Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -71,7 +71,6 @@ def gaussian_kernel(kernel_size=3, sigma=1.0, channels=3):
|
|
71 |
gaussian_1d = gaussian_1d / gaussian_1d.sum()
|
72 |
gaussian_2d = gaussian_1d[:, None] * gaussian_1d[None, :]
|
73 |
kernel = gaussian_2d[None, None, :, :].repeat(channels, 1, 1, 1)
|
74 |
-
|
75 |
return kernel
|
76 |
|
77 |
def gaussian_filter(latents, kernel_size=3, sigma=1.0):
|
@@ -88,9 +87,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
|
|
88 |
"""
|
89 |
std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
|
90 |
std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
|
91 |
-
# rescale the results from guidance (fixes overexposure)
|
92 |
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
|
93 |
-
# mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
|
94 |
noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
|
95 |
return noise_cfg
|
96 |
|
@@ -144,7 +141,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
144 |
add_watermarker: Optional[bool] = None,
|
145 |
):
|
146 |
super().__init__()
|
147 |
-
|
148 |
self.register_modules(
|
149 |
vae=vae,
|
150 |
text_encoder=text_encoder,
|
@@ -158,27 +154,21 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
158 |
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
|
159 |
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
|
160 |
self.default_sample_size = self.unet.config.sample_size
|
161 |
-
|
162 |
add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
|
163 |
-
|
164 |
if add_watermarker:
|
165 |
self.watermark = StableDiffusionXLWatermarker()
|
166 |
else:
|
167 |
self.watermark = None
|
168 |
|
169 |
-
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
|
170 |
def enable_vae_slicing(self):
|
171 |
self.vae.enable_slicing()
|
172 |
|
173 |
-
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
|
174 |
def disable_vae_slicing(self):
|
175 |
self.vae.disable_slicing()
|
176 |
|
177 |
-
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
|
178 |
def enable_vae_tiling(self):
|
179 |
self.vae.enable_tiling()
|
180 |
|
181 |
-
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
|
182 |
def disable_vae_tiling(self):
|
183 |
self.vae.disable_tiling()
|
184 |
|
@@ -198,25 +188,20 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
198 |
lora_scale: Optional[float] = None,
|
199 |
):
|
200 |
device = device or self._execution_device
|
201 |
-
|
202 |
-
# set lora scale so that monkey patched LoRA function of text encoder can correctly access it
|
203 |
if lora_scale is not None and isinstance(self, LoraLoaderMixin):
|
204 |
self._lora_scale = lora_scale
|
205 |
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
|
206 |
adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
|
207 |
-
|
208 |
if prompt is not None and isinstance(prompt, str):
|
209 |
batch_size = 1
|
210 |
elif prompt is not None and isinstance(prompt, list):
|
211 |
batch_size = len(prompt)
|
212 |
else:
|
213 |
batch_size = prompt_embeds.shape[0]
|
214 |
-
|
215 |
tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
|
216 |
text_encoders = (
|
217 |
[self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
|
218 |
)
|
219 |
-
|
220 |
if prompt_embeds is None:
|
221 |
prompt_2 = prompt_2 or prompt
|
222 |
prompt_embeds_list = []
|
@@ -224,7 +209,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
224 |
for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
|
225 |
if isinstance(self, TextualInversionLoaderMixin):
|
226 |
prompt = self.maybe_convert_prompt(prompt, tokenizer)
|
227 |
-
|
228 |
text_inputs = tokenizer(
|
229 |
prompt,
|
230 |
padding="max_length",
|
@@ -232,10 +216,8 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
232 |
truncation=True,
|
233 |
return_tensors="pt",
|
234 |
)
|
235 |
-
|
236 |
text_input_ids = text_inputs.input_ids
|
237 |
untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
|
238 |
-
|
239 |
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
|
240 |
text_input_ids, untruncated_ids
|
241 |
):
|
@@ -244,19 +226,14 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
244 |
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
245 |
f" {tokenizer.model_max_length} tokens: {removed_text}"
|
246 |
)
|
247 |
-
|
248 |
prompt_embeds = text_encoder(
|
249 |
text_input_ids.to(device),
|
250 |
output_hidden_states=True,
|
251 |
)
|
252 |
-
|
253 |
pooled_prompt_embeds = prompt_embeds[0]
|
254 |
prompt_embeds = prompt_embeds.hidden_states[-2]
|
255 |
-
|
256 |
prompt_embeds_list.append(prompt_embeds)
|
257 |
-
|
258 |
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
|
259 |
-
|
260 |
zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
|
261 |
if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
|
262 |
negative_prompt_embeds = torch.zeros_like(prompt_embeds)
|
@@ -264,7 +241,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
264 |
elif do_classifier_free_guidance and negative_prompt_embeds is None:
|
265 |
negative_prompt = negative_prompt or ""
|
266 |
negative_prompt_2 = negative_prompt_2 or negative_prompt
|
267 |
-
|
268 |
if prompt is not None and type(prompt) is not type(negative_prompt):
|
269 |
raise TypeError(
|
270 |
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
@@ -280,12 +256,10 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
280 |
)
|
281 |
else:
|
282 |
uncond_tokens = [negative_prompt, negative_prompt_2]
|
283 |
-
|
284 |
negative_prompt_embeds_list = []
|
285 |
for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
|
286 |
if isinstance(self, TextualInversionLoaderMixin):
|
287 |
negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
|
288 |
-
|
289 |
max_length = prompt_embeds.shape[1]
|
290 |
uncond_input = tokenizer(
|
291 |
negative_prompt,
|
@@ -294,29 +268,23 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
294 |
truncation=True,
|
295 |
return_tensors="pt",
|
296 |
)
|
297 |
-
|
298 |
negative_prompt_embeds = text_encoder(
|
299 |
uncond_input.input_ids.to(device),
|
300 |
output_hidden_states=True,
|
301 |
)
|
302 |
negative_pooled_prompt_embeds = negative_prompt_embeds[0]
|
303 |
negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
|
304 |
-
|
305 |
negative_prompt_embeds_list.append(negative_prompt_embeds)
|
306 |
-
|
307 |
negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
|
308 |
-
|
309 |
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
|
310 |
bs_embed, seq_len, _ = prompt_embeds.shape
|
311 |
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
312 |
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
313 |
-
|
314 |
if do_classifier_free_guidance:
|
315 |
seq_len = negative_prompt_embeds.shape[1]
|
316 |
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
|
317 |
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
318 |
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
319 |
-
|
320 |
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
|
321 |
bs_embed * num_images_per_prompt, -1
|
322 |
)
|
@@ -324,7 +292,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
324 |
negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
|
325 |
bs_embed * num_images_per_prompt, -1
|
326 |
)
|
327 |
-
|
328 |
return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
|
329 |
|
330 |
def prepare_extra_step_kwargs(self, generator, eta):
|
@@ -332,7 +299,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
332 |
extra_step_kwargs = {}
|
333 |
if accepts_eta:
|
334 |
extra_step_kwargs["eta"] = eta
|
335 |
-
|
336 |
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
337 |
if accepts_generator:
|
338 |
extra_step_kwargs["generator"] = generator
|
@@ -355,14 +321,12 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
355 |
):
|
356 |
if height % 8 != 0 or width % 8 != 0:
|
357 |
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
|
358 |
-
|
359 |
if (callback_steps is None) or (
|
360 |
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
|
361 |
):
|
362 |
raise ValueError(
|
363 |
f"`callback_steps` has to be a positive integer but is {callback_steps} of type {type(callback_steps)}."
|
364 |
)
|
365 |
-
|
366 |
if prompt is not None and prompt_embeds is not None:
|
367 |
raise ValueError(
|
368 |
f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to only forward one of the two."
|
@@ -379,7 +343,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
379 |
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
380 |
elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
|
381 |
raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
|
382 |
-
|
383 |
if negative_prompt is not None and negative_prompt_embeds is not None:
|
384 |
raise ValueError(
|
385 |
f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to only forward one of the two."
|
@@ -388,26 +351,21 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
388 |
raise ValueError(
|
389 |
f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to only forward one of the two."
|
390 |
)
|
391 |
-
|
392 |
if prompt_embeds is not None and negative_prompt_embeds is not None:
|
393 |
if prompt_embeds.shape != negative_prompt_embeds.shape:
|
394 |
raise ValueError(
|
395 |
"`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds` {negative_prompt_embeds.shape}."
|
396 |
)
|
397 |
-
|
398 |
if prompt_embeds is not None and pooled_prompt_embeds is None:
|
399 |
raise ValueError(
|
400 |
"If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
|
401 |
)
|
402 |
-
|
403 |
if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
|
404 |
raise ValueError(
|
405 |
"If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
|
406 |
)
|
407 |
-
|
408 |
if max(height, width) % 1024 != 0:
|
409 |
raise ValueError(f"the larger one of `height` and `width` has to be divisible by 1024 but are {height} and {width}.")
|
410 |
-
|
411 |
if num_images_per_prompt != 1:
|
412 |
warnings.warn("num_images_per_prompt != 1 is not supported by AccDiffusion and will be ignored.")
|
413 |
num_images_per_prompt = 1
|
@@ -418,29 +376,24 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
418 |
raise ValueError(
|
419 |
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch size of {batch_size}. Make sure the batch size matches the length of the generators."
