = control_image[
|
| 1707 |
+
:, :, px_row_init_pixel:px_row_end_pixel, px_col_init_pixel:px_col_end_pixel
|
| 1708 |
+
]
|
| 1709 |
+
|
| 1710 |
+
down_block_res_samples, mid_block_res_sample = self.controlnet(
|
| 1711 |
+
control_model_input,
|
| 1712 |
+
t,
|
| 1713 |
+
encoder_hidden_states=controlnet_prompt_embeds,
|
| 1714 |
+
controlnet_cond=[tile_control_image],
|
| 1715 |
+
control_type=control_type,
|
| 1716 |
+
control_type_idx=control_mode,
|
| 1717 |
+
conditioning_scale=cond_scale,
|
| 1718 |
+
guess_mode=guess_mode,
|
| 1719 |
+
added_cond_kwargs=controlnet_added_cond_kwargs,
|
| 1720 |
+
return_dict=False,
|
| 1721 |
+
)
|
| 1722 |
+
|
| 1723 |
+
if guess_mode and self.do_classifier_free_guidance:
|
| 1724 |
+
# Inferred ControlNet only for the conditional batch.
|
| 1725 |
+
# To apply the output of ControlNet to both the unconditional and conditional batches,
|
| 1726 |
+
# add 0 to the unconditional batch to keep it unchanged.
|
| 1727 |
+
down_block_res_samples = [
|
| 1728 |
+
torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples
|
| 1729 |
+
]
|
| 1730 |
+
mid_block_res_sample = torch.cat(
|
| 1731 |
+
[torch.zeros_like(mid_block_res_sample), mid_block_res_sample]
|
| 1732 |
+
)
|
| 1733 |
+
|
| 1734 |
+
# predict the noise residual
|
| 1735 |
+
with torch.amp.autocast(device.type, dtype=dtype, enabled=dtype != self.unet.dtype):
|
| 1736 |
+
noise_pred = self.unet(
|
| 1737 |
+
latent_model_input,
|
| 1738 |
+
t,
|
| 1739 |
+
encoder_hidden_states=embeddings_and_added_time[row][col][0],
|
| 1740 |
+
cross_attention_kwargs=self.cross_attention_kwargs,
|
| 1741 |
+
down_block_additional_residuals=down_block_res_samples,
|
| 1742 |
+
mid_block_additional_residual=mid_block_res_sample,
|
| 1743 |
+
added_cond_kwargs=added_cond_kwargs,
|
| 1744 |
+
return_dict=False,
|
| 1745 |
+
)[0]
|
| 1746 |
+
|
| 1747 |
+
# perform guidance
|
| 1748 |
+
if self.do_classifier_free_guidance:
|
| 1749 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
| 1750 |
+
noise_pred_tile = noise_pred_uncond + guidance_scale * (
|
| 1751 |
+
noise_pred_text - noise_pred_uncond
|
| 1752 |
+
)
|
| 1753 |
+
noise_preds_row.append(noise_pred_tile)
|
| 1754 |
+
noise_preds.append(noise_preds_row)
|
| 1755 |
+
|
| 1756 |
+
# Stitch noise predictions for all tiles
|
| 1757 |
+
noise_pred = torch.zeros(latents.shape, device=device)
|
| 1758 |
+
contributors = torch.zeros(latents.shape, device=device)
|
| 1759 |
+
|
| 1760 |
+
# Add each tile contribution to overall latents
|
| 1761 |
+
for row in range(grid_rows):
|
| 1762 |
+
for col in range(grid_cols):
|
| 1763 |
+
tile_row_overlap = tile_row_overlaps[row, col]
|
| 1764 |
+
tile_col_overlap = tile_col_overlaps[row, col]
|
| 1765 |
+
px_row_init, px_row_end, px_col_init, px_col_end = _tile2latent_indices(
|
| 1766 |
+
row, col, tile_width, tile_height, tile_row_overlap, tile_col_overlap, width, height
|
| 1767 |
+
)
|
| 1768 |
+
