Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -19,9 +19,13 @@ global vae
|
|
| 19 |
global text_encoder
|
| 20 |
global tokenizer
|
| 21 |
global noise_scheduler
|
|
|
|
|
|
|
|
|
|
| 22 |
device = "cuda:0"
|
| 23 |
generator = torch.Generator(device=device)
|
| 24 |
|
|
|
|
| 25 |
models_path = snapshot_download(repo_id="Snapchat/w2w")
|
| 26 |
|
| 27 |
mean = torch.load(f"{models_path}/mean.pt").bfloat16().to(device)
|
|
@@ -30,6 +34,7 @@ v = torch.load(f"{models_path}/V.pt").bfloat16().to(device)
|
|
| 30 |
proj = torch.load(f"{models_path}/proj_1000pc.pt").bfloat16().to(device)
|
| 31 |
df = torch.load(f"{models_path}/identity_df.pt")
|
| 32 |
weight_dimensions = torch.load(f"{models_path}/weight_dimensions.pt")
|
|
|
|
| 33 |
|
| 34 |
unet, vae, text_encoder, tokenizer, noise_scheduler = load_models(device)
|
| 35 |
global network
|
|
@@ -38,16 +43,64 @@ def sample_model():
|
|
| 38 |
global unet
|
| 39 |
del unet
|
| 40 |
global network
|
|
|
|
| 41 |
unet, _, _, _, _ = load_models(device)
|
| 42 |
network = sample_weights(unet, proj, mean, std, v[:, :1000], device, factor = 1.00)
|
| 43 |
|
| 44 |
-
|
| 45 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
|
|
|
|
|
|
|
|
|
|
| 47 |
|
| 48 |
|
| 49 |
@torch.no_grad()
|
| 50 |
-
def
|
|
|
|
| 51 |
global device
|
| 52 |
global generator
|
| 53 |
global unet
|
|
@@ -55,6 +108,15 @@ def inference( prompt, negative_prompt, guidance_scale, ddim_steps, seed):
|
|
| 55 |
global text_encoder
|
| 56 |
global tokenizer
|
| 57 |
global noise_scheduler
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
generator = generator.manual_seed(seed)
|
| 59 |
latents = torch.randn(
|
| 60 |
(1, unet.in_channels, 512 // 8, 512 // 8),
|
|
@@ -76,11 +138,23 @@ def inference( prompt, negative_prompt, guidance_scale, ddim_steps, seed):
|
|
| 76 |
noise_scheduler.set_timesteps(ddim_steps)
|
| 77 |
latents = latents * noise_scheduler.init_noise_sigma
|
| 78 |
|
|
|
|
|
|
|
| 79 |
for i,t in enumerate(tqdm.tqdm(noise_scheduler.timesteps)):
|
| 80 |
latent_model_input = torch.cat([latents] * 2)
|
| 81 |
latent_model_input = noise_scheduler.scale_model_input(latent_model_input, timestep=t)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
with network:
|
| 83 |
noise_pred = unet(latent_model_input, t, encoder_hidden_states=text_embeddings, timestep_cond= None).sample
|
|
|
|
|
|
|
| 84 |
#guidance
|
| 85 |
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
| 86 |
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
|
@@ -89,13 +163,61 @@ def inference( prompt, negative_prompt, guidance_scale, ddim_steps, seed):
|
|
| 89 |
latents = 1 / 0.18215 * latents
|
| 90 |
image = vae.decode(latents).sample
|
| 91 |
image = (image / 2 + 0.5).clamp(0, 1)
|
|
|
|
| 92 |
image = image.detach().cpu().float().permute(0, 2, 3, 1).numpy()[0]
|
| 93 |
|
| 94 |
image = Image.fromarray((image * 255).round().astype("uint8"))
|
| 95 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 96 |
return [image]
|
| 97 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 98 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
|
| 100 |
css = ''
|
| 101 |
with gr.Blocks(css=css) as demo:
|
|
@@ -103,36 +225,41 @@ with gr.Blocks(css=css) as demo:
|
|
| 103 |
gr.Markdown("Demo for the [h94/IP-Adapter-FaceID model](https://huggingface.co/h94/IP-Adapter-FaceID) - Generate AI images with your own face - Non-commercial license")
|
| 104 |
with gr.Row():
|
| 105 |
with gr.Column():
|
| 106 |
-
files = gr.Files(
|
| 107 |
-
label="Upload a photo of your face to invert, or sample a new model",
|
| 108 |
-
file_types=["image"]
|
| 109 |
-
)
|
| 110 |
-