|
420 |
)
|
421 |
-
|
422 |
if latents is None:
|
423 |
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
424 |
else:
|
425 |
latents = latents.to(device)
|
426 |
-
|
427 |
latents = latents * self.scheduler.init_noise_sigma
|
428 |
return latents
|
429 |
|
430 |
def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
|
431 |
add_time_ids = list(original_size + crops_coords_top_left + target_size)
|
432 |
-
|
433 |
passed_add_embed_dim = (
|
434 |
self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
|
435 |
)
|
436 |
expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
|
437 |
-
|
438 |
if expected_add_embed_dim != passed_add_embed_dim:
|
439 |
raise ValueError(
|
440 |
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. \
|
441 |
The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
|
442 |
)
|
443 |
-
|
444 |
add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
|
445 |
return add_time_ids
|
446 |
|
@@ -456,7 +409,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
456 |
h_end = h_start + window_size
|
457 |
w_start = int((i % num_blocks_width) * stride)
|
458 |
w_end = w_start + window_size
|
459 |
-
|
460 |
if h_end > height:
|
461 |
h_start = int(h_start + height - h_end)
|
462 |
h_end = int(height)
|
@@ -469,7 +421,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
469 |
if w_start < 0:
|
470 |
w_end = int(w_end - w_start)
|
471 |
w_start = 0
|
472 |
-
|
473 |
if random_jitter:
|
474 |
jitter_range = (window_size - stride) // 4
|
475 |
w_jitter = 0
|
@@ -480,7 +431,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
480 |
w_jitter = random.randint(-jitter_range, 0)
|
481 |
elif (w_start != 0) and (w_end == width):
|
482 |
w_jitter = random.randint(0, jitter_range)
|
483 |
-
|
484 |
if (h_start != 0) and (h_end != height):
|
485 |
h_jitter = random.randint(-jitter_range, jitter_range)
|
486 |
elif (h_start == 0) and (h_end != height):
|
@@ -491,7 +441,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
491 |
h_end = h_end + h_jitter + jitter_range
|
492 |
w_start = w_start + w_jitter + jitter_range
|
493 |
w_end = w_end + w_jitter + jitter_range
|
494 |
-
|
495 |
views.append((h_start, h_end, w_start, w_end))
|
496 |
return views
|
497 |
|
@@ -527,7 +476,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
527 |
continue
|
528 |
cross_att_count += 1
|
529 |
attn_procs[name] = P2PCrossAttnProcessor(controller=controller, place_in_unet=place_in_unet)
|
530 |
-
|
531 |
self.unet.set_attn_processor(attn_procs)
|
532 |
controller.num_att_layers = cross_att_count
|
533 |
return ori_attn_processors
|
@@ -540,7 +488,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
540 |
from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
|
541 |
else:
|
542 |
raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
|
543 |
-
|
544 |
is_model_cpu_offload = False
|
545 |
is_sequential_cpu_offload = False
|
546 |
recursive = False
|
@@ -560,7 +507,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
560 |
**kwargs,
|
561 |
)
|
562 |
self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet)
|
563 |
-
|
564 |
text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
|
565 |
if len(text_encoder_state_dict) > 0:
|
566 |
self.load_lora_into_text_encoder(
|
@@ -570,7 +516,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
570 |
prefix="text_encoder",
|
571 |
lora_scale=self.lora_scale,
|
572 |
)
|
573 |
-
|
574 |
text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
|
575 |
if len(text_encoder_2_state_dict) > 0:
|
576 |
self.load_lora_into_text_encoder(
|
@@ -580,7 +525,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
580 |
prefix="text_encoder_2",
|
581 |
lora_scale=self.lora_scale,
|
582 |
)
|
583 |
-
|
584 |
if is_model_cpu_offload:
|
585 |
self.enable_model_cpu_offload()
|
586 |
elif is_sequential_cpu_offload:
|
@@ -599,24 +543,19 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
599 |
safe_serialization: bool = True,
|
600 |
):
|
601 |
state_dict = {}
|
602 |
-
|
603 |
def pack_weights(layers, prefix):
|
604 |
layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
|
605 |
layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
|
606 |
return layers_state_dict
|
607 |
-
|
608 |
if not (unet_lora_layers or text_encoder_lora_layers or text_encoder_2_lora_layers):
|
609 |
raise ValueError(
|
610 |
"You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers` or `text_encoder_2_lora_layers`."
|
611 |
)
|
612 |
-
|
613 |
if unet_lora_layers:
|
614 |
state_dict.update(pack_weights(unet_lora_layers, "unet"))
|
615 |
-
|
616 |
if text_encoder_lora_layers and text_encoder_2_lora_layers:
|
617 |
state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder"))
|
618 |
state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))
|
619 |
-
|
620 |
self.write_lora_layers(
|
621 |
state_dict=state_dict,
|
622 |
save_directory=save_directory,
|
@@ -697,39 +636,29 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
697 |
|
698 |
Examples:
|
699 |
"""
|
700 |
-
|
701 |
if debug:
|
702 |
num_inference_steps = 1
|
703 |
-
|
704 |
height = height or self.default_sample_size * self.vae_scale_factor
|
705 |
width = width or self.default_sample_size * self.vae_scale_factor
|
706 |
-
|
707 |
x1_size = self.default_sample_size * self.vae_scale_factor
|
708 |
-
|
709 |
height_scale = height / x1_size
|
710 |
width_scale = width / x1_size
|
711 |
scale_num = int(max(height_scale, width_scale))
|
712 |
aspect_ratio = min(height_scale, width_scale) / max(height_scale, width_scale)
|
713 |
-
|
714 |
original_size = original_size or (height, width)
|
715 |
target_size = target_size or (height, width)
|
716 |
-
|
717 |
if attn_res is None:
|
718 |
-
attn_res = int(np.ceil(self.default_sample_size * self.vae_scale_factor / 32)), int(np.ceil(self.default_sample_size * self.vae_scale_factor / 32))
|
719 |
self.attn_res = attn_res
|
720 |
-
|
721 |
if lowvram:
|
722 |
attention_map_device = torch.device("cpu")
|
723 |
else:
|
724 |
attention_map_device = self.device
|
725 |
-
|
726 |
self.controller = create_controller(
|
727 |
prompt, cross_attention_kwargs, num_inference_steps, tokenizer=self.tokenizer, device=attention_map_device, attn_res=self.attn_res
|
728 |
)
|
729 |
-
|
730 |
if save_attention_map or use_md_prompt:
|
731 |
ori_attn_processors = self.register_attention_control(self.controller)
|
732 |
-
|
733 |
self.check_inputs(
|
734 |
prompt,
|
735 |
prompt_2,
|
@@ -744,14 +673,12 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
744 |
negative_pooled_prompt_embeds,
|
745 |
num_images_per_prompt,
|
746 |
)
|
747 |
-
|
748 |
if prompt is not None and isinstance(prompt, str):
|
749 |
batch_size = 1
|
750 |
elif prompt is not None and isinstance(prompt, list):
|
751 |
batch_size = len(prompt)
|
752 |
else:
|
753 |
batch_size = prompt_embeds.shape[0]
|
754 |
-
|
755 |
device = self._execution_device
|
756 |
self.lowvram = lowvram
|
757 |
if self.lowvram:
|
@@ -759,13 +686,8 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
759 |
self.unet.cpu()
|
760 |
self.text_encoder.to(device)
|
761 |
self.text_encoder_2.to(device)
|
762 |
-
|
763 |
do_classifier_free_guidance = guidance_scale > 1.0
|
764 |
-
|
765 |
-
text_encoder_lora_scale = (
|
766 |
-
cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
|
767 |
-
)
|
768 |
-
|
769 |
(
|
770 |
prompt_embeds,
|
771 |
negative_prompt_embeds,
|
@@ -785,10 +707,8 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
785 |
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
786 |
lora_scale=text_encoder_lora_scale,
|
787 |
)
|
788 |
-
|
789 |
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
790 |
timesteps = self.scheduler.timesteps
|
791 |
-
|
792 |
num_channels_latents = self.unet.config.in_channels
|
793 |
latents = self.prepare_latents(
|
794 |
batch_size * num_images_per_prompt,
|
@@ -800,15 +720,11 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
800 |
generator,
|
801 |
latents,
|
802 |
)
|
803 |
-
|
804 |
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
805 |
-
|
806 |
add_text_embeds = pooled_prompt_embeds
|
807 |
-
|
808 |
add_time_ids = self._get_add_time_ids(
|
809 |
original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
|
810 |
)
|
811 |
-
|
812 |
if negative_original_size is not None and negative_target_size is not None:
|
813 |
negative_add_time_ids = self._get_add_time_ids(
|
814 |
negative_original_size,
|
@@ -818,16 +734,12 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
818 |
)
|
819 |
else:
|
820 |
negative_add_time_ids = add_time_ids
|
821 |
-
|
822 |
if do_classifier_free_guidance:
|
823 |
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0).to(device)
|
824 |
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0).to(device)
|
825 |
add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0).to(device).repeat(batch_size * num_images_per_prompt, 1)