tile_weights_resized = tile_weights[row, col]
|
| 1769 |
+
|
| 1770 |
+
noise_pred[:, :, px_row_init:px_row_end, px_col_init:px_col_end] += (
|
| 1771 |
+
noise_preds[row][col] * tile_weights_resized
|
| 1772 |
+
)
|
| 1773 |
+
contributors[:, :, px_row_init:px_row_end, px_col_init:px_col_end] += tile_weights_resized
|
| 1774 |
+
|
| 1775 |
+
# Average overlapping areas with more than 1 contributor
|
| 1776 |
+
noise_pred /= contributors
|
| 1777 |
+
noise_pred = noise_pred.to(dtype)
|
| 1778 |
+
|
| 1779 |
+
# compute the previous noisy sample x_t -> x_t-1
|
| 1780 |
+
latents_dtype = latents.dtype
|
| 1781 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
| 1782 |
+
if latents.dtype != latents_dtype:
|
| 1783 |
+
if torch.backends.mps.is_available():
|
| 1784 |
+
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
|
| 1785 |
+
latents = latents.to(latents_dtype)
|
| 1786 |
+
|
| 1787 |
+
# update progress bar
|
| 1788 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
| 1789 |
+
progress_bar.update()
|
| 1790 |
+
|
| 1791 |
+
if XLA_AVAILABLE:
|
| 1792 |
+
xm.mark_step()
|
| 1793 |
+
|
| 1794 |
+
# If we do sequential model offloading, let's offload unet and controlnet
|
| 1795 |
+
# manually for max memory savings
|
| 1796 |
+
if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
|
| 1797 |
+
self.unet.to("cpu")
|
| 1798 |
+
self.controlnet.to("cpu")
|
| 1799 |
+
torch.cuda.empty_cache()
|
| 1800 |
+
|
| 1801 |
+
if not output_type == "latent":
|
| 1802 |
+
# make sure the VAE is in float32 mode, as it overflows in float16
|
| 1803 |
+
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
|
| 1804 |
+
|
| 1805 |
+
if needs_upcasting:
|
| 1806 |
+
self.upcast_vae()
|
| 1807 |
+
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
|
| 1808 |
+
|
| 1809 |
+
# unscale/denormalize the latents
|
| 1810 |
+
# denormalize with the mean and std if available and not None
|
| 1811 |
+
has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
|
| 1812 |
+
has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
|
| 1813 |
+
if has_latents_mean and has_latents_std:
|
| 1814 |
+
latents_mean = (
|
| 1815 |
+
torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
|
| 1816 |
+
)
|
| 1817 |
+
latents_std = (
|
| 1818 |
+
torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
|
| 1819 |
+
)
|
| 1820 |
+
latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
|
| 1821 |
+
else:
|
| 1822 |
+
latents = latents / self.vae.config.scaling_factor
|
| 1823 |
+
|
| 1824 |
+
image = self.vae.decode(latents, return_dict=False)[0]
|
| 1825 |
+
|
| 1826 |
+
# cast back to fp16 if needed
|
| 1827 |
+
if needs_upcasting:
|
| 1828 |
+
self.vae.to(dtype=torch.float16)
|
| 1829 |
+
|
| 1830 |
+
# apply watermark if available
|
| 1831 |
+
if self.watermark is not None:
|
| 1832 |
+
image = self.watermark.apply_watermark(image)
|
| 1833 |
+
|
| 1834 |
+
image = self.image_processor.postprocess(image, output_type=output_type)
