Spaces:
Running
on
Zero
Running
on
Zero
shuanholmes
commited on
Commit
·
bf00c4c
1
Parent(s):
d429710
[FireFlow] Init Commit
Browse files- app.py +84 -74
- flux/modules/layers.py +38 -12
- flux/sampling.py +19 -15
app.py
CHANGED
|
@@ -45,24 +45,26 @@ def encode(init_image, torch_device):
|
|
| 45 |
init_image = ae.encode(init_image.to()).to(torch.bfloat16)
|
| 46 |
return init_image
|
| 47 |
|
| 48 |
-
|
|
|
|
| 49 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 50 |
name = 'flux-dev'
|
| 51 |
-
ae = load_ae(name, device)
|
| 52 |
t5 = load_t5(device, max_length=256 if name == "flux-schnell" else 512)
|
| 53 |
clip = load_clip(device)
|
| 54 |
-
model = load_flow_model(name, device=
|
| 55 |
-
offload
|
| 56 |
-
|
|
|
|
|
|
|
| 57 |
is_schnell = False
|
| 58 |
-
feature_path = 'feature'
|
| 59 |
output_dir = 'result'
|
| 60 |
add_sampling_metadata = True
|
| 61 |
|
| 62 |
@spaces.GPU(duration=120)
|
| 63 |
@torch.inference_mode()
|
| 64 |
-
def edit(init_image, source_prompt, target_prompt, num_steps, inject_step, guidance, seed):
|
| 65 |
-
|
| 66 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 67 |
torch.cuda.empty_cache()
|
| 68 |
seed = None
|
|
@@ -76,15 +78,12 @@ def edit(init_image, source_prompt, target_prompt, num_steps, inject_step, guida
|
|
| 76 |
|
| 77 |
width, height = init_image.shape[0], init_image.shape[1]
|
| 78 |
|
| 79 |
-
|
| 80 |
init_image = torch.from_numpy(init_image).permute(2, 0, 1).float() / 127.5 - 1
|
| 81 |
init_image = init_image.unsqueeze(0)
|
| 82 |
init_image = init_image.to(device)
|
| 83 |
with torch.no_grad():
|
| 84 |
init_image = ae.encode(init_image.to()).to(torch.bfloat16)
|
| 85 |
|
| 86 |
-
print(init_image.shape)
|
| 87 |
-
|
| 88 |
rng = torch.Generator(device="cpu")
|
| 89 |
opts = SamplingOptions(
|
| 90 |
source_prompt=source_prompt,
|
|
@@ -97,6 +96,11 @@ def edit(init_image, source_prompt, target_prompt, num_steps, inject_step, guida
|
|
| 97 |
)
|
| 98 |
if opts.seed is None:
|
| 99 |
opts.seed = torch.Generator(device="cpu").seed()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 100 |
|
| 101 |
print(f"Generating with seed {opts.seed}:\n{opts.source_prompt}")
|
| 102 |
t0 = time.perf_counter()
|
|
@@ -106,12 +110,23 @@ def edit(init_image, source_prompt, target_prompt, num_steps, inject_step, guida
|
|
| 106 |
#############inverse#######################
|
| 107 |
info = {}
|
| 108 |
info['feature'] = {}
|
| 109 |
-
info['inject_step'] = inject_step
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 110 |
|
| 111 |
with torch.no_grad():
|
| 112 |
inp = prepare(t5, clip, init_image, prompt=opts.source_prompt)
|
| 113 |
inp_target = prepare(t5, clip, init_image, prompt=opts.target_prompt)
|
| 114 |
timesteps = get_schedule(opts.num_steps, inp["img"].shape[1], shift=(name != "flux-schnell"))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
|
| 116 |
# inversion initial noise
|
| 117 |
with torch.no_grad():
|
|