|
826 |
-
|
827 |
del negative_prompt_embeds, negative_pooled_prompt_embeds, negative_add_time_ids
|
828 |
-
|
829 |
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
830 |
-
|
831 |
if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
|
832 |
discrete_timestep_cutoff = int(
|
833 |
round(
|
@@ -836,35 +748,26 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
836 |
)
|
837 |
num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
|
838 |
timesteps = timesteps[:num_inference_steps]
|
839 |
-
|
840 |
output_images = []
|
841 |
-
|
842 |
###################################################### Phase Initialization ########################################################
|
843 |
-
|
844 |
if self.lowvram:
|
845 |
self.text_encoder.cpu()
|
846 |
self.text_encoder_2.cpu()
|
847 |
-
|
848 |
if image_lr == None:
|
849 |
print("### Phase 1 Denoising ###")
|
850 |
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
851 |
for i, t in enumerate(timesteps):
|
852 |
-
|
853 |
if self.lowvram:
|
854 |
self.vae.cpu()
|
855 |
self.unet.to(device)
|
856 |
-
|
857 |
latents_for_view = latents
|
858 |
-
|
859 |
latent_model_input = (
|
860 |
latents.repeat_interleave(2, dim=0)
|
861 |
if do_classifier_free_guidance
|
862 |
else latents
|
863 |
)
|
864 |
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
865 |
-
|
866 |
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
|
867 |
-
|
868 |
noise_pred = self.unet(
|
869 |
latent_model_input,
|
870 |
t,
|
@@ -872,25 +775,19 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
872 |
added_cond_kwargs=added_cond_kwargs,
|
873 |
return_dict=False,
|
874 |
)[0]
|
875 |
-
|
876 |
if do_classifier_free_guidance:
|
877 |
noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
|
878 |
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
879 |
-
|
880 |
if do_classifier_free_guidance and guidance_rescale > 0.0:
|
881 |
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
|
882 |
-
|
883 |
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
884 |
-
|
885 |
if t == 1 and use_md_prompt:
|
886 |
md_prompts, views_attention = get_multidiffusion_prompts(tokenizer=self.tokenizer, prompts=[prompt], threthod=c, attention_store=self.controller, height=height//scale_num, width=width//scale_num, from_where=["up","down"], random_jitter=True, scale_num=scale_num)
|
887 |
-
|
888 |
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
889 |
progress_bar.update()
|
890 |
if callback is not None and i % callback_steps == 0:
|
891 |
step_idx = i // getattr(self.scheduler, "order", 1)
|
892 |
callback(step_idx, t, latents)
|
893 |
-
|
894 |
del latents_for_view, latent_model_input, noise_pred, noise_pred_text, noise_pred_uncond
|
895 |
if use_md_prompt or save_attention_map:
|
896 |
self.recover_attention_control(ori_attn_processors=ori_attn_processors)
|
@@ -900,7 +797,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
900 |
print("### Encoding Real Image ###")
|
901 |
latents = self.vae.encode(image_lr)
|
902 |
latents = latents.latent_dist.sample() * self.vae.config.scaling_factor
|
903 |
-
|
904 |
anchor_mean = latents.mean()
|
905 |
anchor_std = latents.std()
|
906 |
if self.lowvram:
|
@@ -908,12 +804,10 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
908 |
torch.cuda.empty_cache()
|
909 |
if not output_type == "latent":
|
910 |
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
|
911 |
-
|
912 |
if self.lowvram:
|
913 |
needs_upcasting = False
|
914 |
self.unet.cpu()
|
915 |
self.vae.to(device)
|
916 |
-
|
917 |
if needs_upcasting:
|
918 |
self.upcast_vae()
|
919 |
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
|
@@ -924,15 +818,12 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
924 |
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
925 |
if needs_upcasting:
|
926 |
self.vae.to(dtype=torch.float16)
|
927 |
-
|
928 |
image = self.image_processor.postprocess(image, output_type=output_type)
|
929 |
if not os.path.exists(f'{result_path}'):
|
930 |
os.makedirs(f'{result_path}')
|
931 |
-
|
932 |
image_lr_save_path = f'{result_path}/{image[0].size[0]}_{image[0].size[1]}.png'
|
933 |
image[0].save(image_lr_save_path)
|
934 |
output_images.append(image[0])
|
935 |
-
|
936 |
####################################################### Phase Upscaling #####################################################
|
937 |
if use_progressive_upscaling:
|
938 |
if image_lr == None:
|
@@ -941,26 +832,21 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
941 |
starting_scale = 1
|
942 |
else:
|
943 |
starting_scale = scale_num
|
944 |
-
|
945 |
for current_scale_num in range(starting_scale, scale_num + 1):
|
946 |
if self.lowvram:
|
947 |
latents = latents.to(device)
|
948 |
self.unet.to(device)
|
949 |
torch.cuda.empty_cache()
|
950 |
-
|
951 |
current_height = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
|
952 |
current_width = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
|
953 |
-
|
954 |
if height > width:
|
955 |
current_width = int(current_width * aspect_ratio)
|
956 |
else:
|
957 |
current_height = int(current_height * aspect_ratio)
|
958 |
-
|
959 |
if upscale_mode == "bicubic_latent" or debug:
|
960 |
latents = F.interpolate(latents.to(device), size=(int(current_height / self.vae_scale_factor), int(current_width / self.vae_scale_factor)), mode='bicubic')
|
961 |
else:
|
962 |
raise NotImplementedError
|
963 |
-
|
964 |
print("### Phase {} Denoising ###".format(current_scale_num))
|
965 |
noise_latents = []
|
966 |
noise = torch.randn_like(latents)
|
@@ -968,7 +854,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
968 |
noise_latent = self.scheduler.add_noise(latents, noise, timestep.unsqueeze(0))
|
969 |
noise_latents.append(noise_latent)
|
970 |
latents = noise_latents[0]
|
971 |
-
|
972 |
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
973 |
for i, t in enumerate(timesteps):
|
974 |
count = torch.zeros_like(latents)
|
@@ -977,7 +862,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
977 |
if use_skip_residual:
|
978 |
c1 = cosine_factor ** cosine_scale_1
|
979 |
latents = latents * (1 - c1) + noise_latents[i] * c1
|
980 |
-
|
981 |
if use_multidiffusion:
|
982 |
if use_md_prompt:
|
983 |
md_prompt_embeds_list = []
|
@@ -1005,46 +889,33 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
1005 |
md_prompt_embeds_list.append(torch.cat([md_negative_prompt_embeds, md_prompt_embeds], dim=0).to(device))
|
1006 |
md_add_text_embeds_list.append(torch.cat([md_negative_pooled_prompt_embeds, md_pooled_prompt_embeds], dim=0).to(device))
|
1007 |
del md_negative_prompt_embeds, md_negative_pooled_prompt_embeds
|
1008 |
-
|
1009 |
if use_md_prompt:
|
1010 |
random_jitter = True
|
1011 |
views = [(h_start*4, h_end*4, w_start*4, w_end*4) for h_start, h_end, w_start, w_end in views_attention[current_scale_num]]
|
1012 |
else:
|
1013 |
random_jitter = True
|
1014 |
views = self.get_views(current_height, current_width, stride=stride, window_size=self.unet.config.sample_size, random_jitter=random_jitter)
|
1015 |
-
|
1016 |
views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
|
1017 |
-
|
1018 |
if use_md_prompt:
|
1019 |
views_prompt_embeds_input = [md_prompt_embeds_list[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
|
1020 |
views_add_text_embeds_input = [md_add_text_embeds_list[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
|
1021 |
-
|
1022 |
if random_jitter:
|
1023 |
jitter_range = int((self.unet.config.sample_size - stride) // 4)
|
1024 |
latents_ = F.pad(latents, (jitter_range, jitter_range, jitter_range, jitter_range), 'constant', 0)
|
1025 |
else:
|
1026 |
latents_ = latents
|
1027 |
-
|
1028 |
count_local = torch.zeros_like(latents_)
|
1029 |
value_local = torch.zeros_like(latents_)
|
1030 |
-
|
1031 |
for j, batch_view in enumerate(views_batch):
|
1032 |
vb_size = len(batch_view)
|
1033 |
latents_for_view = torch.cat(
|
1034 |
-
[
|
1035 |
-
latents_[:, :, h_start:h_end, w_start:w_end]
|
1036 |
-
for h_start, h_end, w_start, w_end in batch_view
|
1037 |
-
]
|
1038 |
)
|
1039 |
-
|
1040 |
latent_model_input = latents_for_view
|
1041 |
-
latent_model_input = (
|
1042 |
-
|
1043 |
-
|
1044 |
-
else latent_model_input
|
1045 |
-
)
|
1046 |
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
1047 |
-
|
1048 |
add_time_ids_input = []
|
1049 |
for h_start, h_end, w_start, w_end in batch_view:
|
1050 |
add_time_ids_ = add_time_ids.clone()
|
@@ -1052,7 +923,6 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
1052 |
add_time_ids_[:, 3] = w_start * self.vae_scale_factor
|
1053 |
add_time_ids_input.append(add_time_ids_)
|
1054 |
add_time_ids_input = torch.cat(add_time_ids_input)
|
1055 |
-
|
1056 |
if not use_md_prompt:
|
1057 |
prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
|
1058 |
add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
|