|
| 1835 |
+
else:
|
| 1836 |
+
image = latents
|
| 1837 |
+
|
| 1838 |
+
# Offload all models
|
| 1839 |
+
self.maybe_free_model_hooks()
|
| 1840 |
+
|
| 1841 |
+
result = StableDiffusionXLPipelineOutput(images=image)
|
| 1842 |
+
if not return_dict:
|
| 1843 |
+
return (image,)
|
| 1844 |
+
|
| 1845 |
+
return result
|
pipeline/util.py
ADDED
|
@@ -0,0 +1,328 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2025 The DEVAIEXP Team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
import gc
|
| 17 |
+
import cv2
|
| 18 |
+
import numpy as np
|
| 19 |
+
import torch
|
| 20 |
+
from PIL import Image
|
| 21 |
+
from gradio.themes import Default
|
| 22 |
+
import gradio as gr
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
MAX_SEED = np.iinfo(np.int32).max
|
| 26 |
+
SAMPLERS = {
|
| 27 |
+
"DDIM": ("DDIMScheduler", {}),
|
| 28 |
+
"DDIM trailing": ("DDIMScheduler", {"timestep_spacing": "trailing"}),
|
| 29 |
+
"DDPM": ("DDPMScheduler", {}),
|
| 30 |
+
"DEIS": ("DEISMultistepScheduler", {}),
|
| 31 |
+
"Heun": ("HeunDiscreteScheduler", {}),
|
| 32 |
+
"Heun Karras": ("HeunDiscreteScheduler", {"use_karras_sigmas": True}),
|
| 33 |
+
"Euler": ("EulerDiscreteScheduler", {}),
|
| 34 |
+
"Euler trailing": ("EulerDiscreteScheduler", {"timestep_spacing": "trailing", "prediction_type": "sample"}),
|
| 35 |
+
"Euler Ancestral": ("EulerAncestralDiscreteScheduler", {}),
|
| 36 |
+
"Euler Ancestral trailing": ("EulerAncestralDiscreteScheduler", {"timestep_spacing": "trailing"}),
|
| 37 |
+
"DPM++ 1S": ("DPMSolverMultistepScheduler", {"solver_order": 1}),
|
| 38 |
+
"DPM++ 1S Karras": ("DPMSolverMultistepScheduler", {"solver_order": 1, "use_karras_sigmas": True}),
|
| 39 |
+
"DPM++ 2S": ("DPMSolverSinglestepScheduler", {"use_karras_sigmas": False}),
|
| 40 |
+
"DPM++ 2S Karras": ("DPMSolverSinglestepScheduler", {"use_karras_sigmas": True}),
|
| 41 |
+
"DPM++ 2M": ("DPMSolverMultistepScheduler", {"use_karras_sigmas": False}),
|
| 42 |
+
"DPM++ 2M Karras": ("DPMSolverMultistepScheduler", {"use_karras_sigmas": True}),
|
| 43 |
+
"DPM++ 2M SDE": ("DPMSolverMultistepScheduler", {"use_karras_sigmas": False, "algorithm_type": "sde-dpmsolver++"}),
|
| 44 |
+
"DPM++ 2M SDE Karras": (
|
| 45 |
+
"DPMSolverMultistepScheduler",
|
| 46 |
+
{"use_karras_sigmas": True, "algorithm_type": "sde-dpmsolver++"},
|
| 47 |
+
),
|
| 48 |
+
"DPM++ 3M": ("DPMSolverMultistepScheduler", {"solver_order": 3}),
|
| 49 |
+
"DPM++ 3M Karras": ("DPMSolverMultistepScheduler", {"solver_order": 3, "use_karras_sigmas": True}),
|
| 50 |
+
"DPM++ SDE": ("DPMSolverSDEScheduler", {"use_karras_sigmas": False}),
|
| 51 |
+
"DPM++ SDE Karras": ("DPMSolverSDEScheduler", {"use_karras_sigmas": True}),
|
| 52 |
+
"DPM2": ("KDPM2DiscreteScheduler", {}),
|
| 53 |
+
"DPM2 Karras": ("KDPM2DiscreteScheduler", {"use_karras_sigmas": True}),
|
| 54 |
+
"DPM2 Ancestral": ("KDPM2AncestralDiscreteScheduler", {}),
|
| 55 |
+