@@ -137,6 +152,11 @@ def edit(init_image, source_prompt, target_prompt, num_steps, inject_step, guida
|
|
| 137 |
idx = max(int(fn.split("_")[-1].split(".")[0]) for fn in fns) + 1
|
| 138 |
else:
|
| 139 |
idx = 0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 140 |
|
| 141 |
device = torch.device("cuda")
|
| 142 |
with torch.autocast(device_type=device.type, dtype=torch.bfloat16):
|
|
@@ -166,97 +186,87 @@ def edit(init_image, source_prompt, target_prompt, num_steps, inject_step, guida
|
|
| 166 |
return img
|
| 167 |
|
| 168 |
|
| 169 |
-
|
| 170 |
-
def create_demo(model_name: str, device: str = "cuda:0" if torch.cuda.is_available() else "cpu", offload: bool = False):
|
| 171 |
is_schnell = model_name == "flux-schnell"
|
| 172 |
title = r"""
|
| 173 |
-
<h1 align="center"
|
| 174 |
"""
|
| 175 |
-
|
| 176 |
description = r"""
|
| 177 |
-
<b>Official 🤗 Gradio
|
| 178 |
-
|
| 179 |
-
❗️❗️❗️[<b>Important</b>] Editing steps:<br>
|
| 180 |
-
1️⃣ Upload images you want to edit (The resolution is expected be less than 1360*768, or the memory of GPU may be not enough.) <br>
|
| 181 |
-
2️⃣ Enter the source prompt, which describes the content of the image you unload. The source prompt is not mandatory; you can also leave it to null. <br>
|
| 182 |
-
3️⃣ Enter the target prompt which describes the expected content of the edited image. <br>
|
| 183 |
-
4️⃣ Click the <b>Generate</b> button to start editing. <br>
|
| 184 |
-
5️⃣ We suggest to adjust the value of **feature sharing steps** for better results.<br>
|
| 185 |
-
"""
|
| 186 |
-
article = r"""
|
| 187 |
-
If our work is helpful, please help to ⭐ the <a href='https://github.com/wangjiangshan0725/RF-Solver-Edit' target='_blank'>Github Repo</a>. Thanks!
|
| 188 |
"""
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
[](https://github.com/wangjiangshan0725/RF-Solver-Edit)
|
| 192 |
"""
|
| 193 |
-
|
| 194 |
css = '''
|
| 195 |
.gradio-container {width: 85% !important}
|
| 196 |
'''
|
| 197 |
with gr.Blocks(css=css) as demo:
|
| 198 |
-
#
|
| 199 |
-
|
| 200 |
gr.HTML(title)
|
| 201 |
gr.Markdown(description)
|
| 202 |
gr.Markdown(article)
|
| 203 |
-
gr.Markdown(badge)
|
| 204 |
|
|
|
|
| 205 |
with gr.Row():
|
|
|
|
| 206 |
with gr.Column():
|
| 207 |
-
source_prompt = gr.Textbox(label="Source Prompt", value="")
|
| 208 |
-
target_prompt = gr.Textbox(label="Target Prompt", value="")
|
| 209 |
-
# source_prompt = gr.Text(
|
| 210 |
-
# label="Source Prompt",
|
| 211 |
-
# show_label=False,
|
| 212 |
-
# max_lines=1,
|
| 213 |
-
# placeholder="Enter your source prompt",
|
| 214 |
-
# container=False,
|
| 215 |
-
# value=""
|
| 216 |
-
# )
|
| 217 |
-
# target_prompt = gr.Text(
|
| 218 |
-
# label="Target Prompt",
|
| 219 |
-
# show_label=False,
|
| 220 |
-
# max_lines=1,
|
| 221 |
-
# placeholder="Enter your target prompt",
|
| 222 |
-
# container=False,
|
| 223 |
-
# value=""
|
| 224 |
-
# )
|
| 225 |
init_image = gr.Image(label="Input Image", visible=True)
|