@@ -1075,30 +945,20 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
1075 |
added_cond_kwargs=md_added_cond_kwargs,
|
1076 |
return_dict=False,
|
1077 |
)[0]
|
1078 |
-
|
1079 |
if do_classifier_free_guidance:
|
1080 |
noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
|
1081 |
noise_pred = noise_pred_uncond + multi_guidance_scale * (noise_pred_text - noise_pred_uncond)
|
1082 |
-
|
1083 |
if do_classifier_free_guidance and guidance_rescale > 0.0:
|
1084 |
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
|
1085 |
-
|
1086 |
self.scheduler._init_step_index(t)
|
1087 |
-
latents_denoised_batch = self.scheduler.step(
|
1088 |
-
|
1089 |
-
|
1090 |
-
for latents_view_denoised, (h_start, h_end, w_start, w_end) in zip(
|
1091 |
-
latents_denoised_batch.chunk(vb_size), batch_view
|
1092 |
-
):
|
1093 |
value_local[:, :, h_start:h_end, w_start:w_end] += latents_view_denoised
|
1094 |
count_local[:, :, h_start:h_end, w_start:w_end] += 1
|
1095 |
-
|
1096 |
if random_jitter:
|
1097 |
value_local = value_local[:, :, jitter_range:jitter_range + current_height // self.vae_scale_factor, jitter_range:jitter_range + current_width // self.vae_scale_factor]
|
1098 |
count_local = count_local[:, :, jitter_range:jitter_range + current_height // self.vae_scale_factor, jitter_range:jitter_range + current_width // self.vae_scale_factor]
|
1099 |
-
|
1100 |
noise_index = i + 1 if i != (len(timesteps) - 1) else i
|
1101 |
-
|
1102 |
value_local = torch.where(count_local == 0, noise_latents[noise_index], value_local)
|
1103 |
count_local = torch.where(count_local == 0, torch.ones_like(count_local), count_local)
|
1104 |
if use_dilated_sampling:
|
@@ -1107,19 +967,15 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
1107 |
count += torch.ones_like(value_local) * (1 - c2)
|
1108 |
else:
|
1109 |
value += value_local / count_local
|
1110 |
-
count += torch.ones_like(value_local)
|
1111 |
-
|
1112 |
if use_dilated_sampling:
|
1113 |
views = [[h, w] for h in range(current_scale_num) for w in range(current_scale_num)]
|
1114 |
views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
|
1115 |
-
|
1116 |
h_pad = (current_scale_num - (latents.size(2) % current_scale_num)) % current_scale_num
|
1117 |
w_pad = (current_scale_num - (latents.size(3) % current_scale_num)) % current_scale_num
|
1118 |
latents_ = F.pad(latents, (w_pad, 0, h_pad, 0), 'constant', 0)
|
1119 |
-
|
1120 |
count_global = torch.zeros_like(latents_)
|
1121 |
value_global = torch.zeros_like(latents_)
|
1122 |
-
|
1123 |
if use_guassian:
|
1124 |
c3 = 0.99 * cosine_factor ** cosine_scale_3 + 1e-2
|
1125 |
std_, mean_ = latents_.std(), latents_.mean()
|
@@ -1127,47 +983,31 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
1127 |
latents_gaussian = (latents_gaussian - latents_gaussian.mean()) / latents_gaussian.std() * std_ + mean_
|
1128 |
else:
|
1129 |
latents_gaussian = latents_
|
1130 |
-
|
1131 |
for j, batch_view in enumerate(views_batch):
|
1132 |
-
|
1133 |
latents_for_view = torch.cat(
|
1134 |
-
[
|
1135 |
-
latents_[:, :, h::current_scale_num, w::current_scale_num]
|
1136 |
-
for h, w in batch_view
|
1137 |
-
]
|
1138 |
)
|
1139 |
-
|
1140 |
latents_for_view_gaussian = torch.cat(
|
1141 |
-
[
|
1142 |
-
latents_gaussian[:, :, h::current_scale_num, w::current_scale_num]
|
1143 |
-
for h, w in batch_view
|
1144 |
-
]
|
1145 |
)
|
1146 |
-
|
1147 |
if shuffle:
|
1148 |
shape = latents_for_view.shape
|
1149 |
-
# 수정: range(...)
|
1150 |
shuffle_index = torch.stack([torch.randperm(shape[0]) for _ in range(latents_for_view.reshape(-1).shape[0]//shape[0])])
|
1151 |
shuffle_index = shuffle_index.view(shape[1], shape[2], shape[3], shape[0])
|
1152 |
original_index = torch.zeros_like(shuffle_index).scatter_(3, shuffle_index, torch.arange(shape[0]).repeat(shape[1], shape[2], shape[3], 1))
|
1153 |
shuffle_index = shuffle_index.permute(3, 0, 1, 2).to(device)
|
1154 |
original_index = original_index.permute(3, 0, 1, 2).to(device)
|
1155 |
latents_for_view_gaussian = latents_for_view_gaussian.gather(0, shuffle_index)
|
1156 |
-
|
1157 |
vb_size = latents_for_view.size(0)
|
1158 |
-
|
1159 |
latent_model_input = latents_for_view_gaussian
|
1160 |
-
latent_model_input = (
|
1161 |
-
|
1162 |
-
|
1163 |
-
else latent_model_input
|
1164 |
-
)
|
1165 |
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
1166 |
-
|
1167 |
prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
|
1168 |
add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
|
1169 |
add_time_ids_input = torch.cat([add_time_ids] * vb_size)
|
1170 |
-
|
1171 |
added_cond_kwargs = {"text_embeds": add_text_embeds_input, "time_ids": add_time_ids_input}
|
1172 |
noise_pred = self.unet(
|
1173 |
latent_model_input,
|
@@ -1176,28 +1016,19 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
1176 |
added_cond_kwargs=added_cond_kwargs,
|
1177 |
return_dict=False,
|
1178 |
)[0]
|
1179 |
-
|
1180 |
if do_classifier_free_guidance:
|
1181 |
noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
|
1182 |
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
1183 |
-
|
1184 |
if do_classifier_free_guidance and guidance_rescale > 0.0:
|
1185 |
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
|
1186 |
-
|
1187 |
if shuffle:
|
1188 |
noise_pred = noise_pred.gather(0, original_index)
|
1189 |
-
|
1190 |
self.scheduler._init_step_index(t)
|
1191 |
latents_denoised_batch = self.scheduler.step(noise_pred, t, latents_for_view, **extra_step_kwargs, return_dict=False)[0]
|
1192 |
-
|
1193 |
-
for latents_view_denoised, (h, w) in zip(
|
1194 |
-
latents_denoised_batch.chunk(vb_size), batch_view
|
1195 |
-
):
|
1196 |
value_global[:, :, h::current_scale_num, w::current_scale_num] += latents_view_denoised
|
1197 |
count_global[:, :, h::current_scale_num, w::current_scale_num] += 1
|
1198 |
-
|
1199 |
value_global = value_global[:, :, h_pad:, w_pad:]
|
1200 |
-
|
1201 |
if use_multidiffusion:
|
1202 |
c2 = cosine_factor ** cosine_scale_2
|
1203 |
value += value_global * c2
|
@@ -1205,15 +1036,12 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
1205 |
else:
|
1206 |
value += value_global
|
1207 |
count += torch.ones_like(value_global)
|
1208 |
-
|
1209 |
latents = torch.where(count > 0, value / count, value)
|
1210 |
-
|
1211 |
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
1212 |
progress_bar.update()
|
1213 |
if callback is not None and i % callback_steps == 0:
|
1214 |
step_idx = i // getattr(self.scheduler, "order", 1)
|
1215 |
callback(step_idx, t, latents)
|
1216 |
-
|
1217 |
latents = (latents - latents.mean()) / latents.std() * anchor_std + anchor_mean
|
1218 |
if self.lowvram:
|
1219 |
latents = latents.cpu()
|
@@ -1224,29 +1052,23 @@ class AccDiffusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoade
|
|
1224 |
needs_upcasting = False
|
1225 |
self.unet.cpu()
|
1226 |
self.vae.to(device)
|
1227 |
-
|
1228 |
if needs_upcasting:
|
1229 |
self.upcast_vae()
|
1230 |
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
|
1231 |
-
|
1232 |
print("### Phase {} Decoding ###".format(current_scale_num))
|
1233 |
if current_height > 2048 or current_width > 2048:
|
1234 |
self.enable_vae_tiling()
|
1235 |
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
1236 |
else:
|
1237 |
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
1238 |
-
|
1239 |
image = self.image_processor.postprocess(image, output_type=output_type)
|
1240 |
image[0].save(f'{result_path}/AccDiffusion_{current_scale_num}.png')
|
1241 |
output_images.append(image[0])
|
1242 |
-
|
1243 |
if needs_upcasting:
|
1244 |
self.vae.to(dtype=torch.float16)
|
1245 |
else:
|
1246 |
image = latents
|
1247 |
-
|
1248 |
self.maybe_free_model_hooks()
|
1249 |
-
|
1250 |
return output_images
|
1251 |
|
1252 |
|
@@ -1280,9 +1102,7 @@ if __name__ == "__main__":
|
|
1280 |
## others ##
|
1281 |
parser.add_argument('--debug', default=False, action='store_true')
|
1282 |
parser.add_argument('--experiment_name', default="AccDiffusion")
|
1283 |
-
|
1284 |
args = parser.parse_args()
|
1285 |
-
|
1286 |
pipe = AccDiffusionSDXLPipeline.from_pretrained(args.model_ckpt, torch_dtype=torch.float16).to("cuda")
|
1287 |
|
1288 |
@spaces.GPU(duration=200)
|
@@ -1294,11 +1114,9 @@ if __name__ == "__main__":
|
|
1294 |
"n_cross_replace": {"default_": 1.0, "confetti": 0.8},
|
1295 |
}
|
1296 |
generator = torch.Generator(device='cuda').manual_seed(seed)
|
1297 |
-
|
1298 |
print(f"Prompt: {prompt}")
|
1299 |
md5_hash = hashlib.md5(prompt.encode()).hexdigest()
|
1300 |
result_path = f"./output/{args.experiment_name}/{md5_hash}/{width}_{height}_{seed}/"
|
1301 |
-
|
1302 |
images = pipe(
|
1303 |
prompt,
|
1304 |
negative_prompt=args.negative_prompt,
|
@@ -1326,11 +1144,9 @@ if __name__ == "__main__":
|
|
1326 |
debug=args.debug, save_attention_map=args.save_attention_map, use_md_prompt=use_md_prompt, c=args.c
|
1327 |