"DPM2 Ancestral Karras": ("KDPM2AncestralDiscreteScheduler", {"use_karras_sigmas": True}),
|
| 56 |
+
"LMS": ("LMSDiscreteScheduler", {}),
|
| 57 |
+
"LMS Karras": ("LMSDiscreteScheduler", {"use_karras_sigmas": True}),
|
| 58 |
+
"UniPC": ("UniPCMultistepScheduler", {}),
|
| 59 |
+
"UniPC Karras": ("UniPCMultistepScheduler", {"use_karras_sigmas": True}),
|
| 60 |
+
"PNDM": ("PNDMScheduler", {}),
|
| 61 |
+
"Euler EDM": ("EDMEulerScheduler", {}),
|
| 62 |
+
"Euler EDM Karras": ("EDMEulerScheduler", {"use_karras_sigmas": True}),
|
| 63 |
+
"DPM++ 2M EDM": (
|
| 64 |
+
"EDMDPMSolverMultistepScheduler",
|
| 65 |
+
{"solver_order": 2, "solver_type": "midpoint", "final_sigmas_type": "zero", "algorithm_type": "dpmsolver++"},
|
| 66 |
+
),
|
| 67 |
+
"DPM++ 2M EDM Karras": (
|
| 68 |
+
"EDMDPMSolverMultistepScheduler",
|
| 69 |
+
{
|
| 70 |
+
"use_karras_sigmas": True,
|
| 71 |
+
"solver_order": 2,
|
| 72 |
+
"solver_type": "midpoint",
|
| 73 |
+
"final_sigmas_type": "zero",
|
| 74 |
+
"algorithm_type": "dpmsolver++",
|
| 75 |
+
},
|
| 76 |
+
),
|
| 77 |
+
"DPM++ 2M Lu": ("DPMSolverMultistepScheduler", {"use_lu_lambdas": True}),
|
| 78 |
+
"DPM++ 2M Ef": ("DPMSolverMultistepScheduler", {"euler_at_final": True}),
|
| 79 |
+
"DPM++ 2M SDE Lu": ("DPMSolverMultistepScheduler", {"use_lu_lambdas": True, "algorithm_type": "sde-dpmsolver++"}),
|
| 80 |
+
"DPM++ 2M SDE Ef": ("DPMSolverMultistepScheduler", {"algorithm_type": "sde-dpmsolver++", "euler_at_final": True}),
|
| 81 |
+
"LCM": ("LCMScheduler", {}),
|
| 82 |
+
"LCM trailing": ("LCMScheduler", {"timestep_spacing": "trailing"}),
|
| 83 |
+
"TCD": ("TCDScheduler", {}),
|
| 84 |
+
"TCD trailing": ("TCDScheduler", {"timestep_spacing": "trailing"}),
|
| 85 |
+
}
|
| 86 |
+
|
| 87 |
+
class Platinum(Default):
|
| 88 |
+
def __init__(
|
| 89 |
+
self,
|
| 90 |
+
):
|
| 91 |
+
super().__init__(
|
| 92 |
+
font = (
|
| 93 |
+
gr.themes.GoogleFont("Karla"), 'Segoe UI Emoji', 'Public Sans', 'system-ui', 'sans-serif'
|
| 94 |
+
)
|
| 95 |
+
)
|
| 96 |
+
self.name = "Diffusers"
|
| 97 |
+
super().set(
|
| 98 |
+
block_border_width='1px',
|
| 99 |
+
block_border_width_dark='1px',
|
| 100 |
+
block_info_text_size='13px',
|
| 101 |
+
block_info_text_weight='450',
|
| 102 |
+
block_info_text_color='#474a50',
|
| 103 |
+
block_label_background_fill='*background_fill_secondary',
|
| 104 |
+
block_label_text_color='*neutral_700',
|
| 105 |
+
block_title_text_color='black',
|
| 106 |
+
block_title_text_weight='600',
|
| 107 |
+
block_background_fill='#fcfcfc',
|
| 108 |
+
body_background_fill='*background_fill_secondary',
|
| 109 |
+
body_text_color='black',
|
| 110 |
+
background_fill_secondary='#f8f8f8',
|
| 111 |
+
border_color_accent='*primary_50',
|
| 112 |
+
border_color_primary='#ededed',
|
| 113 |
+
color_accent='#7367f0',
|
| 114 |
+
color_accent_soft='#fcfcfc',
|
| 115 |
+
panel_background_fill='#fcfcfc',
|
| 116 |
+
section_header_text_weight='600',
|