| 226 |
-
|
| 227 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 228 |
generate_btn = gr.Button("Generate")
|
| 229 |
|
|
|
|
| 230 |
with gr.Column():
|
| 231 |
with gr.Accordion("Advanced Options", open=True):
|
| 232 |
-
num_steps = gr.Slider(
|
| 233 |
-
|
| 234 |
-
|
| 235 |
-
|
| 236 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 237 |
|
|
|
|
| 238 |
output_image = gr.Image(label="Generated Image")
|
| 239 |
|
|
|
|
| 240 |
generate_btn.click(
|
| 241 |
fn=edit,
|
| 242 |
-
inputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 243 |
outputs=[output_image]
|
| 244 |
)
|
| 245 |
-
|
| 246 |
-
|
| 247 |
return demo
|
| 248 |
|
| 249 |
-
|
| 250 |
-
# if __name__ == "__main__":
|
| 251 |
-
# import argparse
|
| 252 |
-
# parser = argparse.ArgumentParser(description="Flux")
|
| 253 |
-
# parser.add_argument("--name", type=str, default="flux-dev", choices=list(configs.keys()), help="Model name")
|
| 254 |
-
# parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu", help="Device to use")
|
| 255 |
-
# parser.add_argument("--offload", action="store_true", help="Offload model to CPU when not in use")
|
| 256 |
-
# parser.add_argument("--share", action="store_true", help="Create a public link to your demo")
|
| 257 |
-
|
| 258 |
-
# parser.add_argument("--port", type=int, default=41035)
|
| 259 |
-
# args = parser.parse_args()
|
| 260 |
-
|
| 261 |
demo = create_demo("flux-dev", "cuda")
|
| 262 |
demo.launch()
|
|
|
|
| 45 |
init_image = ae.encode(init_image.to()).to(torch.bfloat16)
|
| 46 |
return init_image
|
| 47 |
|
| 48 |
+
torch_device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 49 |
+
offload = True
|
| 50 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 51 |
name = 'flux-dev'
|
| 52 |
+
ae = load_ae(name, device="cpu" if offload else torch_device)
|
| 53 |
t5 = load_t5(device, max_length=256 if name == "flux-schnell" else 512)
|
| 54 |
clip = load_clip(device)
|
| 55 |
+
model = load_flow_model(name, device="cpu" if offload else torch_device)
|
| 56 |
+
if offload:
|
| 57 |
+
model.cpu()
|
| 58 |
+
torch.cuda.empty_cache()
|
| 59 |
+
ae.encoder.to(torch_device)
|
| 60 |
is_schnell = False
|
|
|
|
| 61 |
output_dir = 'result'
|
| 62 |
add_sampling_metadata = True
|
| 63 |
|
| 64 |
@spaces.GPU(duration=120)
|
| 65 |
@torch.inference_mode()
|
| 66 |
+
def edit(init_image, source_prompt, target_prompt, editing_strategy, num_steps, inject_step, guidance, seed):
|
| 67 |
+
global ae, t5, clip, model, name, is_schnell, output_dir, add_sampling_metadata
|
| 68 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 69 |
torch.cuda.empty_cache()
|
| 70 |
seed = None
|
|
|
|
| 78 |
|
| 79 |
width, height = init_image.shape[0], init_image.shape[1]
|
| 80 |
|
|
|
|
| 81 |
init_image = torch.from_numpy(init_image).permute(2, 0, 1).float() / 127.5 - 1
|
| 82 |
init_image = init_image.unsqueeze(0)
|
| 83 |
init_image = init_image.to(device)
|
| 84 |
with torch.no_grad():
|
| 85 |
