Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -106,65 +106,10 @@ models_b = WurstCoreB.Models(
|
|
106 |
)
|
107 |
models_b.generator.bfloat16().eval().requires_grad_(False)
|
108 |
|
109 |
-
# Off-load old generator (low VRAM mode)
|
110 |
-
if low_vram:
|
111 |
-
models.generator.to("cpu")
|
112 |
-
torch.cuda.empty_cache()
|
113 |
-
|
114 |
-
# Load and configure new generator
|
115 |
-
generator_rbm = StageCRBM()
|
116 |
-
for param_name, param in load_or_fail(core.config.generator_checkpoint_path).items():
|
117 |
-
set_module_tensor_to_device(generator_rbm, param_name, "cpu", value=param)
|
118 |
-
|
119 |
-
generator_rbm = generator_rbm.to(getattr(torch, core.config.dtype)).to(device)
|
120 |
-
generator_rbm = core.load_model(generator_rbm, 'generator')
|
121 |
-
|
122 |
-
# Create models_rbm instance
|
123 |
-
models_rbm = core.Models(
|
124 |
-
effnet=models.effnet,
|
125 |
-
text_model=models.text_model,
|
126 |
-
tokenizer=models.tokenizer,
|
127 |
-
generator=generator_rbm,
|
128 |
-
previewer=models.previewer,
|
129 |
-
image_model=models.image_model # Add this line
|
130 |
-
)
|
131 |
-
|
132 |
-
def unload_models_and_clear_cache():
|
133 |
-
global models_rbm, models_b, sam_model, extras, extras_b
|
134 |
-
|
135 |
-
# Move all models to CPU
|
136 |
-
models_to(models_rbm, device="cpu")
|
137 |
-
|
138 |
-
# Move SAM model components to CPU if they exist
|
139 |
-
if 'sam_model' in globals():
|
140 |
-
models_to(sam_model, device="cpu")
|
141 |
-
models_to(sam_model.sam, device="cpu")
|
142 |
-
|
143 |
-
# Clear CUDA cache
|
144 |
-
torch.cuda.empty_cache()
|
145 |
-
gc.collect()
|
146 |
-
|
147 |
-
# Ensure all models are in eval mode and don't require gradients
|
148 |
-
for model in [models_rbm.generator, models_b.generator]:
|
149 |
-
model.eval()
|
150 |
-
for param in model.parameters():
|
151 |
-
param.requires_grad = False
|
152 |
-
|
153 |
-
# Clear CUDA cache again
|
154 |
-
torch.cuda.empty_cache()
|
155 |
-
gc.collect()
|
156 |
-
|
157 |
-
def reset_inference_state():
|
158 |
-
global models_rbm, models_b, extras, extras_b, device, core, core_b
|
159 |
-
|
160 |
-
# Clear CUDA cache
|
161 |
-
torch.cuda.empty_cache()
|
162 |
-
gc.collect()
|
163 |
-
|
164 |
-
models_to(models_rbm, device=device, excepts=["generator", "previewer"])
|
165 |
-
|
166 |
def infer(ref_style_file, style_description, caption):
|
167 |
-
global models_rbm, models_b
|
|
|
|
|
168 |
try:
|
169 |
caption = f"{caption} in {style_description}"
|
170 |
height=1024
|
@@ -189,7 +134,7 @@ def infer(ref_style_file, style_description, caption):
|
|
189 |
batch = {'captions': [caption] * batch_size}
|
190 |
batch['style'] = ref_style
|
191 |
|
192 |
-
x0_style_forward = models_rbm.effnet(extras.effnet_preprocess(ref_style))
|
193 |
|
194 |
conditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=False, eval_image_embeds=True, eval_style=True, eval_csd=False)
|
195 |
unconditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=True, eval_image_embeds=False)
|
@@ -246,85 +191,48 @@ def infer(ref_style_file, style_description, caption):
|
|
246 |
return output_file # Return the path to the saved image
|
247 |
|
248 |
finally:
|
249 |
-
#
|
250 |
-
|
251 |
-
# Unload models and clear cache after inference
|
252 |
-
# unload_models_and_clear_cache()
|
253 |
-
|
254 |
-
def reset_compo_inference_state():
|
255 |
-
global models_rbm, models_b, extras, extras_b, device, core, core_b, sam_model
|
256 |
-
|
257 |
-
# Reset sampling configurations
|
258 |
-
extras.sampling_configs['cfg'] = 4