| 117 |
+
checkbox_background_color='*background_fill_secondary',
|
| 118 |
+
input_background_fill='white',
|
| 119 |
+
input_placeholder_color='*neutral_300',
|
| 120 |
+
loader_color = '#7367f0',
|
| 121 |
+
slider_color='#7367f0',
|
| 122 |
+
table_odd_background_fill='*neutral_100',
|
| 123 |
+
button_small_radius='*radius_sm',
|
| 124 |
+
button_primary_background_fill='linear-gradient(to bottom right, #7367f0, #9c93f4)',
|
| 125 |
+
button_primary_background_fill_hover='linear-gradient(to bottom right, #9c93f4, #9c93f4)',
|
| 126 |
+
button_primary_background_fill_hover_dark='linear-gradient(to bottom right, #5e50ee, #5e50ee)',
|
| 127 |
+
button_cancel_background_fill='linear-gradient(to bottom right, #fc0379, #ff88ac)',
|
| 128 |
+
button_cancel_background_fill_dark='linear-gradient(to bottom right, #dc2626, #b91c1c)',
|
| 129 |
+
button_cancel_background_fill_hover='linear-gradient(to bottom right, #f592c9, #f592c9)',
|
| 130 |
+
button_cancel_background_fill_hover_dark='linear-gradient(to bottom right, #dc2626, #dc2626)',
|
| 131 |
+
button_primary_border_color='#5949ed',
|
| 132 |
+
button_primary_text_color='white',
|
| 133 |
+
button_cancel_text_color='white',
|
| 134 |
+
button_cancel_text_color_dark='#dc2626',
|
| 135 |
+
button_cancel_border_color='#f04668',
|
| 136 |
+
button_cancel_border_color_dark='#dc2626',
|
| 137 |
+
button_cancel_border_color_hover='#fe6565',
|
| 138 |
+
button_cancel_border_color_hover_dark='#dc2626',
|
| 139 |
+
form_gap_width='1px',
|
| 140 |
+
layout_gap='5px'
|
| 141 |
+
)
|
| 142 |
+
|
| 143 |
+
|
| 144 |
+
def select_scheduler(pipe, selected_sampler):
|
| 145 |
+
import diffusers
|
| 146 |
+
|
| 147 |
+
scheduler_class_name, add_kwargs = SAMPLERS[selected_sampler]
|
| 148 |
+
config = pipe.scheduler.config
|
| 149 |
+
scheduler = getattr(diffusers, scheduler_class_name)
|
| 150 |
+
if selected_sampler in ("LCM", "LCM trailing"):
|
| 151 |
+
config = {
|
| 152 |
+
x: config[x] for x in config if x not in ("skip_prk_steps", "interpolation_type", "use_karras_sigmas")
|
| 153 |
+
}
|
| 154 |
+
elif selected_sampler in ("TCD", "TCD trailing"):
|
| 155 |
+
config = {x: config[x] for x in config if x not in ("skip_prk_steps")}
|
| 156 |
+
|
| 157 |
+
return scheduler.from_config(config, **add_kwargs)
|
| 158 |
+
|
| 159 |
+
|
| 160 |
+
def calculate_overlap(width, height, base_overlap=128):
|
| 161 |
+
"""
|
| 162 |
+
Calculates dynamic overlap based on the image's aspect ratio.
|
| 163 |
+
|
| 164 |
+
Args:
|
| 165 |
+
width (int): Width of the image in pixels.
|
| 166 |
+
height (int): Height of the image in pixels.
|
| 167 |
+
base_overlap (int, optional): Base overlap value in pixels. Defaults to 128.
|
| 168 |
+
|
| 169 |
+
Returns:
|
| 170 |
+
tuple: A tuple containing:
|
| 171 |
+
- row_overlap (int): Overlap between tiles in consecutive rows.
|
| 172 |
+
- col_overlap (int): Overlap between tiles in consecutive columns.
|
| 173 |
+
"""