init_image = ae.encode(init_image.to()).to(torch.bfloat16)
|
| 86 |
|
|
|
|
|
|
|
| 87 |
rng = torch.Generator(device="cpu")
|
| 88 |
opts = SamplingOptions(
|
| 89 |
source_prompt=source_prompt,
|
|
|
|
| 96 |
)
|
| 97 |
if opts.seed is None:
|
| 98 |
opts.seed = torch.Generator(device="cpu").seed()
|
| 99 |
+
|
| 100 |
+
if offload:
|
| 101 |
+
ae = ae.cpu()
|
| 102 |
+
torch.cuda.empty_cache()
|
| 103 |
+
t5, clip = t5.to(torch_device), clip.to(torch_device)
|
| 104 |
|
| 105 |
print(f"Generating with seed {opts.seed}:\n{opts.source_prompt}")
|
| 106 |
t0 = time.perf_counter()
|
|
|
|
| 110 |
#############inverse#######################
|
| 111 |
info = {}
|
| 112 |
info['feature'] = {}
|
| 113 |
+
info['inject_step'] = min(inject_step, num_steps)
|
| 114 |
+
info['reuse_v']= False
|
| 115 |
+
info['editing_strategy']= " ".join(editing_strategy)
|
| 116 |
+
info['start_layer_index'] = 20
|
| 117 |
+
info['end_layer_index'] = 37
|
| 118 |
+
qkv_ratio = '1.0,1.0,1.0'
|
| 119 |
+
info['qkv_ratio'] = list(map(float, qkv_ratio.split(',')))
|
| 120 |
|
| 121 |
with torch.no_grad():
|
| 122 |
inp = prepare(t5, clip, init_image, prompt=opts.source_prompt)
|
| 123 |
inp_target = prepare(t5, clip, init_image, prompt=opts.target_prompt)
|
| 124 |
timesteps = get_schedule(opts.num_steps, inp["img"].shape[1], shift=(name != "flux-schnell"))
|
| 125 |
+
|
| 126 |
+
if offload:
|
| 127 |
+
t5, clip = t5.cpu(), clip.cpu()
|
| 128 |
+
torch.cuda.empty_cache()
|
| 129 |
+
model = model.to(torch_device)
|
| 130 |
|
| 131 |
# inversion initial noise
|
| 132 |
with torch.no_grad():
|
|
|
|
| 152 |
idx = max(int(fn.split("_")[-1].split(".")[0]) for fn in fns) + 1
|
| 153 |
else:
|
| 154 |
idx = 0
|
| 155 |
+
|
| 156 |
+
if offload:
|
| 157 |
+
model.cpu()
|
| 158 |
+
torch.cuda.empty_cache()
|
| 159 |
+
ae.decoder.to(x.device)
|
| 160 |
|
| 161 |
device = torch.device("cuda")
|
| 162 |
with torch.autocast(device_type=device.type, dtype=torch.bfloat16):
|
|
|
|
| 186 |
return img
|
| 187 |
|
| 188 |
|
| 189 |
+
def create_demo(model_name: str, device: str = "cuda:0" if torch.cuda.is_available() else "cpu"):
|
|
|
|
| 190 |
is_schnell = model_name == "flux-schnell"
|
| 191 |
title = r"""
|
| 192 |
+
<h1 align="center">🔥FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing</h1>
|
| 193 |
"""
|
|
|
|
| 194 |
description = r"""
|
| 195 |
+
<b>Official 🤗 Gradio Demo</b> for <a href='https://github.com/HolmesShuan/FireFlow-Fast-Inversion-of-Rectified-Flow-for-Image-Semantic-Editing' target='_blank'><b>🔥FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing</b></a>.<br>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 196 |
"""
|
| 197 |
+
article = r"""
|
| 198 |
+
If you find our work helpful, we would greatly appreciate it if you could ⭐ our <a href='https://github.com/HolmesShuan/FireFlow-Fast-Inversion-of-Rectified-Flow-for-Image-Semantic-Editing' target='_blank'>GitHub repository</a>. Thank you for your support!
