Update README.md
Browse files
README.md
CHANGED
|
@@ -64,19 +64,147 @@ image = pipe(prompt=prompt, num_inference_steps=4, guidance_scale=0).images[0]
|
|
| 64 |
|
| 65 |
### Image-to-Image
|
| 66 |
|
| 67 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 68 |
|
| 69 |
### Inpainting
|
| 70 |
|
| 71 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 72 |
|
| 73 |
### ControlNet
|
| 74 |
|
| 75 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 76 |
|
| 77 |
-
|
| 78 |
|
| 79 |
-
Works as well! TODO docs
|
| 80 |
|
| 81 |
## Speed Benchmark
|
| 82 |
|
|
|
|
| 64 |
|
| 65 |
### Image-to-Image
|
| 66 |
|
| 67 |
+
LCM-LoRA can be applied to image-to-image tasks too. Let's look at how we can perform image-to-image generation with LCMs. For this example we'll use the [dreamshaper-7](https://huggingface.co/Lykon/dreamshaper-7) model and the LCM-LoRA for `stable-diffusion-v1-5 `.
|
| 68 |
+
|
| 69 |
+
```python
|
| 70 |
+
import torch
|
| 71 |
+
from diffusers import AutoPipelineForImage2Image, LCMScheduler
|
| 72 |
+
from diffusers.utils import make_image_grid, load_image
|
| 73 |
+
|
| 74 |
+
pipe = AutoPipelineForImage2Image.from_pretrained(
|
| 75 |
+
"Lykon/dreamshaper-7",
|
| 76 |
+
torch_dtype=torch.float16,
|
| 77 |
+
variant="fp16",
|
| 78 |
+
).to("cuda")
|
| 79 |
+
|
| 80 |
+
# set scheduler
|
| 81 |
+
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
|
| 82 |
+
|
| 83 |
+
# load LCM-LoRA
|
| 84 |
+
pipe.load_lora_weights("latent-consistency/lcm-lora-sdv1-5")
|
| 85 |
+
pipe.fuse_lora()
|
| 86 |
+
|
| 87 |
+
# prepare image
|
| 88 |
+
url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/img2img-init.png"
|
| 89 |
+
init_image = load_image(url)
|
| 90 |
+
prompt = "Astronauts in a jungle, cold color palette, muted colors, detailed, 8k"
|
| 91 |
+
|
| 92 |
+
# pass prompt and image to pipeline
|
| 93 |
+
generator = torch.manual_seed(0)
|
| 94 |
+
image = pipe(
|
| 95 |
+
prompt,
|
| 96 |
+
image=init_image,
|
| 97 |
+
num_inference_steps=4,
|
| 98 |
+
guidance_scale=1,
|
| 99 |
+
strength=0.6,
|
| 100 |
+
generator=generator
|
| 101 |
+
).images[0]
|
| 102 |
+
make_image_grid([init_image, image], rows=1, cols=2)
|
| 103 |
+
```
|
| 104 |
+
|
| 105 |
+

