Create pyramid_dit_for_video_gen_pipeline.py
Browse files
pyramid_dit_for_video_gen_pipeline.py
ADDED
@@ -0,0 +1,518 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import os
|
3 |
+
import sys
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.nn.functional as F
|
6 |
+
|
7 |
+
from collections import OrderedDict
|
8 |
+
from einops import rearrange
|
9 |
+
from diffusers.utils.torch_utils import randn_tensor
|
10 |
+
import numpy as np
|
11 |
+
import math
|
12 |
+
import random
|
13 |
+
import PIL
|
14 |
+
from PIL import Image
|
15 |
+
from tqdm import tqdm
|
16 |
+
from torchvision import transforms
|
17 |
+
from copy import deepcopy
|
18 |
+
from typing import Any, Callable, Dict, List, Optional, Union
|
19 |
+
from accelerate import Accelerator
|
20 |
+
from diffusion_schedulers import PyramidFlowMatchEulerDiscreteScheduler
|
21 |
+
from video_vae.modeling_causal_vae import CausalVideoVAE
|
22 |
+
|
23 |
+
from trainer_misc import (
|
24 |
+
all_to_all,
|
25 |
+
is_sequence_parallel_initialized,
|
26 |
+
get_sequence_parallel_group,
|
27 |
+
get_sequence_parallel_group_rank,
|
28 |
+
get_sequence_parallel_rank,
|
29 |
+
get_sequence_parallel_world_size,
|
30 |
+
get_rank,
|
31 |
+
)
|
32 |
+
|
33 |
+
from .modeling_pyramid_mmdit import PyramidDiffusionMMDiT
|
34 |
+
from .modeling_text_encoder import SD3TextEncoderWithMask
|
35 |
+
|
36 |
+
def compute_density_for_timestep_sampling(
|
37 |
+
weighting_scheme: str, batch_size: int, logit_mean: float = None, logit_std: float = None, mode_scale: float = None
|
38 |
+
):
|
39 |
+
if weighting_scheme == "logit_normal":
|
40 |
+
u = torch.normal(mean=logit_mean, std=logit_std, size=(batch_size,), device="cpu")
|
41 |
+
u = torch.nn.functional.sigmoid(u)
|
42 |
+
elif weighting_scheme == "mode":
|
43 |
+
u = torch.rand(size=(batch_size,), device="cpu")
|
44 |
+
u = 1 - u - mode_scale * (torch.cos(math.pi * u / 2) ** 2 - 1 + u)
|
45 |
+
else:
|
46 |
+
u = torch.rand(size=(batch_size,), device="cpu")
|
47 |
+
return u
|
48 |
+
|
49 |
+
class PyramidDiTForVideoGeneration:
|
50 |
+
def __init__(self, model_path, model_dtype='bf16', use_gradient_checkpointing=False, return_log=True,
|
51 |
+
model_variant="diffusion_transformer_768p", timestep_shift=1.0, stage_range=[0, 1/3, 2/3, 1],
|
52 |
+
sample_ratios=[1, 1, 1], scheduler_gamma=1/3, use_mixed_training=False, use_flash_attn=False,
|
53 |
+
load_text_encoder=True, load_vae=True, max_temporal_length=31, frame_per_unit=1, use_temporal_causal=True,
|
54 |
+
corrupt_ratio=1/3, interp_condition_pos=True, stages=[1, 2, 4], **kwargs,
|
55 |
+
):
|
56 |
+
super().__init__()
|
57 |
+
|
58 |
+
if model_dtype == 'bf16':
|
59 |
+
torch_dtype = torch.bfloat16
|
60 |
+
elif model_dtype == 'fp16':
|
61 |
+
torch_dtype = torch.float16
|
62 |
+
else:
|
63 |
+
torch_dtype = torch.float32
|
64 |
+
|
65 |
+
self.stages = stages
|
66 |
+
self.sample_ratios = sample_ratios
|
67 |
+
self.corrupt_ratio = corrupt_ratio
|
68 |
+
|
69 |
+
dit_path = os.path.join(model_path, model_variant)
|
70 |
+
|
71 |
+
# The dit
|
72 |
+
if use_mixed_training:
|
73 |
+
print("using mixed precision training, do not explicitly casting models")
|
74 |
+
self.dit = PyramidDiffusionMMDiT.from_pretrained(
|
75 |
+
dit_path, use_gradient_checkpointing=use_gradient_checkpointing,
|
76 |
+
use_flash_attn=use_flash_attn, use_t5_mask=True,
|
77 |
+
add_temp_pos_embed=True, temp_pos_embed_type='rope',
|
78 |
+
use_temporal_causal=use_temporal_causal, interp_condition_pos=interp_condition_pos,
|
79 |
+
)
|
80 |
+
else:
|
81 |
+
print("using half precision")
|
82 |
+
self.dit = PyramidDiffusionMMDiT.from_pretrained(
|
83 |
+