|
|
|
|
| 199 |
"""
|
|
|
|
| 200 |
css = '''
|
| 201 |
.gradio-container {width: 85% !important}
|
| 202 |
'''
|
| 203 |
with gr.Blocks(css=css) as demo:
|
| 204 |
+
# Add a title, description, and additional information
|
|
|
|
| 205 |
gr.HTML(title)
|
| 206 |
gr.Markdown(description)
|
| 207 |
gr.Markdown(article)
|
|
|
|
| 208 |
|
| 209 |
+
# Layout: Two columns
|
| 210 |
with gr.Row():
|
| 211 |
+
# Left Column: Inputs
|
| 212 |
with gr.Column():
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 213 |
init_image = gr.Image(label="Input Image", visible=True)
|
| 214 |
+
source_prompt = gr.Textbox(label="Source Prompt", value="", placeholder="(Optional) Describe the content of the uploaded image.")
|
| 215 |
+
target_prompt = gr.Textbox(label="Target Prompt", value="", placeholder="(Required) Describe the desired content of the edited image.")
|
| 216 |
+
# CheckboxGroup for editing strategies
|
| 217 |
+
editing_strategy = gr.CheckboxGroup(
|
| 218 |
+
label="Editing Technique",
|
| 219 |
+
choices=['replace_v', 'add_q', 'add_k'],
|
| 220 |
+
value=['replace_v'], # Default: none selected
|
| 221 |
+
interactive=True
|
| 222 |
+
)
|
| 223 |
generate_btn = gr.Button("Generate")
|
| 224 |
|
| 225 |
+
# Right Column: Advanced options and output
|
| 226 |
with gr.Column():
|
| 227 |
with gr.Accordion("Advanced Options", open=True):
|
| 228 |
+
num_steps = gr.Slider(
|
| 229 |
+
minimum=1,
|
| 230 |
+
maximum=30,
|
| 231 |
+
value=8,
|
| 232 |
+
step=1,
|
| 233 |
+
label="Total timesteps"
|
| 234 |
+
)
|
| 235 |
+
inject_step = gr.Slider(
|
| 236 |
+
minimum=1,
|
| 237 |
+
maximum=15,
|
| 238 |
+
value=1,
|
| 239 |
+
step=1,
|
| 240 |
+
label="Feature sharing steps"
|
| 241 |
+
)
|
| 242 |
+
guidance = gr.Slider(
|
| 243 |
+
minimum=1.0,
|
| 244 |
+
maximum=8.0,
|
| 245 |
+
value=2.0,
|
| 246 |
+
step=0.1,
|
| 247 |
+
label="Guidance",
|
| 248 |
+
interactive=not is_schnell
|
| 249 |
+
)
|
| 250 |
|
| 251 |
+
# Output display
|
| 252 |
output_image = gr.Image(label="Generated Image")
|
| 253 |
|
| 254 |
+
# Button click event to trigger the edit function
|
| 255 |
generate_btn.click(
|
| 256 |
fn=edit,
|
| 257 |
+
inputs=[
|
| 258 |
+
init_image,
|
| 259 |
+
source_prompt,
|
| 260 |
+
target_prompt,
|
| 261 |
+
editing_strategy, # Include the selected editing strategies
|
| 262 |
+
num_steps,
|
| 263 |
+
inject_step,
|
| 264 |
+
guidance
|
| 265 |
+
],
|
| 266 |
outputs=[output_image]
|
| 267 |
)
|
| 268 |
+
|
|
|
|
| 269 |
return demo
|
| 270 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 271 |
demo = create_demo("flux-dev", "cuda")
|
| 272 |
demo.launch()
|
flux/modules/layers.py
CHANGED
|
@@ -243,21 +243,47 @@ class SingleStreamBlock(nn.Module):
|
|
| 243 |
q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
|
| 244 |
q, k = self.norm(q, k, v)
