Spaces:
Running
on
Zero
Running
on
Zero
AlekseyCalvin
commited on
Update pipeline.py
Browse files- pipeline.py +1 -188
pipeline.py
CHANGED
@@ -66,7 +66,7 @@ def prepare_timesteps(
|
|
66 |
return timesteps, num_inference_steps
|
67 |
|
68 |
# FLUX pipeline function
|
69 |
-
class FluxWithCFGPipeline(
|
70 |
|
71 |
def __init__(
|
72 |
self,
|
@@ -244,193 +244,6 @@ class FluxWithCFGPipeline(StableDiffusion3Pipeline):
|
|
244 |
self.maybe_free_model_hooks()
|
245 |
torch.cuda.empty_cache()
|
246 |
|
247 |
-
def _decode_latents_to_image(self, latents, height, width, output_type, vae=None):
|
248 |
-
"""Decodes the given latents into an image."""
|
249 |
-
vae = vae or self.vae
|
250 |
-
latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
|
251 |
-
latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor
|
252 |
-
image = vae.decode(latents, return_dict=False)[0]
|
253 |
-
return self.image_processor.postprocess(image, output_type=output_type)[0]
|
254 |
-
|
255 |
-
class FluxWithCFGPipeline(StableDiffusion3Pipeline):
|
256 |
-
@torch.inference_mode()
|
257 |
-
def __init__(
|
258 |
-
self,
|
259 |
-
transformer: FluxTransformer2DModel,
|
260 |
-
scheduler: FlowMatchEulerDiscreteScheduler,
|
261 |
-
vae: AutoencoderKL,
|
262 |
-
text_encoder: CLIPTextModelWithProjection,
|
263 |
-
tokenizer: CLIPTokenizer,
|
264 |
-
tokenizer_2: T5TokenizerFast,
|
265 |
-
tokenizer_3: None,
|
266 |
-
text_encoder_2: T5EncoderModel,
|
267 |
-
text_encoder_3: None,
|
268 |
-
):
|
269 |
-
super().__init__()
|
270 |
-
|
271 |
-
self.register_modules(
|
272 |
-
vae=vae,
|
273 |
-
text_encoder=text_encoder,
|
274 |
-
text_encoder_2=text_encoder_2,
|
275 |
-
text_encoder_3=text_encoder_3,
|
276 |
-
tokenizer=tokenizer,
|
277 |
-
tokenizer_2=tokenizer_2,
|
278 |
-
tokenizer_3=tokenizer_3,
|
279 |
-
transformer=transformer,
|
280 |
-
scheduler=scheduler,
|
281 |
-
)
|
282 |
-
self.vae_scale_factor = (
|
283 |
-
2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 16
|
284 |
-
)
|
285 |
-
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
|
286 |
-
self.tokenizer_max_length = (
|
287 |
-
self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77
|
288 |
-
)
|
289 |
-
self.default_sample_size = 64
|
290 |
-
|
291 |
-
def generate_image(
|
292 |
-
self,
|
293 |
-
prompt: Union[str, List[str]] = None,
|
294 |
-
prompt_2: Optional[Union[str, List[str]]] = None,
|
295 |
-
height: Optional[int] = None,
|
296 |
-
width: Optional[int] = None,
|
297 |
-
negative_prompt: Optional[Union[str, List[str]]] = None,
|
298 |
-
negative_prompt_2: Optional[Union[str, List[str]]] = None,
|
299 |
-
num_inference_steps: int = 4,
|
300 |
-
timesteps: List[int] = None,
|
301 |
-
guidance_scale: float = 3.5,
|
302 |
-
num_images_per_prompt: Optional[int] = 1,
|
303 |
-
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
304 |
-
latents: Optional[torch.FloatTensor] = None,
|
305 |
-
prompt_embeds: Optional[torch.FloatTensor] = None,
|
306 |
-
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
307 |
-
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
308 |
-
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
309 |
-
output_type: Optional[str] = "pil",
|
310 |
-
return_dict: bool = True,
|
311 |
-
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
|
312 |
-
max_sequence_length: int = 300,
|
313 |
-
):
|
314 |
-
height = height or self.default_sample_size * self.vae_scale_factor
|
315 |
-
width = width or self.default_sample_size * self.vae_scale_factor
|
316 |
-
|
317 |
-
# 1. Check inputs
|
318 |
-
self.check_inputs(
|
319 |
-
prompt,
|
320 |
-
prompt_2,
|
321 |
-
negative_prompt,
|
322 |
-
height,
|
323 |
-
width,
|
324 |
-
prompt_embeds=prompt_embeds,
|
325 |
-
pooled_prompt_embeds=pooled_prompt_embeds,
|
326 |
-
max_sequence_length=max_sequence_length,
|
327 |
-
)
|
328 |
-
|
329 |
-
self._guidance_scale = guidance_scale
|
330 |
-
self._joint_attention_kwargs = joint_attention_kwargs
|
331 |
-
self._interrupt = False
|
332 |
-
|
333 |
-
# 2. Define call parameters
|
334 |
-
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
335 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
336 |
-
|
337 |
-
# 3. Encode prompt
|
338 |
-