)
|
1328 |
print(images)
|
1329 |
-
|
1330 |
return images
|
1331 |
|
1332 |
MAX_SEED = np.iinfo(np.int32).max
|
1333 |
-
|
1334 |
css = """
|
1335 |
body {
|
1336 |
background: linear-gradient(135deg, #2c3e50, #4ca1af);
|
@@ -1353,7 +1169,6 @@ if __name__ == "__main__":
|
|
1353 |
visibility: hidden;
|
1354 |
}
|
1355 |
"""
|
1356 |
-
|
1357 |
with gr.Blocks(css=css) as demo:
|
1358 |
with gr.Column(elem_id="col-container"):
|
1359 |
gr.Markdown("<h1>AccDiffusion: Advanced AI Art Generator</h1>")
|
@@ -1363,7 +1178,6 @@ if __name__ == "__main__":
|
|
1363 |
with gr.Row():
|
1364 |
prompt = gr.Textbox(label="Prompt", placeholder="예: A surreal landscape with floating islands and vibrant colors.", lines=2, scale=4)
|
1365 |
submit_btn = gr.Button("Generate", scale=1)
|
1366 |
-
|
1367 |
with gr.Accordion("Advanced Settings", open=False):
|
1368 |
with gr.Row():
|
1369 |
resolution = gr.Radio(
|
@@ -1386,8 +1200,7 @@ if __name__ == "__main__":
|
|
1386 |
use_progressive_upscaling = gr.Checkbox(label="Use Progressive Upscaling", value=False)
|
1387 |
shuffle = gr.Checkbox(label="Shuffle", value=False)
|
1388 |
use_md_prompt = gr.Checkbox(label="Use MD Prompt", value=False)
|
1389 |
-
|
1390 |
-
output_images = gr.Gallery(label="Output Images", format="png").style(grid=[2], height="auto")
|
1391 |
gr.Markdown("### Example Prompts")
|
1392 |
gr.Examples(
|
1393 |
examples = [
|
|
|
71 |
gaussian_1d = gaussian_1d / gaussian_1d.sum()
|
72 |
gaussian_2d = gaussian_1d[:, None] * gaussian_1d[None, :]
|
73 |
kernel = gaussian_2d[None, None, :, :].repeat(channels, 1, 1, 1)
|
|
|
74 |
return kernel
|
75 |
|
76 |
def gaussian_filter(latents, kernel_size=3, sigma=1.0):
|
|
|
87 |
"""
|
88 |
std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
|
89 |
std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
|
|
|
90 |
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
|
|
|
91 |
noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
|
92 |
return noise_cfg
|
93 |
|
|
|
141 |
add_watermarker: Optional[bool] = None,
|
142 |
):
|
143 |
super().__init__()
|
|
|
144 |
self.register_modules(
|
145 |
vae=vae,
|
146 |
text_encoder=text_encoder,
|
|
|
154 |
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
|
155 |
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
|
156 |
self.default_sample_size = self.unet.config.sample_size
|
|
|
157 |
add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
|
|
|
158 |
if add_watermarker:
|
159 |
self.watermark = StableDiffusionXLWatermarker()
|
160 |
else:
|
161 |
self.watermark = None
|
162 |
|
|
|
163 |
def enable_vae_slicing(self):
|
164 |
self.vae.enable_slicing()
|
165 |
|
|
|
166 |
def disable_vae_slicing(self):
|
167 |
self.vae.disable_slicing()
|
168 |
|
|
|
169 |
def enable_vae_tiling(self):
|
170 |
self.vae.enable_tiling()
|
171 |
|
|
|
172 |
def disable_vae_tiling(self):
|
173 |
self.vae.disable_tiling()
|
174 |
|
|
|
188 |
lora_scale: Optional[float] = None,
|
189 |
):
|
190 |
device = device or self._execution_device
|
|
|
|
|
191 |
if lora_scale is not None and isinstance(self, LoraLoaderMixin):
|
192 |
self._lora_scale = lora_scale
|
193 |
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
|
194 |
adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
|
|
|
195 |
if prompt is not None and isinstance(prompt, str):
|
196 |
batch_size = 1
|
197 |
elif prompt is not None and isinstance(prompt, list):
|
198 |
batch_size = len(prompt)
|
199 |
else:
|
200 |
batch_size = prompt_embeds.shape[0]
|
|
|
201 |
tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
|
202 |
text_encoders = (
|
203 |
[self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
|
204 |
)
|
|
|
205 |
if prompt_embeds is None:
|
206 |
prompt_2 = prompt_2 or prompt
|
207 |
prompt_embeds_list = []
|
|
|
209 |
for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
|
210 |
if isinstance(self, TextualInversionLoaderMixin):
|
211 |
prompt = self.maybe_convert_prompt(prompt, tokenizer)
|
|
|
212 |
text_inputs = tokenizer(
|
213 |
prompt,
|
214 |
padding="max_length",
|
|
|
216 |
truncation=True,
|
217 |
return_tensors="pt",
|
218 |
)
|
|
|
219 |
text_input_ids = text_inputs.input_ids
|
220 |
untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
|
|
|
221 |
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
|
222 |
text_input_ids, untruncated_ids
|
223 |
):
|
|
|
226 |
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
227 |
f" {tokenizer.model_max_length} tokens: {removed_text}"
|
228 |
)
|
|
|
229 |
prompt_embeds = text_encoder(
|
230 |
text_input_ids.to(device),
|
231 |
output_hidden_states=True,
|
232 |
)
|
|
|
233 |
pooled_prompt_embeds = prompt_embeds[0]
|
234 |
prompt_embeds = prompt_embeds.hidden_states[-2]
|
|
|
235 |
prompt_embeds_list.append(prompt_embeds)
|
|
|
236 |
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
|
|
|
237 |
zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
|
238 |
if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
|
239 |
negative_prompt_embeds = torch.zeros_like(prompt_embeds)
|
|
|
241 |
elif do_classifier_free_guidance and negative_prompt_embeds is None:
|
242 |
negative_prompt = negative_prompt or ""
|
243 |
negative_prompt_2 = negative_prompt_2 or negative_prompt
|
|
|
244 |
if prompt is not None and type(prompt) is not type(negative_prompt):
|
245 |
raise TypeError(
|
246 |
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
|
|
256 |
)
|
257 |
else:
|
258 |
uncond_tokens = [negative_prompt, negative_prompt_2]
|
|
|
259 |
negative_prompt_embeds_list = []
|
260 |
for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
|
261 |
if isinstance(self, TextualInversionLoaderMixin):
|
262 |
negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
|
|
|
263 |
max_length = prompt_embeds.shape[1]
|
264 |
uncond_input = tokenizer(
|
265 |
negative_prompt,
|
|
|
268 |
truncation=True,
|
269 |
return_tensors="pt",
|
270 |
)
|
|
|
271 |
negative_prompt_embeds = text_encoder(
|
272 |
uncond_input.input_ids.to(device),
|
273 |
output_hidden_states=True,
|
274 |
)
|
275 |
negative_pooled_prompt_embeds = negative_prompt_embeds[0]
|
276 |
negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
|
|
|
277 |
negative_prompt_embeds_list.append(negative_prompt_embeds)
|
|
|
278 |
negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
|
|
|
279 |
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
|
280 |
bs_embed, seq_len, _ = prompt_embeds.shape
|
281 |
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
282 |
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
|
|
283 |
if do_classifier_free_guidance:
|
284 |
seq_len = negative_prompt_embeds.shape[1]
|
285 |
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
|
286 |
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
287 |
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
|
|
288 |
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
|
289 |
bs_embed * num_images_per_prompt, -1
|
290 |
)
|
|
|
292 |
negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
|
293 |
bs_embed * num_images_per_prompt, -1
|
294 |
)
|
|
|
295 |
return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
|
296 |
|
297 |
def prepare_extra_step_kwargs(self, generator, eta):
|
|
|
299 |
extra_step_kwargs = {}
|
300 |
if accepts_eta:
|
301 |
extra_step_kwargs["eta"] = eta
|
|
|
302 |
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
303 |
if accepts_generator:
|
304 |
extra_step_kwargs["generator"] = generator
|
|
|
321 |
):
|
322 |
if height % 8 != 0 or width % 8 != 0:
|
323 |
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
|
|
|
324 |
if (callback_steps is None) or (
|
325 |
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
|
326 |
):
|
327 |
raise ValueError(
|
328 |
f"`callback_steps` has to be a positive integer but is {callback_steps} of type {type(callback_steps)}."
|
329 |
)
|
|
|
330 |
if prompt is not None and prompt_embeds is not None:
|
331 |
raise ValueError(
|
332 |
f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to only forward one of the two."
|
|
|
343 |
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
344 |
elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
|
345 |
raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
|
|
|
346 |
if negative_prompt is not None and negative_prompt_embeds is not None:
|
347 |
raise ValueError(
|
348 |
f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to only forward one of the two."
|
|
|
351 |
raise ValueError(
|
352 |
f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to only forward one of the two."
|
353 |
)
|
|
|
354 |
if prompt_embeds is not None and negative_prompt_embeds is not None:
|
355 |
if prompt_embeds.shape != negative_prompt_embeds.shape:
|
356 |
raise ValueError(
|
357 |
"`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds` {negative_prompt_embeds.shape}."