|
| 245 |
|
| 246 |
-
# Note: If the memory of your device is not enough, you may consider uncomment the following code.
|
| 247 |
-
# if info['inject'] and info['id'] > 19:
|
| 248 |
-
# store_path = os.path.join(info['feature_path'], str(info['t']) + '_' + str(info['second_order']) + '_' + str(info['id']) + '_' + info['type'] + '_' + 'V' + '.pth')
|
| 249 |
-
# if info['inverse']:
|
| 250 |
-
# torch.save(v, store_path)
|
| 251 |
-
# if not info['inverse']:
|
| 252 |
-
# v = torch.load(store_path, weights_only=True)
|
| 253 |
-
|
| 254 |
# Save the features in the memory
|
| 255 |
-
if info['inject'] and info['id']
|
| 256 |
-
|
|
|
|
|
|
|
| 257 |
if info['inverse']:
|
| 258 |
-
info['
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 259 |
else:
|
| 260 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 261 |
|
| 262 |
# compute attention
|
| 263 |
attn = attention(q, k, v, pe=pe)
|
|
|
|
| 243 |
q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
|
| 244 |
q, k = self.norm(q, k, v)
|
| 245 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 246 |
# Save the features in the memory
|
| 247 |
+
if info['inject'] and info['id'] <= info['end_layer_index'] and info['id'] >= info['start_layer_index']:
|
| 248 |
+
v_feature_name = str(info['t']) + '_' + str(info['second_order']) + '_' + str(info['id']) + '_' + info['type'] + '_' + 'V'
|
| 249 |
+
k_feature_name = str(info['t']) + '_' + str(info['second_order']) + '_' + str(info['id']) + '_' + info['type'] + '_' + 'K'
|
| 250 |
+
q_feature_name = str(info['t']) + '_' + str(info['second_order']) + '_' + str(info['id']) + '_' + info['type'] + '_' + 'Q'
|
| 251 |
if info['inverse']:
|
| 252 |
+
if info['reuse_v']:
|
| 253 |
+
info['feature'][v_feature_name] = v.cpu()
|
| 254 |
+
else:
|
| 255 |
+
editing_strategy = info['editing_strategy']
|
| 256 |
+
qkv_ratio = info['qkv_ratio']
|
| 257 |
+
if 'q' in editing_strategy:
|
| 258 |
+
info['feature'][q_feature_name] = (q * qkv_ratio[0]).cpu()
|
| 259 |
+
if 'k' in editing_strategy:
|
| 260 |
+
info['feature'][k_feature_name] = (k * qkv_ratio[1]).cpu()
|
| 261 |
+
if 'v' in editing_strategy:
|
| 262 |
+
info['feature'][v_feature_name] = (v * qkv_ratio[2]).cpu()
|
| 263 |
else:
|
| 264 |
+
if info['reuse_v']:
|
| 265 |
+
if v_feature_name in info['feature']:
|
| 266 |
+
v = info['feature'][v_feature_name].cuda()
|
| 267 |
+
else:
|
| 268 |
+
editing_strategy = info['editing_strategy']
|
| 269 |
+
if 'replace_v' in editing_strategy:
|
| 270 |
+
if v_feature_name in info['feature']:
|
| 271 |
+
v = info['feature'][v_feature_name].cuda()
|
| 272 |
+
if 'add_v' in editing_strategy:
|
| 273 |
+
if v_feature_name in info['feature']:
|
| 274 |
+
v += info['feature'][v_feature_name].cuda()
|
| 275 |
+
if 'replace_k' in editing_strategy:
|
| 276 |
+
if k_feature_name in info['feature']:
|
| 277 |
+
k = info['feature'][k_feature_name].cuda()
|
| 278 |
+
if 'add_k' in editing_strategy:
|