uploaded_files = gr.Gallery(label="Your images", visible=False, columns=5, rows=1, height=125)
|
| 111 |
-
|
| 112 |
sample = gr.Button("Sample New Model")
|
|
|
|
| 113 |
|
| 114 |
-
|
| 115 |
-
remove_and_reupload = gr.ClearButton(value="Remove and upload new ones", components=files, size="sm")
|
| 116 |
prompt = gr.Textbox(label="Prompt",
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
negative_prompt = gr.Textbox(label="Negative Prompt", placeholder="low quality, blurry, unfinished, cartoon", value="low quality, blurry, unfinished, cartoon")
|
| 121 |
-
seed = gr.Number(value=5, label="Seed", interactive=True)
|
| 122 |
cfg = gr.Slider(label="CFG", value=3.0, step=0.1, minimum=0, maximum=10, interactive=True)
|
| 123 |
-
steps = gr.Slider(label="Inference Steps", value=50, step=1, minimum=0, maximum=100, interactive=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 124 |
|
| 125 |
|
| 126 |
submit = gr.Button("Submit")
|
| 127 |
|
| 128 |
with gr.Column():
|
| 129 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 130 |
|
| 131 |
-
sample.click(fn=sample_model)
|
| 132 |
|
| 133 |
-
submit.click(fn=
|
| 134 |
-
inputs=[prompt, negative_prompt, cfg, steps, seed],
|
| 135 |
-
outputs=
|
| 136 |
|
| 137 |
|
| 138 |
|
|
@@ -140,3 +267,8 @@ with gr.Blocks(css=css) as demo:
|
|
| 140 |
|
| 141 |
|
| 142 |
demo.launch(share=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
global text_encoder
|
| 20 |
global tokenizer
|
| 21 |
global noise_scheduler
|
| 22 |
+
global young_val
|
| 23 |
+
global pointy_val
|
| 24 |
+
global bags_val
|
| 25 |
device = "cuda:0"
|
| 26 |
generator = torch.Generator(device=device)
|
| 27 |
|
| 28 |
+
|
| 29 |
models_path = snapshot_download(repo_id="Snapchat/w2w")
|
| 30 |
|
| 31 |
mean = torch.load(f"{models_path}/mean.pt").bfloat16().to(device)
|
|
|
|
| 34 |
proj = torch.load(f"{models_path}/proj_1000pc.pt").bfloat16().to(device)
|
| 35 |
df = torch.load(f"{models_path}/identity_df.pt")
|
| 36 |
weight_dimensions = torch.load(f"{models_path}/weight_dimensions.pt")
|
| 37 |
+
pinverse = torch.load(f"{models_path}/pinverse_1000pc.pt").bfloat16().to(device)
|
| 38 |
|
| 39 |
unet, vae, text_encoder, tokenizer, noise_scheduler = load_models(device)
|
| 40 |
global network
|
|
|
|
| 43 |
global unet
|
| 44 |
del unet
|
| 45 |
global network
|
| 46 |
+
|
| 47 |
unet, _, _, _, _ = load_models(device)
|
| 48 |
network = sample_weights(unet, proj, mean, std, v[:, :1000], device, factor = 1.00)
|
| 49 |
|
| 50 |
+
|
| 51 |
+
@torch.no_grad()
|
| 52 |
+
def inference( prompt, negative_prompt, guidance_scale, ddim_steps, seed):
|
| 53 |
+
global device
|
| 54 |
+
global generator
|
| 55 |
+
global unet
|
| 56 |
+
global vae
|
| 57 |
+
global text_encoder
|
| 58 |
+
global tokenizer
|
| 59 |
+
global noise_scheduler
|
| 60 |
+
generator = generator.manual_seed(seed)
|
| 61 |
+
latents = torch.randn(
|
| 62 |
+
(1, unet.in_channels, 512 // 8, 512 // 8),
|
| 63 |
+
generator = generator,
|
| 64 |
+
device = device
|
| 65 |
+
).bfloat16()
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
text_input = tokenizer(prompt, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt")
|
| 69 |
+
|
| 70 |
+
text_embeddings = text_encoder(text_input.input_ids.to(device))[0]
|
| 71 |
+
|
| 72 |
+
max_length = text_input.input_ids.shape[-1]
|
| 73 |
+
uncond_input = tokenizer(
|
| 74 |
+
[negative_prompt], padding="max_length", max_length=max_length, return_tensors="pt"
|
| 75 |
+
)
|
| 76 |
+
uncond_embeddings = text_encoder(uncond_input.input_ids.to(device))[0]
|
| 77 |
+
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
| 78 |
+
noise_scheduler.set_timesteps(ddim_steps)