lora_scale = joint_attention_kwargs.get("scale", None) if joint_attention_kwargs is not None else None
|
339 |
-
prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt(
|
340 |
-
prompt=prompt,
|
341 |
-
prompt_2=prompt_2,
|
342 |
-
prompt_embeds=prompt_embeds,
|
343 |
-
pooled_prompt_embeds=pooled_prompt_embeds,
|
344 |
-
device=device,
|
345 |
-
num_images_per_prompt=num_images_per_prompt,
|
346 |
-
max_sequence_length=max_sequence_length,
|
347 |
-
lora_scale=lora_scale,
|
348 |
-
)
|
349 |
-
negative_prompt_embeds, negative_pooled_prompt_embeds, negative_text_ids = self.encode_prompt(
|
350 |
-
prompt=negative_prompt,
|
351 |
-
prompt_2=negative_prompt_2,
|
352 |
-
prompt_embeds=negative_prompt_embeds,
|
353 |
-
pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
354 |
-
device=device,
|
355 |
-
num_images_per_prompt=num_images_per_prompt,
|
356 |
-
max_sequence_length=max_sequence_length,
|
357 |
-
lora_scale=lora_scale,
|
358 |
-
)
|
359 |
-
|
360 |
-
# 4. Prepare latent variables
|
361 |
-
num_channels_latents = self.transformer.config.in_channels // 4
|
362 |
-
latents, latent_image_ids = self.prepare_latents(
|
363 |
-
batch_size * num_images_per_prompt,
|
364 |
-
num_channels_latents,
|
365 |
-
height,
|
366 |
-
width,
|
367 |
-
prompt_embeds.dtype,
|
368 |
-
negative_prompt_embeds.dtype,
|
369 |
-
device,
|
370 |
-
generator,
|
371 |
-
latents,
|
372 |
-
)
|
373 |
-
|
374 |
-
# 5. Prepare timesteps
|
375 |
-
sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
|
376 |
-
image_seq_len = latents.shape[1]
|
377 |
-
mu = calculate_timestep_shift(image_seq_len)
|
378 |
-
timesteps, num_inference_steps = prepare_timesteps(
|
379 |
-
self.scheduler,
|
380 |
-
num_inference_steps,
|
381 |
-
device,
|
382 |
-
timesteps,
|
383 |
-
sigmas,
|
384 |
-
mu=mu,
|
385 |
-
)
|
386 |
-
self._num_timesteps = len(timesteps)
|
387 |
-
|
388 |
-
# Handle guidance
|
389 |
-
guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float16).expand(latents.shape[0]) if self.transformer.config.guidance_embeds else None
|
390 |
-
|
391 |
-
# 6. Denoising loop
|
392 |
-
for i, t in enumerate(timesteps):
|
393 |
-
if self.interrupt:
|
394 |
-
continue
|
395 |
-
|
396 |
-
timestep = t.expand(latents.shape[0]).to(latents.dtype)
|
397 |
-
|
398 |
-
noise_pred = self.transformer(
|
399 |
-
hidden_states=latents,
|
400 |
-
timestep=timestep / 1000,
|
401 |
-
guidance=guidance,
|
402 |
-
pooled_projections=pooled_prompt_embeds,
|
403 |
-
encoder_hidden_states=prompt_embeds,
|
404 |
-
txt_ids=text_ids,
|
405 |
-
img_ids=latent_image_ids,
|
406 |
-
joint_attention_kwargs=self.joint_attention_kwargs,
|
407 |
-
return_dict=False,
|
408 |
-
)[0]
|
409 |
-
|
410 |
-
noise_pred_uncond = self.transformer(
|
411 |
-
hidden_states=latents,
|
412 |
-
timestep=timestep / 1000,
|
413 |
-
guidance=guidance,
|
414 |
-
pooled_projections=negative_pooled_prompt_embeds,
|
415 |
-
encoder_hidden_states=negative_prompt_embeds,
|
416 |
-
txt_ids=negative_text_ids,
|
417 |
-
img_ids=latent_image_ids,
|
418 |
-
joint_attention_kwargs=self.joint_attention_kwargs,
|
419 |
-
return_dict=False,
|
420 |
-
)[0]
|
421 |
-
|
422 |
-
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
|
423 |
-
|
424 |
-
latents_dtype = latents.dtype
|
425 |
-
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
426 |
-
# Yield intermediate result
|
427 |
-
torch.cuda.empty_cache()
|
428 |
-
|
429 |
-
# Final image
|
430 |
-
return self._decode_latents_to_image(latents, height, width, output_type)
|
431 |
-
self.maybe_free_model_hooks()
|
432 |
-
torch.cuda.empty_cache()
|
433 |
-
|
434 |
def _decode_latents_to_image(self, latents, height, width, output_type, vae=None):
|
435 |
"""Decodes the given latents into an image."""
|
436 |
vae = vae or self.vae
|
|
|
66 |
return timesteps, num_inference_steps
|
67 |
|
68 |
# FLUX pipeline function
|
69 |
+
class FluxWithCFGPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin):
|
70 |
|
71 |
def __init__(
|
72 |
self,
|
|
|
244 |
self.maybe_free_model_hooks()
|
245 |
torch.cuda.empty_cache()
|
246 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
247 |
def _decode_latents_to_image(self, latents, height, width, output_type, vae=None):
|
248 |
"""Decodes the given latents into an image."""
|
249 |
vae = vae or self.vae
|