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 |
+
![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/lcm/lcm_sdv1-5_i2i.png)
|
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 |
+
![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/lcm/lcm_sdv1-5_inpainting.png)
|
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 |
+
![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/lcm/lcm_sdv1-5_controlnet.png)
|
207 |
|
|
|
208 |
|
209 |
## Speed Benchmark
|
210 |
|