|
| 174 |
+
ratio = height / width
|
| 175 |
+
if ratio < 1: # Image is wider than tall
|
| 176 |
+
return base_overlap // 2, base_overlap
|
| 177 |
+
else: # Image is taller than wide
|
| 178 |
+
return base_overlap, base_overlap * 2
|
| 179 |
+
|
| 180 |
+
|
| 181 |
+
# def calculate_overlap(width, height, base_overlap=128, scale=4):
|
| 182 |
+
# """
|
| 183 |
+
# Calculates dynamic overlap based on the image's aspect ratio and resolution.
|
| 184 |
+
# For scales less than 4, the overlap is fixed at 64, 128 (or 128, 256).
|
| 185 |
+
# For scales 4 or greater, the overlap is adjusted proportionally to the scale.
|
| 186 |
+
|
| 187 |
+
# Args:
|
| 188 |
+
# width (int): Width of the image in pixels.
|
| 189 |
+
# height (int): Height of the image in pixels.
|
| 190 |
+
# base_overlap (int, optional): Base overlap value in pixels. Defaults to 128.
|
| 191 |
+
# scale (int, optional): Scale factor for calculating the overlap. Defaults to 4.
|
| 192 |
+
|
| 193 |
+
# Returns:
|
| 194 |
+
# tuple: A tuple containing:
|
| 195 |
+
# - row_overlap (int): Overlap between tiles in consecutive rows.
|
| 196 |
+
# - col_overlap (int): Overlap between tiles in consecutive columns.
|
| 197 |
+
# """
|
| 198 |
+
# # Define the base scale (4)
|
| 199 |
+
# base_scale = 4
|
| 200 |
+
|
| 201 |
+
# # If scale is less than 4, use fixed overlap values
|
| 202 |
+
# if scale < base_scale:
|
| 203 |
+
# ratio = height / width
|
| 204 |
+
# if ratio < 1: # Image is wider than tall
|
| 205 |
+
# return base_overlap // 2, base_overlap
|
| 206 |
+
# else: # Image is taller than wide
|
| 207 |
+
# return base_overlap, base_overlap * 2
|
| 208 |
+
# else:
|
| 209 |
+
# # For scales 4 or greater, adjust overlap proportionally
|
| 210 |
+
# scaling_factor = scale / base_scale
|
| 211 |
+
# base_overlap = int(base_overlap * base_scale)
|
| 212 |
+
# #base_overlap = int(base_overlap * scaling_factor)
|
| 213 |
+
|
| 214 |
+
# ratio = height / width
|
| 215 |
+
# if ratio < 1: # Image is wider than tall
|
| 216 |
+
# return base_overlap // 2, base_overlap
|
| 217 |
+
# else: # Image is taller than wide
|
| 218 |
+
# return base_overlap, base_overlap * 2
|
| 219 |
+
|
| 220 |
+
|
| 221 |
+
# This function was copied and adapted from https://huggingface.co/spaces/gokaygokay/TileUpscalerV2, licensed under Apache 2.0.
|
| 222 |
+
def progressive_upscale(input_image, target_resolution, steps=3):
|
| 223 |
+
"""
|
| 224 |
+
Progressively upscales an image to the target resolution in multiple steps.
|
| 225 |
+
|
| 226 |
+
Args:
|
| 227 |
+
input_image (PIL.Image.Image): The input image to be upscaled.
|
| 228 |
+
target_resolution (int): The target resolution (width or height) in pixels.
|
| 229 |
+
steps (int, optional): The number of upscaling steps. Defaults to 3.
|
| 230 |
+
|
| 231 |
+
Returns:
|
| 232 |
+
PIL.Image.Image: The upscaled image at the target resolution.
|
| 233 |
+
"""
|
| 234 |
+
current_image = input_image.convert("RGB")
|
| 235 |
+
current_size = max(current_image.size)
|
| 236 |
+
|
| 237 |
+
# Upscale in multiple steps
|
| 238 |
+
for _ in range(steps):
|
| 239 |
+
if current_size >= target_resolution:
|
| 240 |
+
break
|
| 241 |
+
scale_factor = min(2, target_resolution / current_size)
|
| 242 |
+
new_size = (int(current_image.width * scale_factor), int(current_image.height * scale_factor))
|
| 243 |
+
current_image = current_image.resize(new_size, Image.LANCZOS)
|
| 244 |
+
current_size = max(current_image.size)
|
| 245 |
+
|
| 246 |
+
# Final resize to exact target resolution
|
| 247 |
+
if current_size != target_resolution:
|
| 248 |
+
aspect_ratio = current_image.width / current_image.height
|
| 249 |
+
if current_image.width > current_image.height:
|
| 250 |
+
new_size = (target_resolution, int(target_resolution / aspect_ratio))
|
| 251 |
+
else:
|
| 252 |
+
new_size = (int(target_resolution * aspect_ratio), target_resolution)
|
| 253 |
+
current_image = current_image.resize(new_size, Image.LANCZOS)
|
| 254 |
+
|
| 255 |
+
return current_image
|
| 256 |
+
|
| 257 |
+
|
| 258 |
+
# This function was copied and adapted from https://huggingface.co/spaces/gokaygokay/TileUpscalerV2, licensed under Apache 2.0.
|
| 259 |
+
def create_hdr_effect(original_image, hdr):
|
| 260 |
+
"""
|
| 261 |
+
Applies an HDR (High Dynamic Range) effect to an image based on the specified intensity.
|
| 262 |
+
|
| 263 |
+
Args:
|
| 264 |
+
original_image (PIL.Image.Image): The original image to which the HDR effect will be applied.