|
358 |
)
|
|
|
359 |
if prompt_embeds is not None and pooled_prompt_embeds is None:
|
360 |
raise ValueError(
|
361 |
"If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
|
362 |
)
|
|
|
363 |
if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
|
364 |
raise ValueError(
|
365 |
"If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
|
366 |
)
|
|
|
367 |
if max(height, width) % 1024 != 0:
|
368 |
raise ValueError(f"the larger one of `height` and `width` has to be divisible by 1024 but are {height} and {width}.")
|
|
|
369 |
if num_images_per_prompt != 1:
|
370 |
warnings.warn("num_images_per_prompt != 1 is not supported by AccDiffusion and will be ignored.")
|
371 |
num_images_per_prompt = 1
|
|
|
376 |
raise ValueError(
|
377 |
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch size of {batch_size}. Make sure the batch size matches the length of the generators."
|
378 |
)
|
|
|
379 |
if latents is None:
|
380 |
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
381 |
else:
|
382 |
latents = latents.to(device)
|
|
|
383 |
latents = latents * self.scheduler.init_noise_sigma
|
384 |
return latents
|
385 |
|
386 |
def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
|
387 |
add_time_ids = list(original_size + crops_coords_top_left + target_size)
|
|
|
388 |
passed_add_embed_dim = (
|
389 |
self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
|
390 |
)
|
391 |
expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
|
|
|
392 |
if expected_add_embed_dim != passed_add_embed_dim:
|
393 |
raise ValueError(
|
394 |
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. \
|
395 |
The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
|
396 |
)
|
|
|
397 |
add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
|
398 |
return add_time_ids
|
399 |
|
|
|
409 |
h_end = h_start + window_size
|
410 |
w_start = int((i % num_blocks_width) * stride)
|
411 |
w_end = w_start + window_size
|
|
|
412 |
if h_end > height:
|
413 |
h_start = int(h_start + height - h_end)
|
414 |
h_end = int(height)
|
|
|
421 |
if w_start < 0:
|
422 |
w_end = int(w_end - w_start)
|
423 |
w_start = 0
|
|
|
424 |
if random_jitter:
|
425 |
jitter_range = (window_size - stride) // 4
|
426 |
w_jitter = 0
|
|
|
431 |
w_jitter = random.randint(-jitter_range, 0)
|
432 |
elif (w_start != 0) and (w_end == width):
|
433 |
w_jitter = random.randint(0, jitter_range)
|
|
|
434 |
if (h_start != 0) and (h_end != height):
|
435 |
h_jitter = random.randint(-jitter_range, jitter_range)
|
436 |
elif (h_start == 0) and (h_end != height):
|
|
|
441 |
h_end = h_end + h_jitter + jitter_range
|
442 |
w_start = w_start + w_jitter + jitter_range
|
443 |
w_end = w_end + w_jitter + jitter_range
|
|
|
444 |
views.append((h_start, h_end, w_start, w_end))
|
445 |
return views
|
446 |
|
|
|
476 |
continue
|
477 |
cross_att_count += 1
|
478 |
attn_procs[name] = P2PCrossAttnProcessor(controller=controller, place_in_unet=place_in_unet)
|
|
|
479 |
self.unet.set_attn_processor(attn_procs)
|
480 |
controller.num_att_layers = cross_att_count
|
481 |
return ori_attn_processors
|
|
|
488 |
from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
|
489 |
else:
|
490 |
raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
|
|
|
491 |
is_model_cpu_offload = False
|
492 |
is_sequential_cpu_offload = False
|
493 |
recursive = False
|
|
|
507 |
**kwargs,
|
508 |
)
|
509 |
self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet)
|
|
|
510 |
text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
|
511 |
if len(text_encoder_state_dict) > 0:
|
512 |
self.load_lora_into_text_encoder(
|
|
|
516 |
prefix="text_encoder",
|
517 |
lora_scale=self.lora_scale,
|
518 |
)
|
|
|
519 |
text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
|
520 |
if len(text_encoder_2_state_dict) > 0:
|
521 |
self.load_lora_into_text_encoder(
|
|
|
525 |
prefix="text_encoder_2",
|
526 |
lora_scale=self.lora_scale,
|
527 |
)
|
|
|
528 |
if is_model_cpu_offload:
|
529 |
self.enable_model_cpu_offload()
|
530 |
elif is_sequential_cpu_offload:
|
|
|
543 |
safe_serialization: bool = True,
|
544 |
):
|
545 |
state_dict = {}
|
|
|
546 |
def pack_weights(layers, prefix):
|
547 |
layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
|
548 |
layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
|
549 |
return layers_state_dict
|
|
|
550 |
if not (unet_lora_layers or text_encoder_lora_layers or text_encoder_2_lora_layers):
|
551 |
raise ValueError(
|
552 |
"You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers` or `text_encoder_2_lora_layers`."
|
553 |
)
|
|
|
554 |
if unet_lora_layers:
|
555 |
state_dict.update(pack_weights(unet_lora_layers, "unet"))
|
|
|
556 |
if text_encoder_lora_layers and text_encoder_2_lora_layers:
|
557 |
state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder"))
|
558 |
state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))
|
|
|
559 |
self.write_lora_layers(
|
560 |
state_dict=state_dict,
|
561 |
save_directory=save_directory,
|
|
|
636 |
|
637 |
Examples:
|
638 |
"""
|
|
|
639 |
if debug:
|
640 |
num_inference_steps = 1
|
|
|
641 |
height = height or self.default_sample_size * self.vae_scale_factor
|
642 |
width = width or self.default_sample_size * self.vae_scale_factor
|
|
|
643 |
x1_size = self.default_sample_size * self.vae_scale_factor
|
|
|
644 |
height_scale = height / x1_size
|
645 |
width_scale = width / x1_size
|
646 |
scale_num = int(max(height_scale, width_scale))
|
647 |
aspect_ratio = min(height_scale, width_scale) / max(height_scale, width_scale)
|
|
|
648 |
original_size = original_size or (height, width)
|
649 |
target_size = target_size or (height, width)
|
|
|
650 |
if attn_res is None:
|
651 |
+
attn_res = (int(np.ceil(self.default_sample_size * self.vae_scale_factor / 32)), int(np.ceil(self.default_sample_size * self.vae_scale_factor / 32)))
|
652 |
self.attn_res = attn_res
|
|
|
653 |
if lowvram:
|
654 |
attention_map_device = torch.device("cpu")
|
655 |
else:
|
656 |
attention_map_device = self.device
|
|
|
657 |
self.controller = create_controller(
|
658 |
prompt, cross_attention_kwargs, num_inference_steps, tokenizer=self.tokenizer, device=attention_map_device, attn_res=self.attn_res
|
659 |
)
|
|
|
660 |
if save_attention_map or use_md_prompt:
|
661 |
ori_attn_processors = self.register_attention_control(self.controller)
|
|
|
662 |
self.check_inputs(
|
663 |
prompt,
|
664 |
prompt_2,
|
|
|
673 |
negative_pooled_prompt_embeds,
|
674 |
num_images_per_prompt,
|
675 |
)
|
|
|
676 |
if prompt is not None and isinstance(prompt, str):
|
677 |
batch_size = 1
|
678 |
elif prompt is not None and isinstance(prompt, list):
|
679 |
batch_size = len(prompt)
|
680 |
else:
|
681 |
batch_size = prompt_embeds.shape[0]
|
|
|
682 |
device = self._execution_device
|
683 |
self.lowvram = lowvram
|
684 |
if self.lowvram:
|
|
|
686 |
self.unet.cpu()
|
687 |
self.text_encoder.to(device)
|
688 |
self.text_encoder_2.to(device)
|
|
|
689 |
do_classifier_free_guidance = guidance_scale > 1.0
|
690 |
+
text_encoder_lora_scale = (cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None)
|
|
|
|
|
|
|
|
|
691 |
(
|
692 |
prompt_embeds,
|
693 |
negative_prompt_embeds,
|
|
|
707 |
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
708 |
lora_scale=text_encoder_lora_scale,
|
709 |
)
|
|
|
710 |
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
711 |
timesteps = self.scheduler.timesteps
|
|
|
712 |
num_channels_latents = self.unet.config.in_channels
|
713 |
latents = self.prepare_latents(
|
714 |
batch_size * num_images_per_prompt,
|
|
|
720 |
generator,
|
721 |
latents,
|
722 |
)
|
|
|
723 |
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
|
|
724 |
add_text_embeds = pooled_prompt_embeds
|
|
|
725 |
add_time_ids = self._get_add_time_ids(
|
726 |
original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
|
727 |
)
|
|
|
728 |
if negative_original_size is not None and negative_target_size is not None:
|
729 |
negative_add_time_ids = self._get_add_time_ids(
|
730 |
negative_original_size,
|
|
|
734 |
)
|
735 |
else:
|
736 |
negative_add_time_ids = add_time_ids
|
|
|
737 |
if do_classifier_free_guidance:
|
738 |
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0).to(device)
|
739 |
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0).to(device)
|
740 |
add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0).to(device).repeat(batch_size * num_images_per_prompt, 1)
|
|
|
741 |
del negative_prompt_embeds, negative_pooled_prompt_embeds, negative_add_time_ids
