Spaces:
Runtime error
Runtime error
File size: 10,633 Bytes
a24b16a 640a27b f71eb42 a24b16a 640a27b 5349660 843b14b 88c9af4 a24b16a f71eb42 a24b16a 88c9af4 f0cf9b0 5349660 a24b16a f0cf9b0 d53dfab 5349660 a24b16a 5349660 a24b16a 5349660 506478b a24b16a 5349660 f0cf9b0 5349660 f0cf9b0 5349660 f0cf9b0 5349660 0166058 5349660 0166058 5349660 0166058 843b14b 5349660 f0cf9b0 843b14b 5349660 843b14b f0cf9b0 5349660 843b14b f0cf9b0 843b14b 88c9af4 a24b16a f71eb42 843b14b a24b16a 843b14b a24b16a f0cf9b0 a24b16a 5349660 843b14b a24b16a f71eb42 a24b16a 5349660 a24b16a 5349660 a24b16a f71eb42 a24b16a 843b14b a24b16a 6a74362 a24b16a 6a74362 a24b16a 958100a e066869 a24b16a 843b14b 5349660 a24b16a 958100a a24b16a 958100a 843b14b 5349660 88c9af4 5349660 0166058 843b14b f71eb42 a24b16a 843b14b a24b16a 843b14b 88c9af4 f71eb42 843b14b f0cf9b0 843b14b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 |
from pathlib import Path
import gradio as gr
import torch
from finetuning import FineTunedModel
from StableDiffuser import StableDiffuser
from tqdm import tqdm
model_map = {
'Car' : 'models/car.pt',
'Van Gogh' : 'models/vangogh.pt',
}
class Demo:
def __init__(self) -> None:
self.training = False
self.generating = False
self.nsteps = 50
self.diffuser = StableDiffuser(scheduler='DDIM', seed=42).to('cuda')
self.finetuner = None
with gr.Blocks() as demo:
self.layout()
self.switch_model(self.model_dropdown.value)
self.finetuner = self.finetuner.eval().half()
self.diffuser = self.diffuser.eval().half()
demo.queue(concurrency_count=2).launch()
def disable(self):
return [gr.update(interactive=False), gr.update(interactive=False)]
def switch_model(self, model_name):
if not model_name:
return
model_path = model_map[model_name]
checkpoint = torch.load(model_path)
self.finetuner = FineTunedModel.from_checkpoint(self.diffuser, checkpoint)
torch.cuda.empty_cache()
def layout(self):
with gr.Row():
with gr.Tab("Test") as inference_column:
with gr.Row():
self.explain_infr = gr.Markdown(interactive=False,
value='This is a demo of [Erasing Concepts from Stable Diffusion](https://erasing.baulab.info/). To try out a model where a concept has been erased, select a model and enter any prompt. For example, if you select the model "Van Gogh" you can generate images for the prompt "A portrait in the style of Van Gogh" and compare the erased and unerased models. We have also provided models with "cars" erased, and with "nudity" erased. You can also train and run your own custom model with a concept erased.')
with gr.Row():
with gr.Column(scale=1):
self.prompt_input_infr = gr.Text(
placeholder="Enter prompt...",
label="Prompt",
info="Prompt to generate"
)
with gr.Row():
self.model_dropdown = gr.Dropdown(
label="ESD Model",
choices=['Van Gogh', 'Car'],
value='Van Gogh',
interactive=True
)
self.seed_infr = gr.Number(
label="Seed",
value=42
)
with gr.Column(scale=2):
self.infr_button = gr.Button(
value="Generate",
interactive=True
)
with gr.Row():
self.image_new = gr.Image(
label="ESD",
interactive=False
)
self.image_orig = gr.Image(
label="SD",
interactive=False
)
with gr.Tab("Train") as training_column:
with gr.Row():
self.explain_train= gr.Markdown(interactive=False,
value='In this part you can erase any concept from Stable Diffusion. Enter a prompt for the concept or style you want to erase, and select ESD-x if you want to focus erasure on prompts that mention the concept explicitly, or ESD-u if you want to erase the concept even for prompts that do not mention the concept. With default settings, it takes about 20 minutes to fine-tune the model; then you can try inference above or download the weights. The training code used here is slightly different than the code tested in the original paper. Code and details are at [github link](https://github.com/rohitgandikota/erasing).')
