Spaces:
Running
on
Zero
Running
on
Zero
Update custom_pipeline.py
Browse files- custom_pipeline.py +53 -147
custom_pipeline.py
CHANGED
@@ -1,16 +1,8 @@
|
|
1 |
-
import numpy as np
|
2 |
import torch
|
3 |
-
|
4 |
-
from
|
5 |
-
from
|
6 |
-
from
|
7 |
-
|
8 |
-
if is_torch_xla_available():
|
9 |
-
import torch_xla.core.xla_model as xm
|
10 |
-
|
11 |
-
XLA_AVAILABLE = True
|
12 |
-
else:
|
13 |
-
XLA_AVAILABLE = False
|
14 |
|
15 |
# Constants for shift calculation
|
16 |
BASE_SEQ_LEN = 256
|
@@ -27,7 +19,7 @@ def calculate_timestep_shift(image_seq_len: int) -> float:
|
|
27 |
return mu
|
28 |
|
29 |
def prepare_timesteps(
|
30 |
-
scheduler,
|
31 |
num_inference_steps: Optional[int] = None,
|
32 |
device: Optional[Union[str, torch.device]] = None,
|
33 |
timesteps: Optional[List[int]] = None,
|
@@ -49,23 +41,24 @@ def prepare_timesteps(
|
|
49 |
num_inference_steps = len(timesteps)
|
50 |
return timesteps, num_inference_steps
|
51 |
|
52 |
-
# FLUX pipeline
|
53 |
class FluxWithCFGPipeline(FluxPipeline):
|
54 |
"""
|
55 |
-
|
56 |
-
|
57 |
-
increasing resolution for faster generation.
|
58 |
"""
|
|
|
|
|
|
|
|
|
59 |
@torch.inference_mode()
|
60 |
-
def
|
61 |
self,
|
62 |
prompt: Union[str, List[str]] = None,
|
63 |
prompt_2: Optional[Union[str, List[str]]] = None,
|
64 |
-
negative_prompt: Optional[Union[str, List[str]]] = None,
|
65 |
-
negative_prompt_2: Optional[Union[str, List[str]]] = None,
|
66 |
height: Optional[int] = None,
|
67 |
width: Optional[int] = None,
|
68 |
-
num_inference_steps: int =
|
69 |
timesteps: List[int] = None,
|
70 |
guidance_scale: float = 3.5,
|
71 |
num_images_per_prompt: Optional[int] = 1,
|
@@ -73,21 +66,16 @@ class FluxWithCFGPipeline(FluxPipeline):
|
|
73 |
latents: Optional[torch.FloatTensor] = None,
|
74 |
prompt_embeds: Optional[torch.FloatTensor] = None,
|
75 |
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
76 |
-
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
77 |
-
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
78 |
output_type: Optional[str] = "pil",
|
79 |
return_dict: bool = True,
|
80 |
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
|
81 |
-
|
82 |
-
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
83 |
-
max_sequence_length: int = 512,
|
84 |
-
yield_intermediates: bool = False, # New parameter for yielding intermediates
|
85 |
):
|
86 |
-
|
87 |
height = height or self.default_sample_size * self.vae_scale_factor
|
88 |
width = width or self.default_sample_size * self.vae_scale_factor
|
89 |
|
90 |
-
# 1. Check inputs
|
91 |
self.check_inputs(
|
92 |
prompt,
|
93 |
prompt_2,
|
@@ -95,7 +83,6 @@ class FluxWithCFGPipeline(FluxPipeline):
|
|
95 |
width,
|
96 |
prompt_embeds=prompt_embeds,
|
97 |
pooled_prompt_embeds=pooled_prompt_embeds,
|
98 |
-
callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
|
99 |
max_sequence_length=max_sequence_length,
|
100 |
)
|
101 |
|
@@ -104,23 +91,12 @@ class FluxWithCFGPipeline(FluxPipeline):
|
|
104 |
self._interrupt = False
|
105 |
|
106 |
# 2. Define call parameters
|
107 |
-
|
108 |
-
batch_size = 1
|
109 |
-
elif prompt is not None and isinstance(prompt, list):
|
110 |
-
batch_size = len(prompt)
|
111 |
-
else:
|
112 |
-
batch_size = prompt_embeds.shape[0]
|
113 |
-
|
114 |
device = self._execution_device
|
115 |
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
(
|
120 |
-
prompt_embeds,
|
121 |
-
pooled_prompt_embeds,
|
122 |
-
text_ids,
|
123 |
-
) = self.encode_prompt(
|
124 |
prompt=prompt,
|
125 |
prompt_2=prompt_2,
|
126 |
prompt_embeds=prompt_embeds,
|
@@ -130,20 +106,6 @@ class FluxWithCFGPipeline(FluxPipeline):
|
|
130 |
max_sequence_length=max_sequence_length,
|
131 |
lora_scale=lora_scale,
|
132 |
)
|
133 |
-
(
|
134 |
-
negative_prompt_embeds,
|
135 |
-
negative_pooled_prompt_embeds,
|
136 |
-
negative_text_ids,
|
137 |
-
) = self.encode_prompt(
|
138 |
-
prompt=negative_prompt,
|
139 |
-
prompt_2=negative_prompt_2,
|
140 |
-
prompt_embeds=negative_prompt_embeds,