|
|
|
742 |
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
|
|
743 |
if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
|
744 |
discrete_timestep_cutoff = int(
|
745 |
round(
|
|
|
748 |
)
|
749 |
num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
|
750 |
timesteps = timesteps[:num_inference_steps]
|
|
|
751 |
output_images = []
|
|
|
752 |
###################################################### Phase Initialization ########################################################
|
|
|
753 |
if self.lowvram:
|
754 |
self.text_encoder.cpu()
|
755 |
self.text_encoder_2.cpu()
|
|
|
756 |
if image_lr == None:
|
757 |
print("### Phase 1 Denoising ###")
|
758 |
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
759 |
for i, t in enumerate(timesteps):
|
|
|
760 |
if self.lowvram:
|
761 |
self.vae.cpu()
|
762 |
self.unet.to(device)
|
|
|
763 |
latents_for_view = latents
|
|
|
764 |
latent_model_input = (
|
765 |
latents.repeat_interleave(2, dim=0)
|
766 |
if do_classifier_free_guidance
|
767 |
else latents
|
768 |
)
|
769 |
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
|
|
770 |
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
|
|
|
771 |
noise_pred = self.unet(
|
772 |
latent_model_input,
|
773 |
t,
|
|
|
775 |
added_cond_kwargs=added_cond_kwargs,
|
776 |
return_dict=False,
|
777 |
)[0]
|
|
|
778 |
if do_classifier_free_guidance:
|
779 |
noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
|
780 |
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
|
|
781 |
if do_classifier_free_guidance and guidance_rescale > 0.0:
|
782 |
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
|
|
|
783 |
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
|
|
784 |
if t == 1 and use_md_prompt:
|
785 |
md_prompts, views_attention = get_multidiffusion_prompts(tokenizer=self.tokenizer, prompts=[prompt], threthod=c, attention_store=self.controller, height=height//scale_num, width=width//scale_num, from_where=["up","down"], random_jitter=True, scale_num=scale_num)
|
|
|
786 |
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
787 |
progress_bar.update()
|
788 |
if callback is not None and i % callback_steps == 0:
|
789 |
step_idx = i // getattr(self.scheduler, "order", 1)
|
790 |
callback(step_idx, t, latents)
|
|
|
791 |
del latents_for_view, latent_model_input, noise_pred, noise_pred_text, noise_pred_uncond
|
792 |
if use_md_prompt or save_attention_map:
|
793 |
self.recover_attention_control(ori_attn_processors=ori_attn_processors)
|
|
|
797 |
print("### Encoding Real Image ###")
|
798 |
latents = self.vae.encode(image_lr)
|
799 |
latents = latents.latent_dist.sample() * self.vae.config.scaling_factor
|
|
|
800 |
anchor_mean = latents.mean()
|
801 |
anchor_std = latents.std()
|
802 |
if self.lowvram:
|
|
|
804 |
torch.cuda.empty_cache()
|
805 |
if not output_type == "latent":
|
806 |
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
|
|
|
807 |
if self.lowvram:
|
808 |
needs_upcasting = False
|
809 |
self.unet.cpu()
|
810 |
self.vae.to(device)
|
|
|
811 |
if needs_upcasting:
|
812 |
self.upcast_vae()
|
813 |
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
|
|
|
818 |
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
819 |
if needs_upcasting:
|
820 |
self.vae.to(dtype=torch.float16)
|
|
|
821 |
image = self.image_processor.postprocess(image, output_type=output_type)
|
822 |
if not os.path.exists(f'{result_path}'):
|
823 |
os.makedirs(f'{result_path}')
|
|
|
824 |
image_lr_save_path = f'{result_path}/{image[0].size[0]}_{image[0].size[1]}.png'
|
825 |
image[0].save(image_lr_save_path)
|
826 |
output_images.append(image[0])
|
|
|
827 |
####################################################### Phase Upscaling #####################################################
|
828 |
if use_progressive_upscaling:
|
829 |
if image_lr == None:
|
|
|
832 |
starting_scale = 1
|
833 |
else:
|
834 |
starting_scale = scale_num
|
|
|
835 |
for current_scale_num in range(starting_scale, scale_num + 1):
|
836 |
if self.lowvram:
|
837 |
latents = latents.to(device)
|
838 |
self.unet.to(device)
|
839 |
torch.cuda.empty_cache()
|
|
|
840 |
current_height = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
|
841 |
current_width = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
|
|
|
842 |
if height > width:
|
843 |
current_width = int(current_width * aspect_ratio)
|
844 |
else:
|
845 |
current_height = int(current_height * aspect_ratio)
|
|
|
846 |
if upscale_mode == "bicubic_latent" or debug:
|
847 |
latents = F.interpolate(latents.to(device), size=(int(current_height / self.vae_scale_factor), int(current_width / self.vae_scale_factor)), mode='bicubic')
|
848 |
else:
|
849 |
raise NotImplementedError
|
|
|
850 |
print("### Phase {} Denoising ###".format(current_scale_num))
|
851 |
noise_latents = []
|
852 |
noise = torch.randn_like(latents)
|
|
|
854 |
noise_latent = self.scheduler.add_noise(latents, noise, timestep.unsqueeze(0))
|
855 |
noise_latents.append(noise_latent)
|
856 |
latents = noise_latents[0]
|
|
|
857 |
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
858 |
for i, t in enumerate(timesteps):
|
859 |
count = torch.zeros_like(latents)
|
|
|
862 |
if use_skip_residual:
|
863 |
c1 = cosine_factor ** cosine_scale_1
|
864 |
latents = latents * (1 - c1) + noise_latents[i] * c1
|
|
|
865 |
if use_multidiffusion:
|
866 |
if use_md_prompt:
|
867 |
md_prompt_embeds_list = []
|
|
|
889 |
md_prompt_embeds_list.append(torch.cat([md_negative_prompt_embeds, md_prompt_embeds], dim=0).to(device))
|
890 |
md_add_text_embeds_list.append(torch.cat([md_negative_pooled_prompt_embeds, md_pooled_prompt_embeds], dim=0).to(device))
|
891 |
del md_negative_prompt_embeds, md_negative_pooled_prompt_embeds
|
|
|
892 |
if use_md_prompt:
|
893 |
random_jitter = True
|
894 |
views = [(h_start*4, h_end*4, w_start*4, w_end*4) for h_start, h_end, w_start, w_end in views_attention[current_scale_num]]
|
895 |
else:
|
896 |
random_jitter = True
|
897 |
views = self.get_views(current_height, current_width, stride=stride, window_size=self.unet.config.sample_size, random_jitter=random_jitter)
|
|
|
898 |
views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
|
|
|
899 |
if use_md_prompt:
|
900 |
views_prompt_embeds_input = [md_prompt_embeds_list[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
|
901 |
views_add_text_embeds_input = [md_add_text_embeds_list[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
|
|
|
902 |
if random_jitter:
|
903 |
jitter_range = int((self.unet.config.sample_size - stride) // 4)
|
904 |
latents_ = F.pad(latents, (jitter_range, jitter_range, jitter_range, jitter_range), 'constant', 0)
|
905 |
else:
|
906 |
latents_ = latents
|
|
|
907 |
count_local = torch.zeros_like(latents_)
|
908 |
value_local = torch.zeros_like(latents_)
|
|
|
909 |
for j, batch_view in enumerate(views_batch):
|
910 |
vb_size = len(batch_view)
|
911 |
latents_for_view = torch.cat(
|
912 |
+
[latents_[:, :, h_start:h_end, w_start:w_end] for h_start, h_end, w_start, w_end in batch_view]
|
|
|
|
|
|
|
913 |
)
|
|
|
914 |
latent_model_input = latents_for_view
|
915 |
+
latent_model_input = (latent_model_input.repeat_interleave(2, dim=0)
|
916 |
+
if do_classifier_free_guidance
|
917 |
+
else latent_model_input)
|
|
|
|
|
918 |
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
|
|
919 |
add_time_ids_input = []
|
920 |
for h_start, h_end, w_start, w_end in batch_view:
|
921 |
add_time_ids_ = add_time_ids.clone()
|
|
|
923 |
add_time_ids_[:, 3] = w_start * self.vae_scale_factor
|
924 |
add_time_ids_input.append(add_time_ids_)
|
925 |
add_time_ids_input = torch.cat(add_time_ids_input)
|
|
|
926 |
if not use_md_prompt:
|
927 |
prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
|
928 |
add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
|
|
|
945 |
added_cond_kwargs=md_added_cond_kwargs,
|
946 |
return_dict=False,
|
947 |
)[0]
|
|
|
948 |
if do_classifier_free_guidance:
|
949 |
noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
|
950 |
noise_pred = noise_pred_uncond + multi_guidance_scale * (noise_pred_text - noise_pred_uncond)
|
|
|
951 |
if do_classifier_free_guidance and guidance_rescale > 0.0:
|
952 |
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
|
|
|
953 |
self.scheduler._init_step_index(t)
|
954 |
+
latents_denoised_batch = self.scheduler.step(noise_pred, t, latents_for_view, **extra_step_kwargs, return_dict=False)[0]
|
955 |
+