|
| 106 |
+
|
| 107 |
|
| 108 |
### Inpainting
|
| 109 |
|
| 110 |
+
LCM-LoRA can be used for inpainting as well.
|
| 111 |
+
|
| 112 |
+
```python
|
| 113 |
+
import torch
|
| 114 |
+
from diffusers import AutoPipelineForInpainting, LCMScheduler
|
| 115 |
+
from diffusers.utils import load_image, make_image_grid
|
| 116 |
+
|
| 117 |
+
pipe = AutoPipelineForInpainting.from_pretrained(
|
| 118 |
+
"runwayml/stable-diffusion-inpainting",
|
| 119 |
+
torch_dtype=torch.float16,
|
| 120 |
+
variant="fp16",
|
| 121 |
+
).to("cuda")
|
| 122 |
+
|
| 123 |
+
# set scheduler
|
| 124 |
+
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
|
| 125 |
+
|
| 126 |
+
# load LCM-LoRA
|
| 127 |
+
pipe.load_lora_weights("latent-consistency/lcm-lora-sdv1-5")
|
| 128 |
+
pipe.fuse_lora()
|
| 129 |
+
|
| 130 |
+
# load base and mask image
|
| 131 |
+
init_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/inpaint.png")
|
| 132 |
+
mask_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/inpaint_mask.png")
|
| 133 |
+
|
| 134 |
+
# generator = torch.Generator("cuda").manual_seed(92)
|
| 135 |
+
prompt = "concept art digital painting of an elven castle, inspired by lord of the rings, highly detailed, 8k"
|
| 136 |
+
generator = torch.manual_seed(0)
|
| 137 |
+
image = pipe(
|
| 138 |
+
prompt=prompt,
|
| 139 |
+
image=init_image,
|
| 140 |
+
mask_image=mask_image,
|
| 141 |
+
generator=generator,
|
| 142 |
+
num_inference_steps=4,
|
| 143 |
+
guidance_scale=4,
|
| 144 |
+
).images[0]
|
| 145 |
+
make_image_grid([init_image, mask_image, image], rows=1, cols=3)
|
| 146 |
+
```
|
| 147 |
+
|
| 148 |
+

|
| 149 |
+
|
| 150 |
|
| 151 |
### ControlNet
|
| 152 |
|
| 153 |
+
For this example, we'll use the SD-v1-5 model and the LCM-LoRA for SD-v1-5 with canny ControlNet.
|
| 154 |
+
|
| 155 |
+
```python
|
| 156 |
+
import torch
|
| 157 |
+
import cv2
|
| 158 |
+
import numpy as np
|
| 159 |
+
from PIL import Image
|
| 160 |
+
|
| 161 |
+
from diffusers import StableDiffusionControlNetPipeline, ControlNetModel, LCMScheduler
|
| 162 |
+
from diffusers.utils import load_image
|
| 163 |
+
|
| 164 |
+
image = load_image(
|
| 165 |
+
"https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png"
|
| 166 |
+
).resize((512, 512))
|
| 167 |
+
|
| 168 |
+
image = np.array(image)
|
| 169 |
+
|
| 170 |
+
low_threshold = 100
|
| 171 |
+
high_threshold = 200
|
| 172 |
+
|
| 173 |
+
image = cv2.Canny(image, low_threshold, high_threshold)
|
| 174 |
+
image = image[:, :, None]
|
| 175 |
+
image = np.concatenate([image, image, image], axis=2)
|
| 176 |
+
canny_image = Image.fromarray(image)
|
| 177 |
+
|
| 178 |
+
controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
|
| 179 |
+
pipe = StableDiffusionControlNetPipeline.from_pretrained(
|
| 180 |
+
"runwayml/stable-diffusion-v1-5",
|
| 181 |
+
controlnet=controlnet,
|
| 182 |
+
torch_dtype=torch.float16,
|
| 183 |
+
safety_checker=None,
|
| 184 |
+
variant="fp16"
|
| 185 |
+
).to("cuda")
|
| 186 |
+
|
| 187 |
+
# set scheduler
|
| 188 |
+
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
|
| 189 |
+
|
| 190 |
+
# load LCM-LoRA
|
| 191 |
+
pipe.load_lora_weights("latent-consistency/lcm-lora-sdv1-5")
|
| 192 |
+
|
| 193 |
+
generator = torch.manual_seed(0)
|
| 194 |
+
image = pipe(
|
| 195 |
+
"the mona lisa",
|
| 196 |
+
image=canny_image,
|
| 197 |
+
num_inference_steps=4,
|
| 198 |
+
guidance_scale=1.5,
|
| 199 |
+
controlnet_conditioning_scale=0.8,
|
| 200 |
+
cross_attention_kwargs={"scale": 1},
|
| 201 |
+
generator=generator,
|
| 202 |
+
).images[0]
|
| 203 |
+
make_image_grid([canny_image, image], rows=1, cols=2)
|
| 204 |
+
```
|
| 205 |
|
| 206 |
+

|
| 207 |
|
|
|
|
| 208 |
|
| 209 |
## Speed Benchmark
|
| 210 |
|