with gr.Row():
with gr.Column(scale=3):
self.prompt_input = gr.Text(
placeholder="Enter prompt...",
label="Prompt to Erase",
info="Prompt corresponding to concept to erase"
)
self.train_method_input = gr.Dropdown(
choices=['ESD-x', 'ESD-u', 'ESD-self'],
value='ESD-x',
label='Train Method',
info='Method of training'
)
self.neg_guidance_input = gr.Number(
value=1,
label="Negative Guidance",
info='Guidance of negative training used to train'
)
self.iterations_input = gr.Number(
value=150,
precision=0,
label="Iterations",
info='iterations used to train'
)
self.lr_input = gr.Number(
value=1e-5,
label="Learning Rate",
info='Learning rate used to train'
)
with gr.Column(scale=1):
self.train_button = gr.Button(
value="Train",
)
self.download = gr.Files()
self.model_dropdown.change(self.switch_model, inputs=[self.model_dropdown])
self.infr_button.click(self.inference, inputs = [
self.prompt_input_infr,
self.seed_infr
],
outputs=[
self.image_new,
self.image_orig
]
)
self.train_button.click(self.disable,
outputs=[self.train_button, self.infr_button]
)
self.train_button.click(self.train, inputs = [
self.prompt_input,
self.train_method_input,
self.neg_guidance_input,
self.iterations_input,
self.lr_input
],
outputs=[self.train_button, self.infr_button, self.download, self.model_dropdown]
)
def train(self, prompt, train_method, neg_guidance, iterations, lr, pbar = gr.Progress(track_tqdm=True)):
if self.training:
return [None, None, None]
else:
self.training = True
del self.finetuner
torch.cuda.empty_cache()
self.diffuser = self.diffuser.train().float()
if train_method == 'ESD-x':
modules = ".*attn2$"
frozen = []
elif train_method == 'ESD-u':
modules = "unet$"
frozen = [".*attn2$", "unet.time_embedding$", "unet.conv_out$"]
elif train_method == 'ESD-self':
modules = ".*attn1$"
frozen = []
finetuner = FineTunedModel(self.diffuser, modules, frozen_modules=frozen)
optimizer = torch.optim.Adam(finetuner.parameters(), lr=lr)
criteria = torch.nn.MSELoss()
pbar = tqdm(range(iterations))
with torch.no_grad():
neutral_text_embeddings = self.diffuser.get_text_embeddings([''],n_imgs=1)
positive_text_embeddings = self.diffuser.get_text_embeddings([prompt],n_imgs=1)
for i in pbar:
with torch.no_grad():
self.diffuser.set_scheduler_timesteps(self.nsteps)
optimizer.zero_grad()
iteration = torch.randint(1, self.nsteps - 1, (1,)).item()
latents = self.diffuser.get_initial_latents(1, 512, 1)
with finetuner:
latents_steps, _ = self.diffuser.diffusion(
latents,
positive_text_embeddings,
start_iteration=0,
end_iteration=iteration,
guidance_scale=3,
show_progress=False
)
self.diffuser.set_scheduler_timesteps(1000)
iteration = int(iteration / self.nsteps * 1000)
positive_latents = self.diffuser.predict_noise(iteration, latents_steps[0], positive_text_embeddings, guidance_scale=1)
neutral_latents = self.diffuser.predict_noise(iteration, latents_steps[0], neutral_text_embeddings, guidance_scale=1)
with finetuner:
negative_latents = self.diffuser.predict_noise(iteration, latents_steps[0], positive_text_embeddings, guidance_scale=1)
positive_latents.requires_grad = False
neutral_latents.requires_grad = False
loss = criteria(negative_latents, neutral_latents - (neg_guidance*(positive_latents - neutral_latents))) #loss = criteria(e_n, e_0) works the best try 5000 epochs
loss.backward()
optimizer.step()
ft_path = f"{prompt.lower().replace(' ', '')}.pt"
torch.save(finetuner.state_dict(), ft_path)
self.finetuner = finetuner.eval().half()
self.diffuser = self.diffuser.eval().half()
torch.cuda.empty_cache()
self.training = False
model_map['Custom'] = ft_path
return [gr.update(interactive=True), gr.update(interactive=True), ft_path, gr.Dropdown.update(choices=list(model_map.keys()), value='Custom')]
def inference(self, prompt, seed, pbar = gr.Progress(track_tqdm=True)):
if self.generating:
return [None, None]
else:
self.generating = True
self.diffuser._seed = seed or 42
images = self.diffuser(
prompt,
n_steps=50,
reseed=True
)
orig_image = images[0][0]
torch.cuda.empty_cache()
with self.finetuner:
images = self.diffuser(
prompt,
n_steps=50,
reseed=True
)
edited_image = images[0][0]
self.generating = False
torch.cuda.empty_cache()
return edited_image, orig_image
demo = Demo()
|