| 279 |
+
if k_feature_name in info['feature']:
|
| 280 |
+
k += info['feature'][k_feature_name].cuda()
|
| 281 |
+
if 'replace_q' in editing_strategy:
|
| 282 |
+
if q_feature_name in info['feature']:
|
| 283 |
+
q = info['feature'][q_feature_name].cuda()
|
| 284 |
+
if 'add_q' in editing_strategy:
|
| 285 |
+
if q_feature_name in info['feature']:
|
| 286 |
+
q += info['feature'][q_feature_name].cuda()
|
| 287 |
|
| 288 |
# compute attention
|
| 289 |
attn = attention(q, k, v, pe=pe)
|
flux/sampling.py
CHANGED
|
@@ -97,6 +97,7 @@ def denoise(
|
|
| 97 |
guidance_vec = torch.full((img.shape[0],), guidance, device=img.device, dtype=img.dtype)
|
| 98 |
|
| 99 |
step_list = []
|
|
|
|
| 100 |
for i, (t_curr, t_prev) in enumerate(zip(timesteps[:-1], timesteps[1:])):
|
| 101 |
t_vec = torch.full((img.shape[0],), t_curr, dtype=img.dtype, device=img.device)
|
| 102 |
info['t'] = t_prev if inverse else t_curr
|
|
@@ -104,20 +105,23 @@ def denoise(
|
|
| 104 |
info['second_order'] = False
|
| 105 |
info['inject'] = inject_list[i]
|
| 106 |
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
|
|
|
|
|
|
|
|
|
| 118 |
img_mid = img + (t_prev - t_curr) / 2 * pred
|
| 119 |
|
| 120 |
-
t_vec_mid = torch.full((img.shape[0],),
|
| 121 |
info['second_order'] = True
|
| 122 |
pred_mid, info = model(
|
| 123 |
img=img_mid,
|
|
@@ -129,9 +133,9 @@ def denoise(
|
|
| 129 |
guidance=guidance_vec,
|
| 130 |
info=info
|
| 131 |
)
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
img = img + (t_prev - t_curr) *
|
| 135 |
|
| 136 |
return img, info
|
| 137 |
|
|
|
|
| 97 |
guidance_vec = torch.full((img.shape[0],), guidance, device=img.device, dtype=img.dtype)
|
| 98 |
|
| 99 |
step_list = []
|
| 100 |
+
next_step_velocity = None
|
| 101 |
for i, (t_curr, t_prev) in enumerate(zip(timesteps[:-1], timesteps[1:])):
|
| 102 |
t_vec = torch.full((img.shape[0],), t_curr, dtype=img.dtype, device=img.device)
|
| 103 |
info['t'] = t_prev if inverse else t_curr
|
|
|
|
| 105 |
info['second_order'] = False
|
| 106 |
info['inject'] = inject_list[i]
|
| 107 |
|
| 108 |
+
if next_step_velocity is None:
|
| 109 |
+
pred, info = model(
|
| 110 |
+
img=img,
|
| 111 |
+
img_ids=img_ids,
|
| 112 |
+
txt=txt,
|
| 113 |
+
txt_ids=txt_ids,
|
| 114 |
+
y=vec,
|
| 115 |
+
timesteps=t_vec,
|
| 116 |
+
guidance=guidance_vec,
|
| 117 |
+
info=info
|
| 118 |
+
)
|
| 119 |
+
else:
|
| 120 |
+
pred = next_step_velocity
|
| 121 |
+
|
| 122 |
img_mid = img + (t_prev - t_curr) / 2 * pred
|
| 123 |
|
| 124 |
+
t_vec_mid = torch.full((img.shape[0],), t_curr + (t_prev - t_curr) / 2, dtype=img.dtype, device=img.device)
|
| 125 |
info['second_order'] = True
|
| 126 |
pred_mid, info = model(
|
| 127 |
img=img_mid,
|
|
|
|
| 133 |
guidance=guidance_vec,
|
| 134 |
info=info
|
| 135 |
)
|
| 136 |
+
next_step_velocity = pred_mid
|
| 137 |
+
|
| 138 |
+
img = img + (t_prev - t_curr) * pred_mid
|
| 139 |
|
| 140 |
return img, info
|
| 141 |
|