Spaces:
Build error
Build error
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,6 +1,7 @@
|
|
| 1 |
import sys
|
| 2 |
import os
|
| 3 |
from pathlib import Path
|
|
|
|
| 4 |
|
| 5 |
# Add the StableCascade and CSD directories to the Python path
|
| 6 |
app_dir = Path(__file__).parent
|
|
@@ -130,17 +131,12 @@ models_rbm = core.Models(
|
|
| 130 |
)
|
| 131 |
models_rbm.generator.eval().requires_grad_(False)
|
| 132 |
|
| 133 |
-
def
|
| 134 |
-
|
| 135 |
-
height=1024
|
| 136 |
-
width=1024
|
| 137 |
-
batch_size=1
|
| 138 |
-
output_file='output.png'
|
| 139 |
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
extras.sampling_configs['
|
| 143 |
-
extras.sampling_configs['shift'] = 2
|
| 144 |
extras.sampling_configs['timesteps'] = 20
|
| 145 |
extras.sampling_configs['t_start'] = 1.0
|
| 146 |
|
|
@@ -149,66 +145,101 @@ def infer(style_description, ref_style_file, caption):
|
|
| 149 |
extras_b.sampling_configs['timesteps'] = 10
|
| 150 |
extras_b.sampling_configs['t_start'] = 1.0
|
| 151 |
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
batch = {'captions': [caption] * batch_size}
|
| 155 |
-
batch['style'] = ref_style
|
| 156 |
-
|
| 157 |
-
x0_style_forward = models_rbm.effnet(extras.effnet_preprocess(ref_style.to(device)))
|
| 158 |
-
|
| 159 |
-
conditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=False, eval_image_embeds=True, eval_style=True, eval_csd=False)
|
| 160 |
-
unconditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=True, eval_image_embeds=False)
|
| 161 |
-
conditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=False)
|
| 162 |
-
unconditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=True)
|
| 163 |
-
|
| 164 |
if low_vram:
|
| 165 |
-
# The sampling process uses more vram, so we offload everything except two modules to the cpu.
|
| 166 |
models_to(models_rbm, device="cpu", excepts=["generator", "previewer"])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 167 |
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
apply_pushforward=False, tau_pushforward=8,
|
| 175 |
-
num_iter=3, eta=0.1, tau=20, eval_csd=True,
|
| 176 |
-
extras=extras, models=models_rbm,
|
| 177 |
-
lam_style=1, lam_txt_alignment=1.0,
|
| 178 |
-
use_ddim_sampler=True,
|
| 179 |
-
)
|
| 180 |
-
for (sampled_c, _, _) in tqdm(sampling_c, total=extras.sampling_configs['timesteps']):
|
| 181 |
-
sampled_c = sampled_c
|
| 182 |
-
|
| 183 |
-
# Stage B reverse process.
|
| 184 |
-
with torch.no_grad(), torch.cuda.amp.autocast(dtype=torch.bfloat16):
|
| 185 |
-
conditions_b['effnet'] = sampled_c
|
| 186 |
-
unconditions_b['effnet'] = torch.zeros_like(sampled_c)
|
| 187 |
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 191 |
)
|
| 192 |
-
for (
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 212 |
|
| 213 |
import gradio as gr
|
| 214 |
|
|
|
|
| 1 |
import sys
|
| 2 |
import os
|
| 3 |
from pathlib import Path
|
| 4 |
+
import gc
|
| 5 |
|
| 6 |
# Add the StableCascade and CSD directories to the Python path
|
| 7 |
app_dir = Path(__file__).parent
|
|
|
|
| 131 |
)
|
| 132 |
models_rbm.generator.eval().requires_grad_(False)
|
| 133 |
|
| 134 |
+
def reset_inference_state():
|
| 135 |
+
global models_rbm, models_b, extras, extras_b
|
|
|
|
|
|
|
|
|
|
|
|
|
| 136 |
|
| 137 |
+
# Reset sampling configurations
|
| 138 |
+
extras.sampling_configs['cfg'] = 5
|
| 139 |
+
extras.sampling_configs['shift'] = 1
|
|
|
|
| 140 |
extras.sampling_configs['timesteps'] = 20
|
| 141 |
extras.sampling_configs['t_start'] = 1.0
|
| 142 |
|
|
|
|
| 145 |
extras_b.sampling_configs['timesteps'] = 10
|
| 146 |
extras_b.sampling_configs['t_start'] = 1.0
|
| 147 |
|
| 148 |
+
# Move models back to initial state
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 149 |
if low_vram:
|
|
|
|
| 150 |
models_to(models_rbm, device="cpu", excepts=["generator", "previewer"])
|
| 151 |
+
models_b.generator.to("cpu")
|
| 152 |
+
else:
|
| 153 |
+
models_to(models_rbm, device="cuda")
|
| 154 |
+
models_b.generator.to("cuda")
|
| 155 |
+
|
| 156 |
+
# Clear CUDA cache
|
| 157 |
+
torch.cuda.empty_cache()
|
| 158 |
+
gc.collect()
|
| 159 |
|
| 160 |
+