|
259 |
-
extras.sampling_configs['shift'] = 2
|
260 |
-
extras.sampling_configs['timesteps'] = 20
|
261 |
-
extras.sampling_configs['t_start'] = 1.0
|
262 |
-
|
263 |
-
extras_b.sampling_configs['cfg'] = 1.1
|
264 |
-
extras_b.sampling_configs['shift'] = 1
|
265 |
-
extras_b.sampling_configs['timesteps'] = 10
|
266 |
-
extras_b.sampling_configs['t_start'] = 1.0
|
267 |
-
|
268 |
-
# Move models to CPU to free up GPU memory
|
269 |
-
models_to(models_rbm, device="cpu")
|
270 |
-
models_b.generator.to("cpu")
|
271 |
-
|
272 |
-
# Clear CUDA cache
|
273 |
-
torch.cuda.empty_cache()
|
274 |
-
gc.collect()
|
275 |
-
|
276 |
-
# Move SAM model components to CPU if they exist
|
277 |
-
models_to(sam_model, device="cpu")
|
278 |
-
models_to(sam_model.sam, device="cpu")
|
279 |
-
|
280 |
-
# Clear CUDA cache
|
281 |
-
torch.cuda.empty_cache()
|
282 |
-
gc.collect()
|
283 |
-
|
284 |
-
# Ensure all models are in eval mode and don't require gradients
|
285 |
-
for model in [models_rbm.generator, models_b.generator]:
|
286 |
-
model.eval()
|
287 |
-
for param in model.parameters():
|
288 |
-
param.requires_grad = False
|
289 |
-
|
290 |
-
# Clear CUDA cache again
|
291 |
-
torch.cuda.empty_cache()
|
292 |
-
gc.collect()
|
293 |
|
294 |
def infer_compo(style_description, ref_style_file, caption, ref_sub_file):
|
295 |
global models_rbm, models_b, device, sam_model
|
|
|
|
|
|
|
|
|
296 |
try:
|
297 |
caption = f"{caption} in {style_description}"
|
298 |
sam_prompt = f"{caption}"
|
299 |
use_sam_mask = False
|
300 |
-
|
301 |
-
# Ensure all models are on the correct device
|
302 |
-
models_to(models_rbm, device)
|
303 |
-
models_b.generator.to(device)
|
304 |
|
305 |
batch_size = 1
|
306 |
height, width = 1024, 1024
|
307 |
stage_c_latent_shape, stage_b_latent_shape = calculate_latent_sizes(height, width, batch_size=batch_size)
|
308 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
309 |
ref_style = resize_image(PIL.Image.open(ref_style_file).convert("RGB")).unsqueeze(0).expand(batch_size, -1, -1, -1).to(device)
|
310 |
ref_images = resize_image(PIL.Image.open(ref_sub_file).convert("RGB")).unsqueeze(0).expand(batch_size, -1, -1, -1).to(device)
|
311 |
|
312 |
-
batch = {'captions': [caption] * batch_size
|
|
|
|
|
313 |
|
314 |
-
x0_forward = models_rbm.effnet(extras.effnet_preprocess(ref_images))
|
315 |
-
x0_style_forward = models_rbm.effnet(extras.effnet_preprocess(ref_style))
|
316 |
|
317 |
## SAM Mask for sub
|
318 |
use_sam_mask = False
|
319 |
x0_preview = models_rbm.previewer(x0_forward)
|
320 |
sam_model = LangSAM()
|
321 |
-
|
322 |
-
# Move SAM model components to the correct device
|
323 |
-
models_to(sam_model, device)
|
324 |
-
models_to(sam_model.sam, device)
|
325 |
-
|
326 |
-
x0_preview_pil = T.ToPILImage()(x0_preview[0].cpu())
|
327 |
-
sam_mask, boxes, phrases, logits = sam_model.predict(x0_preview_pil, sam_prompt)
|
328 |
sam_mask = sam_mask.detach().unsqueeze(dim=0).to(device)
|
329 |
|
330 |
conditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=False, eval_image_embeds=True, eval_subject_style=True, eval_csd=False)
|
@@ -389,11 +297,8 @@ def infer_compo(style_description, ref_style_file, caption, ref_sub_file):
|
|
389 |
return output_file # Return the path to the saved image
|
390 |
|
391 |
finally:
|
392 |
-
#
|
393 |
-
|
394 |
-
# reset_inference_state()
|
395 |
-
# Unload models and clear cache after inference
|
396 |
-
unload_models_and_clear_cache()
|
397 |
|
398 |
def run(style_reference_image, style_description, subject_prompt, subject_reference, use_subject_ref):