|
141 |
-
pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
142 |
-
device=device,
|
143 |
-
num_images_per_prompt=num_images_per_prompt,
|
144 |
-
max_sequence_length=max_sequence_length,
|
145 |
-
lora_scale=lora_scale,
|
146 |
-
)
|
147 |
|
148 |
# 4. Prepare latent variables
|
149 |
num_channels_latents = self.transformer.config.in_channels // 4
|
@@ -170,97 +132,41 @@ class FluxWithCFGPipeline(FluxPipeline):
|
|
170 |
sigmas,
|
171 |
mu=mu,
|
172 |
)
|
173 |
-
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
174 |
self._num_timesteps = len(timesteps)
|
175 |
|
176 |
-
#
|
177 |
-
|
178 |
-
for i, t in enumerate(timesteps):
|
179 |
-
if self.interrupt:
|
180 |
-
continue
|
181 |
-
|
182 |
-
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
183 |
-
timestep = t.expand(latents.shape[0]).to(latents.dtype)
|
184 |
-
|
185 |
-
# handle guidance
|
186 |
-
if self.transformer.config.guidance_embeds:
|
187 |
-
guidance = torch.tensor([guidance_scale], device=device)
|
188 |
-
guidance = guidance.expand(latents.shape[0])
|
189 |
-
else:
|
190 |
-
guidance = None
|
191 |
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
latents_dtype = latents.dtype
|
220 |
-
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
221 |
-
|
222 |
-
if latents.dtype != latents_dtype:
|
223 |
-
if torch.backends.mps.is_available():
|
224 |
-
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
|
225 |
-
latents = latents.to(latents_dtype)
|
226 |
-
|
227 |
-
if callback_on_step_end is not None:
|
228 |
-
callback_kwargs = {}
|
229 |
-
for k in callback_on_step_end_tensor_inputs:
|
230 |
-
callback_kwargs[k] = locals()[k]
|
231 |
-
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
|
232 |
-
|
233 |
-
latents = callback_outputs.pop("latents", latents)
|
234 |
-
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
|
235 |
-
|
236 |
-
# call the callback, if provided
|
237 |
-
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
238 |
-
progress_bar.update()
|
239 |
-
|
240 |
-
# Yield intermediate images if requested
|
241 |
-
if yield_intermediates:
|
242 |
-
yield self._decode_latents_to_image(latents, height, width, output_type)
|
243 |
-
|
244 |
-
if XLA_AVAILABLE:
|
245 |
-
xm.mark_step()
|
246 |
-
|
247 |
-
# Final image decoding
|
248 |
-
if output_type == "latent":
|
249 |
-
image = latents
|
250 |
-
else:
|
251 |
-
image = self._decode_latents_to_image(latents, height, width, output_type)
|
252 |
-
|
253 |
-
# Offload all models
|
254 |
-
self.maybe_free_model_hooks()
|
255 |
-
|
256 |
-
if not return_dict:
|
257 |
-
return (image,)
|
258 |
-
|
259 |
-
return FluxPipelineOutput(images=image)
|
260 |
-
|
261 |
-
def _decode_latents_to_image(self, latents, height, width, output_type):
|
262 |
"""Decodes the given latents into an image."""
|
|
|
263 |
latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
|
264 |
-
latents = (latents /
|
265 |
-
image =
|
266 |
return self.image_processor.postprocess(image, output_type=output_type)[0]
|
|
|
|
|
1 |
import torch
|
2 |
+
import numpy as np
|
3 |
+
from diffusers import FluxPipeline, FlowMatchEulerDiscreteScheduler
|
4 |
+
from typing import Any, Dict, List, Optional, Union
|
5 |
+
from PIL import Image
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
# Constants for shift calculation
|
8 |
BASE_SEQ_LEN = 256
|
|
|
19 |
return mu
|
20 |
|
21 |
def prepare_timesteps(
|
22 |
+
scheduler: FlowMatchEulerDiscreteScheduler,
|
23 |
num_inference_steps: Optional[int] = None,
|
24 |
device: Optional[Union[str, torch.device]] = None,
|
25 |
timesteps: Optional[List[int]] = None,
|
|
|
41 |
num_inference_steps = len(timesteps)
|
42 |
return timesteps, num_inference_steps
|
43 |
|
44 |
+
# FLUX pipeline function
|
45 |
class FluxWithCFGPipeline(FluxPipeline):
|
46 |
"""
|
47 |
+
Extends the FluxPipeline to yield intermediate images during the denoising process
|
48 |
+
with progressively increasing resolution for faster generation.