|
| 79 |
+
latents = latents * noise_scheduler.init_noise_sigma
|
| 80 |
+
|
| 81 |
+
for i,t in enumerate(tqdm.tqdm(noise_scheduler.timesteps)):
|
| 82 |
+
latent_model_input = torch.cat([latents] * 2)
|
| 83 |
+
latent_model_input = noise_scheduler.scale_model_input(latent_model_input, timestep=t)
|
| 84 |
+
with network:
|
| 85 |
+
noise_pred = unet(latent_model_input, t, encoder_hidden_states=text_embeddings, timestep_cond= None).sample
|
| 86 |
+
#guidance
|
| 87 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
| 88 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
| 89 |
+
latents = noise_scheduler.step(noise_pred, t, latents).prev_sample
|
| 90 |
+
|
| 91 |
+
latents = 1 / 0.18215 * latents
|
| 92 |
+
image = vae.decode(latents).sample
|
| 93 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
| 94 |
+
image = image.detach().cpu().float().permute(0, 2, 3, 1).numpy()[0]
|
| 95 |
|
| 96 |
+
image = Image.fromarray((image * 255).round().astype("uint8"))
|
| 97 |
+
|
| 98 |
+
return [image]
|
| 99 |
|
| 100 |
|
| 101 |
@torch.no_grad()
|
| 102 |
+
def edit_inference(prompt, negative_prompt, guidance_scale, ddim_steps, seed, start_noise, a1, a2, a3):
|
| 103 |
+
|
| 104 |
global device
|
| 105 |
global generator
|
| 106 |
global unet
|
|
|
|
| 108 |
global text_encoder
|
| 109 |
global tokenizer
|
| 110 |
global noise_scheduler
|
| 111 |
+
global young
|
| 112 |
+
global pointy
|
| 113 |
+
global bags
|
| 114 |
+
|
| 115 |
+
original_weights = network.proj.clone()
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
edited_weights = original_weights+a1*young+a2*pointy+a3*bags
|
| 119 |
+
|
| 120 |
generator = generator.manual_seed(seed)
|
| 121 |
latents = torch.randn(
|
| 122 |
(1, unet.in_channels, 512 // 8, 512 // 8),
|
|
|
|
| 138 |
noise_scheduler.set_timesteps(ddim_steps)
|
| 139 |
latents = latents * noise_scheduler.init_noise_sigma
|
| 140 |
|
| 141 |
+
|
| 142 |
+
|
| 143 |
for i,t in enumerate(tqdm.tqdm(noise_scheduler.timesteps)):
|
| 144 |
latent_model_input = torch.cat([latents] * 2)
|
| 145 |
latent_model_input = noise_scheduler.scale_model_input(latent_model_input, timestep=t)
|
| 146 |
+
|
| 147 |
+
if t>start_noise:
|
| 148 |
+
pass
|
| 149 |
+
elif t<=start_noise:
|
| 150 |
+
network.proj = torch.nn.Parameter(edited_weights)
|
| 151 |
+
network.reset()
|
| 152 |
+
|
| 153 |
+
|
| 154 |
with network:
|
| 155 |
noise_pred = unet(latent_model_input, t, encoder_hidden_states=text_embeddings, timestep_cond= None).sample
|
| 156 |
+
|
| 157 |
+
|
| 158 |
#guidance
|
| 159 |
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
| 160 |
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
|
|
|
| 163 |
latents = 1 / 0.18215 * latents
|
| 164 |
image = vae.decode(latents).sample
|
| 165 |
image = (image / 2 + 0.5).clamp(0, 1)
|
| 166 |
+
|
| 167 |
image = image.detach().cpu().float().permute(0, 2, 3, 1).numpy()[0]
|
| 168 |
|
| 169 |
image = Image.fromarray((image * 255).round().astype("uint8"))
|
| 170 |
|
| 171 |
+
#reset weights back to original
|
| 172 |
+
network.proj = torch.nn.Parameter(original_weights)
|
| 173 |
+
network.reset()
|
| 174 |
+
|
| 175 |
return [image]
|
| 176 |
|
| 177 |
+
|
| 178 |
+
|
| 179 |
+
|
| 180 |
+
def sample_then_run():
|
| 181 |
+
global young_val
|
| 182 |
+
global pointy_val
|
| 183 |
+
global bags_val
|
| 184 |
+
global young
|
| 185 |
+
global pointy
|
| 186 |
+
global bags
|
| 187 |
+
|
| 188 |
+
sample_model()
|
| 189 |
+
|
| 190 |
+
young_val = network.proj@young[0]/(torch.norm(young)**2).item()
|
| 191 |
+
pointy_val = network.proj@pointy[0]/(torch.norm(pointy)**2).item()
|
| 192 |
+