for latents_view_denoised, (h_start, h_end, w_start, w_end) in zip(latents_denoised_batch.chunk(vb_size), batch_view):
|
|
|
|
|
|
|
|
|
956 |
value_local[:, :, h_start:h_end, w_start:w_end] += latents_view_denoised
|
957 |
count_local[:, :, h_start:h_end, w_start:w_end] += 1
|
|
|
958 |
if random_jitter:
|
959 |
value_local = value_local[:, :, jitter_range:jitter_range + current_height // self.vae_scale_factor, jitter_range:jitter_range + current_width // self.vae_scale_factor]
|
960 |
count_local = count_local[:, :, jitter_range:jitter_range + current_height // self.vae_scale_factor, jitter_range:jitter_range + current_width // self.vae_scale_factor]
|
|
|
961 |
noise_index = i + 1 if i != (len(timesteps) - 1) else i
|
|
|
962 |
value_local = torch.where(count_local == 0, noise_latents[noise_index], value_local)
|
963 |
count_local = torch.where(count_local == 0, torch.ones_like(count_local), count_local)
|
964 |
if use_dilated_sampling:
|
|
|
967 |
count += torch.ones_like(value_local) * (1 - c2)
|
968 |
else:
|
969 |
value += value_local / count_local
|
970 |
+
count += torch.ones_like(value_local)
|
|
|
971 |
if use_dilated_sampling:
|
972 |
views = [[h, w] for h in range(current_scale_num) for w in range(current_scale_num)]
|
973 |
views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
|
|
|
974 |
h_pad = (current_scale_num - (latents.size(2) % current_scale_num)) % current_scale_num
|
975 |
w_pad = (current_scale_num - (latents.size(3) % current_scale_num)) % current_scale_num
|
976 |
latents_ = F.pad(latents, (w_pad, 0, h_pad, 0), 'constant', 0)
|
|
|
977 |
count_global = torch.zeros_like(latents_)
|
978 |
value_global = torch.zeros_like(latents_)
|
|
|
979 |
if use_guassian:
|
980 |
c3 = 0.99 * cosine_factor ** cosine_scale_3 + 1e-2
|
981 |
std_, mean_ = latents_.std(), latents_.mean()
|
|
|
983 |
latents_gaussian = (latents_gaussian - latents_gaussian.mean()) / latents_gaussian.std() * std_ + mean_
|
984 |
else:
|
985 |
latents_gaussian = latents_
|
|
|
986 |
for j, batch_view in enumerate(views_batch):
|
|
|
987 |
latents_for_view = torch.cat(
|
988 |
+
[latents_[:, :, h::current_scale_num, w::current_scale_num] for h, w in batch_view]
|
|
|
|
|
|
|
989 |
)
|
|
|
990 |
latents_for_view_gaussian = torch.cat(
|
991 |
+
[latents_gaussian[:, :, h::current_scale_num, w::current_scale_num] for h, w in batch_view]
|
|
|
|
|
|
|
992 |
)
|
|
|
993 |
if shuffle:
|
994 |
shape = latents_for_view.shape
|
995 |
+
# 수정: range(...) 괄호 추가
|
996 |
shuffle_index = torch.stack([torch.randperm(shape[0]) for _ in range(latents_for_view.reshape(-1).shape[0]//shape[0])])
|
997 |
shuffle_index = shuffle_index.view(shape[1], shape[2], shape[3], shape[0])
|
998 |
original_index = torch.zeros_like(shuffle_index).scatter_(3, shuffle_index, torch.arange(shape[0]).repeat(shape[1], shape[2], shape[3], 1))
|
999 |
shuffle_index = shuffle_index.permute(3, 0, 1, 2).to(device)
|
1000 |
original_index = original_index.permute(3, 0, 1, 2).to(device)
|
1001 |
latents_for_view_gaussian = latents_for_view_gaussian.gather(0, shuffle_index)
|
|
|
1002 |
vb_size = latents_for_view.size(0)
|
|
|
1003 |
latent_model_input = latents_for_view_gaussian
|
1004 |
+
latent_model_input = (latent_model_input.repeat_interleave(2, dim=0)
|
1005 |
+
if do_classifier_free_guidance
|
1006 |
+
else latent_model_input)
|
|
|
|
|
1007 |
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
|
|
1008 |
prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
|
1009 |
add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
|
1010 |
add_time_ids_input = torch.cat([add_time_ids] * vb_size)
|
|
|
1011 |
added_cond_kwargs = {"text_embeds": add_text_embeds_input, "time_ids": add_time_ids_input}
|
1012 |
noise_pred = self.unet(
|
1013 |
latent_model_input,
|
|
|
1016 |
added_cond_kwargs=added_cond_kwargs,
|
1017 |
return_dict=False,
|
1018 |
)[0]
|
|
|
1019 |
if do_classifier_free_guidance:
|
1020 |
noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
|
1021 |
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
|
|
1022 |
if do_classifier_free_guidance and guidance_rescale > 0.0:
|
1023 |
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
|
|
|
1024 |
if shuffle:
|
1025 |
noise_pred = noise_pred.gather(0, original_index)
|
|
|
1026 |
self.scheduler._init_step_index(t)
|
1027 |
latents_denoised_batch = self.scheduler.step(noise_pred, t, latents_for_view, **extra_step_kwargs, return_dict=False)[0]
|
1028 |
+
for latents_view_denoised, (h, w) in zip(latents_denoised_batch.chunk(vb_size), batch_view):
|
|
|
|
|
|
|
1029 |
value_global[:, :, h::current_scale_num, w::current_scale_num] += latents_view_denoised
|
1030 |
count_global[:, :, h::current_scale_num, w::current_scale_num] += 1
|
|
|
1031 |
value_global = value_global[:, :, h_pad:, w_pad:]
|
|
|
1032 |
if use_multidiffusion:
|
1033 |
c2 = cosine_factor ** cosine_scale_2
|
1034 |
value += value_global * c2
|
|
|
1036 |
else:
|
1037 |
value += value_global
|
1038 |
count += torch.ones_like(value_global)
|
|
|
1039 |
latents = torch.where(count > 0, value / count, value)
|
|
|
1040 |
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
1041 |
progress_bar.update()
|
1042 |
if callback is not None and i % callback_steps == 0:
|
1043 |
step_idx = i // getattr(self.scheduler, "order", 1)
|
1044 |
callback(step_idx, t, latents)
|
|
|
1045 |
latents = (latents - latents.mean()) / latents.std() * anchor_std + anchor_mean
|
1046 |
if self.lowvram:
|
1047 |
latents = latents.cpu()
|
|
|
1052 |
needs_upcasting = False
|
1053 |
self.unet.cpu()
|
1054 |
self.vae.to(device)
|
|
|
1055 |
if needs_upcasting:
|
1056 |
self.upcast_vae()
|
1057 |
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
|
|
|
1058 |
print("### Phase {} Decoding ###".format(current_scale_num))
|
1059 |
if current_height > 2048 or current_width > 2048:
|
1060 |
self.enable_vae_tiling()
|
1061 |
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
1062 |
else:
|
1063 |
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
|
|
1064 |
image = self.image_processor.postprocess(image, output_type=output_type)
|
1065 |
image[0].save(f'{result_path}/AccDiffusion_{current_scale_num}.png')
|
1066 |
output_images.append(image[0])
|
|
|
1067 |
if needs_upcasting:
|
1068 |
self.vae.to(dtype=torch.float16)
|
1069 |
else:
|
1070 |
image = latents
|
|
|
1071 |
self.maybe_free_model_hooks()
|
|
|
1072 |
return output_images
|
1073 |
|
1074 |
|
|
|
1102 |
## others ##
|
1103 |
parser.add_argument('--debug', default=False, action='store_true')
|
1104 |
parser.add_argument('--experiment_name', default="AccDiffusion")
|
|
|
1105 |
args = parser.parse_args()
|
|
|
1106 |
pipe = AccDiffusionSDXLPipeline.from_pretrained(args.model_ckpt, torch_dtype=torch.float16).to("cuda")
|
1107 |
|
1108 |
@spaces.GPU(duration=200)
|
|
|
1114 |
"n_cross_replace": {"default_": 1.0, "confetti": 0.8},
|
1115 |
}
|
1116 |
generator = torch.Generator(device='cuda').manual_seed(seed)
|
|
|
1117 |
print(f"Prompt: {prompt}")
|
1118 |
md5_hash = hashlib.md5(prompt.encode()).hexdigest()
|
1119 |
result_path = f"./output/{args.experiment_name}/{md5_hash}/{width}_{height}_{seed}/"
|
|
|
1120 |
images = pipe(
|
1121 |
prompt,
|
1122 |
negative_prompt=args.negative_prompt,
|
|
|
1144 |
debug=args.debug, save_attention_map=args.save_attention_map, use_md_prompt=use_md_prompt, c=args.c
|
1145 |
)
|
1146 |
print(images)
|
|
|
1147 |
return images
|
1148 |
|
1149 |
MAX_SEED = np.iinfo(np.int32).max
|
|
|
1150 |
css = """
|
1151 |
body {
|
1152 |
background: linear-gradient(135deg, #2c3e50, #4ca1af);
|
|
|
1169 |
visibility: hidden;
|
1170 |
}
|
1171 |
"""
|
|
|
1172 |
with gr.Blocks(css=css) as demo:
|
1173 |
with gr.Column(elem_id="col-container"):
|
1174 |
gr.Markdown("<h1>AccDiffusion: Advanced AI Art Generator</h1>")
|
|
|
1178 |
with gr.Row():
|
1179 |
prompt = gr.Textbox(label="Prompt", placeholder="예: A surreal landscape with floating islands and vibrant colors.", lines=2, scale=4)
|
1180 |
submit_btn = gr.Button("Generate", scale=1)
|
|
|
1181 |
with gr.Accordion("Advanced Settings", open=False):
|
1182 |
with gr.Row():
|
1183 |
resolution = gr.Radio(
|
|
|
1200 |
use_progressive_upscaling = gr.Checkbox(label="Use Progressive Upscaling", value=False)
|
1201 |
shuffle = gr.Checkbox(label="Shuffle", value=False)
|
1202 |
use_md_prompt = gr.Checkbox(label="Use MD Prompt", value=False)
|
1203 |
+
output_images = gr.Gallery(label="Output Images", format="png")
|
|
|
1204 |
gr.Markdown("### Example Prompts")
|
1205 |
gr.Examples(
|
1206 |
examples = [
|