|
| 265 |
+
hdr (float): The intensity of the HDR effect, ranging from 0 (no effect) to 1 (maximum effect).
|
| 266 |
+
|
| 267 |
+
Returns:
|
| 268 |
+
PIL.Image.Image: The image with the HDR effect applied.
|
| 269 |
+
"""
|
| 270 |
+
if hdr == 0:
|
| 271 |
+
return original_image # No effect applied if hdr is 0
|
| 272 |
+
|
| 273 |
+
# Convert the PIL image to a NumPy array in BGR format (OpenCV format)
|
| 274 |
+
cv_original = cv2.cvtColor(np.array(original_image), cv2.COLOR_RGB2BGR)
|
| 275 |
+
|
| 276 |
+
# Define scaling factors for creating multiple exposures
|
| 277 |
+
factors = [
|
| 278 |
+
1.0 - 0.9 * hdr,
|
| 279 |
+
1.0 - 0.7 * hdr,
|
| 280 |
+
1.0 - 0.45 * hdr,
|
| 281 |
+
1.0 - 0.25 * hdr,
|
| 282 |
+
1.0,
|
| 283 |
+
1.0 + 0.2 * hdr,
|
| 284 |
+
1.0 + 0.4 * hdr,
|
| 285 |
+
1.0 + 0.6 * hdr,
|
| 286 |
+
1.0 + 0.8 * hdr,
|
| 287 |
+
]
|
| 288 |
+
|
| 289 |
+
# Generate multiple exposure images by scaling the original image
|
| 290 |
+
images = [cv2.convertScaleAbs(cv_original, alpha=factor) for factor in factors]
|
| 291 |
+
|
| 292 |
+
# Merge the images using the Mertens algorithm to create an HDR effect
|
| 293 |
+
merge_mertens = cv2.createMergeMertens()
|
| 294 |
+
hdr_image = merge_mertens.process(images)
|
| 295 |
+
|
| 296 |
+
# Convert the HDR image to 8-bit format (0-255 range)
|
| 297 |
+
hdr_image_8bit = np.clip(hdr_image * 255, 0, 255).astype("uint8")
|
| 298 |
+
|
| 299 |
+
# Convert the image back to RGB format and return as a PIL image
|
| 300 |
+
return Image.fromarray(cv2.cvtColor(hdr_image_8bit, cv2.COLOR_BGR2RGB))
|
| 301 |
+
|
| 302 |
+
|
| 303 |
+
def torch_gc():
|
| 304 |
+
if torch.cuda.is_available():
|
| 305 |
+
with torch.cuda.device("cuda"):
|
| 306 |
+
torch.cuda.empty_cache()
|
| 307 |
+
torch.cuda.ipc_collect()
|
| 308 |
+
|
| 309 |
+
gc.collect()
|
| 310 |
+
|
| 311 |
+
|
| 312 |
+
def quantize_8bit(unet):
|
| 313 |
+
if unet is None:
|
| 314 |
+
return
|
| 315 |
+
|
| 316 |
+
from peft.tuners.tuners_utils import BaseTunerLayer
|
| 317 |
+
|
| 318 |
+
dtype = unet.dtype
|
| 319 |
+
unet.to(torch.float8_e4m3fn)
|
| 320 |
+
for module in unet.modules(): # revert lora modules to prevent errors with fp8
|
| 321 |
+
if isinstance(module, BaseTunerLayer):
|
| 322 |
+
module.to(dtype)
|
| 323 |
+
|
| 324 |
+
if hasattr(unet, "encoder_hid_proj"): # revert ip adapter modules to prevent errors with fp8
|
| 325 |
+
if unet.encoder_hid_proj is not None:
|
| 326 |
+
for module in unet.encoder_hid_proj.modules():
|
| 327 |
+
module.to(dtype)
|
| 328 |
+
torch_gc()
|
requirements.txt
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
torch
|
| 2 |
+
peft
|
| 3 |
+
opencv-python
|
| 4 |
+
spaces
|
| 5 |
+
scipy
|
| 6 |
+
gradio==5.15.0
|
| 7 |
+
numpy==1.26.4
|
| 8 |
+
transformers
|
| 9 |
+
accelerate
|
| 10 |
+
diffusers
|
| 11 |
+
fastapi>=0.115.2
|