bags_val = network.proj@bags[0]/(torch.norm(bags)**2).item()
|
| 193 |
|
| 194 |
+
prompt = "sks person"
|
| 195 |
+
negative_prompt = "low quality, blurry, unfinished, cartoon"
|
| 196 |
+
seed = 5
|
| 197 |
+
cfg = 3.0
|
| 198 |
+
steps = 50
|
| 199 |
+
image = inference( prompt, negative_prompt, cfg, steps, seed)
|
| 200 |
+
return image
|
| 201 |
+
|
| 202 |
+
|
| 203 |
+
#directions
|
| 204 |
+
global young
|
| 205 |
+
global pointy
|
| 206 |
+
global bags
|
| 207 |
+
young = get_direction(df, "Young", pinverse, 1000, device)
|
| 208 |
+
young = debias(young, "Male", df, pinverse, device)
|
| 209 |
+
young_max = torch.max(proj@young[0]/(torch.norm(young))**2).item()
|
| 210 |
+
young_min = torch.min(proj@young[0]/(torch.norm(young))**2).item()
|
| 211 |
+
|
| 212 |
+
pointy = get_direction(df, "Pointy_Nose", pinverse, 1000, device)
|
| 213 |
+
pointy_max = torch.max(proj@pointy[0]/(torch.norm(pointy))**2).item()
|
| 214 |
+
pointy_min = torch.min(proj@pointy[0]/(torch.norm(pointy))**2).item()
|
| 215 |
+
|
| 216 |
+
bags = get_direction(df, "Bags_Under_Eyes", pinverse, 1000, device)
|
| 217 |
+
bags_max = torch.max(proj@bags[0]/(torch.norm(bags))**2).item()
|
| 218 |
+
bags_min = torch.min(proj@bags[0]/(torch.norm(bags))**2).item()
|
| 219 |
+
|
| 220 |
+
|
| 221 |
|
| 222 |
css = ''
|
| 223 |
with gr.Blocks(css=css) as demo:
|
|
|
|
| 225 |
gr.Markdown("Demo for the [h94/IP-Adapter-FaceID model](https://huggingface.co/h94/IP-Adapter-FaceID) - Generate AI images with your own face - Non-commercial license")
|
| 226 |
with gr.Row():
|
| 227 |
with gr.Column():
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 228 |
sample = gr.Button("Sample New Model")
|
| 229 |
+
gallery1 = gr.Gallery(label="Identity from Sampled Model")
|
| 230 |
|
| 231 |
+
with gr.Column():
|
|
|
|
| 232 |
prompt = gr.Textbox(label="Prompt",
|
| 233 |
+
info="Make sure to include 'sks person'" ,
|
| 234 |
+
placeholder="sks person",
|
| 235 |
+
value="sks person")
|
| 236 |
negative_prompt = gr.Textbox(label="Negative Prompt", placeholder="low quality, blurry, unfinished, cartoon", value="low quality, blurry, unfinished, cartoon")
|
| 237 |
+
seed = gr.Number(value=5, precision=0, label="Seed", interactive=True)
|
| 238 |
cfg = gr.Slider(label="CFG", value=3.0, step=0.1, minimum=0, maximum=10, interactive=True)
|
| 239 |
+
steps = gr.Slider(label="Inference Steps", precision=0, value=50, step=1, minimum=0, maximum=100, interactive=True)
|
| 240 |
+
injection_step = gr.Slider(label="Injection Step", precision=0, value=800, step=1, minimum=0, maximum=1000, interactive=True)
|
| 241 |
+
|
| 242 |
+
|
| 243 |
+
with gr.Row():
|
| 244 |
+
a1 = gr.Slider(label="Young", value=0, step=1, minimum=-1000000, maximum=1000000, interactive=True)
|
| 245 |
+
a2 = gr.Slider(label="Pointy Nose", value=0, step=1, minimum=-1000000, maximum=1000000, interactive=True)
|
| 246 |
+
a3 = gr.Slider(label="Undereye Bags", value=0, step=1, minimum=-1000000, maximum=1000000, interactive=True)
|
| 247 |
|
| 248 |
|
| 249 |
submit = gr.Button("Submit")
|
| 250 |
|
| 251 |
with gr.Column():
|
| 252 |
+
gallery2 = gr.Gallery(label="Identity from Edited Model")
|
| 253 |
+
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
|
| 257 |
+
sample.click(fn=sample_then_run, outputs=gallery1)
|
| 258 |
|
|
|
|
| 259 |
|
| 260 |
+
submit.click(fn=edit_inference,
|
| 261 |
+
inputs=[prompt, negative_prompt, cfg, steps, seed, injection_step, a1, a2, a3],
|
| 262 |
+
outputs=gallery2)
|
| 263 |
|
| 264 |
|
| 265 |
|
|
|
|
| 267 |
|
| 268 |
|
| 269 |
demo.launch(share=True)
|
| 270 |
+
|
| 271 |
+
|
| 272 |
+
|
| 273 |
+
|
| 274 |
+
|