def infer(style_description, ref_style_file, caption):
|
| 161 |
+
try:
|
| 162 |
+
height=1024
|
| 163 |
+
width=1024
|
| 164 |
+
batch_size=1
|
| 165 |
+
output_file='output.png'
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 166 |
|
| 167 |
+
stage_c_latent_shape, stage_b_latent_shape = calculate_latent_sizes(height, width, batch_size=batch_size)
|
| 168 |
+
|
| 169 |
+
extras.sampling_configs['cfg'] = 4
|
| 170 |
+
extras.sampling_configs['shift'] = 2
|
| 171 |
+
extras.sampling_configs['timesteps'] = 20
|
| 172 |
+
extras.sampling_configs['t_start'] = 1.0
|
| 173 |
+
|
| 174 |
+
extras_b.sampling_configs['cfg'] = 1.1
|
| 175 |
+
extras_b.sampling_configs['shift'] = 1
|
| 176 |
+
extras_b.sampling_configs['timesteps'] = 10
|
| 177 |
+
extras_b.sampling_configs['t_start'] = 1.0
|
| 178 |
+
|
| 179 |
+
ref_style = resize_image(PIL.Image.open(ref_style_file).convert("RGB")).unsqueeze(0).expand(batch_size, -1, -1, -1).to(device)
|
| 180 |
+
|
| 181 |
+
batch = {'captions': [caption] * batch_size}
|
| 182 |
+
batch['style'] = ref_style
|
| 183 |
+
|
| 184 |
+
x0_style_forward = models_rbm.effnet(extras.effnet_preprocess(ref_style.to(device)))
|
| 185 |
+
|
| 186 |
+
conditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=False, eval_image_embeds=True, eval_style=True, eval_csd=False)
|
| 187 |
+
unconditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=True, eval_image_embeds=False)
|
| 188 |
+
conditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=False)
|
| 189 |
+
unconditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=True)
|
| 190 |
+
|
| 191 |
+
if low_vram:
|
| 192 |
+
# The sampling process uses more vram, so we offload everything except two modules to the cpu.
|
| 193 |
+
models_to(models_rbm, device="cpu", excepts=["generator", "previewer"])
|
| 194 |
+
|
| 195 |
+
# Stage C reverse process.
|
| 196 |
+
sampling_c = extras.gdf.sample(
|
| 197 |
+
models_rbm.generator, conditions, stage_c_latent_shape,
|
| 198 |
+
unconditions, device=device,
|
| 199 |
+
**extras.sampling_configs,
|
| 200 |
+
x0_style_forward=x0_style_forward,
|
| 201 |
+
apply_pushforward=False, tau_pushforward=8,
|
| 202 |
+
num_iter=3, eta=0.1, tau=20, eval_csd=True,
|
| 203 |
+
extras=extras, models=models_rbm,
|
| 204 |
+
lam_style=1, lam_txt_alignment=1.0,
|
| 205 |
+
use_ddim_sampler=True,
|
| 206 |
)
|
| 207 |
+
for (sampled_c, _, _) in tqdm(sampling_c, total=extras.sampling_configs['timesteps']):
|
| 208 |
+
sampled_c = sampled_c
|
| 209 |
+
|
| 210 |
+
# Stage B reverse process.
|
| 211 |
+
with torch.no_grad(), torch.cuda.amp.autocast(dtype=torch.bfloat16):
|
| 212 |
+
conditions_b['effnet'] = sampled_c
|
| 213 |
+
unconditions_b['effnet'] = torch.zeros_like(sampled_c)
|
| 214 |
+
|
| 215 |
+
sampling_b = extras_b.gdf.sample(
|
| 216 |
+
models_b.generator, conditions_b, stage_b_latent_shape,
|
| 217 |
+
unconditions_b, device=device, **extras_b.sampling_configs,
|
| 218 |
+
)
|
| 219 |
+
for (sampled_b, _, _) in tqdm(sampling_b, total=extras_b.sampling_configs['timesteps']):
|
| 220 |
+
sampled_b = sampled_b
|
| 221 |
+
sampled = models_b.stage_a.decode(sampled_b).float()
|
| 222 |
+
|
| 223 |
+
sampled = torch.cat([
|
| 224 |
+
torch.nn.functional.interpolate(ref_style.cpu(), size=(height, width)),
|
| 225 |
+
sampled.cpu(),
|
| 226 |
+
], dim=0)
|
| 227 |
+
|
| 228 |
+
# Remove the batch dimension and keep only the generated image
|
| 229 |
+
sampled = sampled[1] # This selects the generated image, discarding the reference style image
|
| 230 |
+
|
| 231 |
+
# Ensure the tensor is in [C, H, W] format
|
| 232 |
+
if sampled.dim() == 3 and sampled.shape[0] == 3:
|
| 233 |
+
sampled_image = T.ToPILImage()(sampled) # Convert tensor to PIL image
|
| 234 |
+
sampled_image.save(output_file) # Save the image as a PNG
|
| 235 |
+
else:
|
| 236 |
+
raise ValueError(f"Expected tensor of shape [3, H, W] but got {sampled.shape}")
|
| 237 |
+
|
| 238 |
+
return output_file # Return the path to the saved image
|
| 239 |
+
|
| 240 |
+
finally:
|
| 241 |
+
# Reset the state after inference, regardless of success or failure
|
| 242 |
+
reset_inference_state()
|
| 243 |
|
| 244 |
import gradio as gr
|
| 245 |
|