|
399 |
result = None
|
|
|
106 |
)
|
107 |
models_b.generator.bfloat16().eval().requires_grad_(False)
|
108 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
109 |
def infer(ref_style_file, style_description, caption):
|
110 |
+
global models_rbm, models_b, device
|
111 |
+
if low_vram:
|
112 |
+
models_to(models_rbm, device=device)
|
113 |
try:
|
114 |
caption = f"{caption} in {style_description}"
|
115 |
height=1024
|
|
|
134 |
batch = {'captions': [caption] * batch_size}
|
135 |
batch['style'] = ref_style
|
136 |
|
137 |
+
x0_style_forward = models_rbm.effnet(extras.effnet_preprocess(ref_style.to(device)))
|
138 |
|
139 |
conditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=False, eval_image_embeds=True, eval_style=True, eval_csd=False)
|
140 |
unconditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=True, eval_image_embeds=False)
|
|
|
191 |
return output_file # Return the path to the saved image
|
192 |
|
193 |
finally:
|
194 |
+
# Clear CUDA cache
|
195 |
+
torch.cuda.empty_cache()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
196 |
|
197 |
def infer_compo(style_description, ref_style_file, caption, ref_sub_file):
|
198 |
global models_rbm, models_b, device, sam_model
|
199 |
+
if low_vram:
|
200 |
+
models_to(models_rbm, device=device)
|
201 |
+
models_to(sam_model, device=device)
|
202 |
+
models_to(sam_model.sam, device=device)
|
203 |
try:
|
204 |
caption = f"{caption} in {style_description}"
|
205 |
sam_prompt = f"{caption}"
|
206 |
use_sam_mask = False
|
|
|
|
|
|
|
|
|
207 |
|
208 |
batch_size = 1
|
209 |
height, width = 1024, 1024
|
210 |
stage_c_latent_shape, stage_b_latent_shape = calculate_latent_sizes(height, width, batch_size=batch_size)
|
211 |
+
|
212 |
+
extras.sampling_configs['cfg'] = 4
|
213 |
+
extras.sampling_configs['shift'] = 2
|
214 |
+
extras.sampling_configs['timesteps'] = 20
|
215 |
+
extras.sampling_configs['t_start'] = 1.0
|
216 |
+
extras_b.sampling_configs['cfg'] = 1.1
|
217 |
+
extras_b.sampling_configs['shift'] = 1
|
218 |
+
extras_b.sampling_configs['timesteps'] = 10
|
219 |
+
extras_b.sampling_configs['t_start'] = 1.0
|
220 |
+
|
221 |
ref_style = resize_image(PIL.Image.open(ref_style_file).convert("RGB")).unsqueeze(0).expand(batch_size, -1, -1, -1).to(device)
|
222 |
ref_images = resize_image(PIL.Image.open(ref_sub_file).convert("RGB")).unsqueeze(0).expand(batch_size, -1, -1, -1).to(device)
|
223 |
|
224 |
+
batch = {'captions': [caption] * batch_size}
|
225 |
+
batch['style'] = ref_style
|
226 |
+
batch['images'] = ref_images
|
227 |
|
228 |
+
x0_forward = models_rbm.effnet(extras.effnet_preprocess(ref_images.to(device)))
|
229 |
+
x0_style_forward = models_rbm.effnet(extras.effnet_preprocess(ref_style.to(device)))
|
230 |
|
231 |
## SAM Mask for sub
|
232 |
use_sam_mask = False
|
233 |
x0_preview = models_rbm.previewer(x0_forward)
|
234 |
sam_model = LangSAM()
|
235 |
+
sam_mask, boxes, phrases, logits = sam_model.predict(transform(x0_preview[0]), sam_prompt)
|
|
|
|
|
|
|
|
|
|
|
|
|
236 |
sam_mask = sam_mask.detach().unsqueeze(dim=0).to(device)
|
237 |
|
238 |
conditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=False, eval_image_embeds=True, eval_subject_style=True, eval_csd=False)
|
|
|
297 |
return output_file # Return the path to the saved image
|
298 |
|
299 |
finally:
|
300 |
+
# Clear CUDA cache
|
301 |
+
torch.cuda.empty_cache()
|
|
|
|
|
|
|
302 |
|
303 |
def run(style_reference_image, style_description, subject_prompt, subject_reference, use_subject_ref):
|
304 |
result = None
|