|
|
|
49 |
"""
|
50 |
+
def __init__(self, *args, **kwargs):
|
51 |
+
super().__init__(*args, **kwargs)
|
52 |
+
self.default_sample_size = 512 # Default sample size from the first pipeline
|
53 |
+
|
54 |
@torch.inference_mode()
|
55 |
+
def generate_images(
|
56 |
self,
|
57 |
prompt: Union[str, List[str]] = None,
|
58 |
prompt_2: Optional[Union[str, List[str]]] = None,
|
|
|
|
|
59 |
height: Optional[int] = None,
|
60 |
width: Optional[int] = None,
|
61 |
+
num_inference_steps: int = 4,
|
62 |
timesteps: List[int] = None,
|
63 |
guidance_scale: float = 3.5,
|
64 |
num_images_per_prompt: Optional[int] = 1,
|
|
|
66 |
latents: Optional[torch.FloatTensor] = None,
|
67 |
prompt_embeds: Optional[torch.FloatTensor] = None,
|
68 |
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
|
|
|
|
69 |
output_type: Optional[str] = "pil",
|
70 |
return_dict: bool = True,
|
71 |
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
|
72 |
+
max_sequence_length: int = 300,
|
|
|
|
|
|
|
73 |
):
|
74 |
+
"""Generates images and yields intermediate results during the denoising process."""
|
75 |
height = height or self.default_sample_size * self.vae_scale_factor
|
76 |
width = width or self.default_sample_size * self.vae_scale_factor
|
77 |
|
78 |
+
# 1. Check inputs
|
79 |
self.check_inputs(
|
80 |
prompt,
|
81 |
prompt_2,
|
|
|
83 |
width,
|
84 |
prompt_embeds=prompt_embeds,
|
85 |
pooled_prompt_embeds=pooled_prompt_embeds,
|
|
|
86 |
max_sequence_length=max_sequence_length,
|
87 |
)
|
88 |
|
|
|
91 |
self._interrupt = False
|
92 |
|
93 |
# 2. Define call parameters
|
94 |
+
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
device = self._execution_device
|
96 |
|
97 |
+
# 3. Encode prompt
|
98 |
+
lora_scale = joint_attention_kwargs.get("scale", None) if joint_attention_kwargs is not None else None
|
99 |
+
prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt(
|
|
|
|
|
|
|
|
|
|
|
100 |
prompt=prompt,
|
101 |
prompt_2=prompt_2,
|
102 |
prompt_embeds=prompt_embeds,
|
|
|
106 |
max_sequence_length=max_sequence_length,
|
107 |
lora_scale=lora_scale,
|
108 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
109 |
|
110 |
# 4. Prepare latent variables
|
111 |
num_channels_latents = self.transformer.config.in_channels // 4
|
|
|
132 |
sigmas,
|
133 |
mu=mu,
|
134 |
)
|
|
|
135 |
self._num_timesteps = len(timesteps)
|
136 |
|
137 |
+
# Handle guidance
|
138 |
+
guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float16).expand(latents.shape[0]) if self.transformer.config.guidance_embeds else None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
139 |
|
140 |
+
# 6. Denoising loop
|
141 |
+
for i, t in enumerate(timesteps):
|
142 |
+
if self.interrupt:
|
143 |
+
continue
|
144 |
+
|
145 |
+
timestep = t.expand(latents.shape[0]).to(latents.dtype)
|
146 |
+
|
147 |
+
noise_pred = self.transformer(
|
148 |
+
hidden_states=latents,
|
149 |
+
timestep=timestep / 1000,
|
150 |
+
guidance=guidance,
|
151 |
+
pooled_projections=pooled_prompt_embeds,
|
152 |
+
encoder_hidden_states=prompt_embeds,
|
153 |
+
txt_ids=text_ids,
|
154 |
+
img_ids=latent_image_ids,
|
155 |
+
joint_attention_kwargs=self.joint_attention_kwargs,
|
156 |
+
return_dict=False,
|
157 |
+
)[0]
|
158 |
+
|
159 |
+
# Yield intermediate result
|
160 |
+
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
161 |
+
torch.cuda.empty_cache()
|
162 |
+
|
163 |
+
# Final image
|
164 |
+
return self._decode_latents_to_image(latents, height, width, output_type)
|
165 |
+
|
166 |
+
def _decode_latents_to_image(self, latents, height, width, output_type, vae=None):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
167 |
"""Decodes the given latents into an image."""
|
168 |
+
vae = vae or self.vae
|
169 |
latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
|
170 |
+
latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor
|
171 |
+
image = vae.decode(latents, return_dict=False)[0]
|
172 |
return self.image_processor.postprocess(image, output_type=output_type)[0]
|