dit_path, torch_dtype=torch_dtype,
|
84 |
+
use_gradient_checkpointing=use_gradient_checkpointing,
|
85 |
+
use_flash_attn=use_flash_attn, use_t5_mask=True,
|
86 |
+
add_temp_pos_embed=True, temp_pos_embed_type='rope',
|
87 |
+
use_temporal_causal=use_temporal_causal, interp_condition_pos=interp_condition_pos,
|
88 |
+
)
|
89 |
+
|
90 |
+
# The text encoder
|
91 |
+
if load_text_encoder:
|
92 |
+
self.text_encoder = SD3TextEncoderWithMask(model_path, torch_dtype=torch_dtype)
|
93 |
+
else:
|
94 |
+
self.text_encoder = None
|
95 |
+
|
96 |
+
# The base video vae decoder
|
97 |
+
if load_vae:
|
98 |
+
self.vae = CausalVideoVAE.from_pretrained(os.path.join(model_path, 'causal_video_vae'), torch_dtype=torch_dtype, interpolate=False)
|
99 |
+
# Freeze vae
|
100 |
+
for parameter in self.vae.parameters():
|
101 |
+
parameter.requires_grad = False
|
102 |
+
else:
|
103 |
+
self.vae = None
|
104 |
+
|
105 |
+
# For the image latent
|
106 |
+
self.vae_shift_factor = 0.1490
|
107 |
+
self.vae_scale_factor = 1 / 1.8415
|
108 |
+
|
109 |
+
# For the video latent
|
110 |
+
self.vae_video_shift_factor = -0.2343
|
111 |
+
self.vae_video_scale_factor = 1 / 3.0986
|
112 |
+
|
113 |
+
self.downsample = 8
|
114 |
+
|
115 |
+
# Configure the video training hyper-parameters
|
116 |
+
# The video sequence: one frame + N * unit
|
117 |
+
self.frame_per_unit = frame_per_unit
|
118 |
+
self.max_temporal_length = max_temporal_length
|
119 |
+
assert (max_temporal_length - 1) % frame_per_unit == 0, "The frame number should be divided by the frame number per unit"
|
120 |
+
self.num_units_per_video = 1 + ((max_temporal_length - 1) // frame_per_unit) + int(sum(sample_ratios))
|
121 |
+
|
122 |
+
self.scheduler = PyramidFlowMatchEulerDiscreteScheduler(
|
123 |
+
shift=timestep_shift, stages=len(self.stages),
|
124 |
+
stage_range=stage_range, gamma=scheduler_gamma,
|
125 |
+
)
|
126 |
+
print(f"The start sigmas and end sigmas of each stage is Start: {self.scheduler.start_sigmas}, End: {self.scheduler.end_sigmas}, Ori_start: {self.scheduler.ori_start_sigmas}")
|
127 |
+
|
128 |
+
self.cfg_rate = 0.1
|
129 |
+
self.return_log = return_log
|
130 |
+
self.use_flash_attn = use_flash_attn
|
131 |
+
|
132 |
+
# Initialize scaler for mixed precision
|
133 |
+
self.scaler = torch.cuda.amp.GradScaler()
|
134 |
+
|
135 |
+
# ... [other methods remain the same] ...
|
136 |
+
|
137 |
+
@torch.cuda.amp.autocast()
|
138 |
+
def generate(
|
139 |
+
self,
|
140 |
+
prompt: Union[str, List[str]] = None,
|
141 |
+
height: Optional[int] = None,
|
142 |
+
width: Optional[int] = None,
|
143 |
+
temp: int = 1,
|
144 |
+
num_inference_steps: Optional[Union[int, List[int]]] = 28,
|
145 |
+
video_num_inference_steps: Optional[Union[int, List[int]]] = 28,
|
146 |
+
guidance_scale: float = 7.0,
|
147 |
+
video_guidance_scale: float = 7.0,
|
148 |
+
min_guidance_scale: float = 2.0,
|
149 |
+
use_linear_guidance: bool = False,
|
150 |
+
alpha: float = 0.5,
|
151 |
+
negative_prompt: Optional[Union[str, List[str]]]="cartoon style, worst quality, low quality, blurry, absolute black, absolute white, low res, extra limbs, extra digits, misplaced objects, mutated anatomy, monochrome, horror",
|
152 |
+
num_images_per_prompt: Optional[int] = 1,
|
153 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
154 |
+
output_type: Optional[str] = "pil",
|
155 |
+
save_memory: bool = True,
|
156 |
+
cpu_offloading: bool = False,
|
157 |
+
):
|
158 |
+
device = self.device if not cpu_offloading else "cuda"
|
159 |
+
dtype = self.dtype
|
160 |
+
if cpu_offloading:
|
161 |
+
if str(self.dit.device) != "cpu":
|
162 |
+
print("(dit) Warning: Do not preload pipeline components (i.e. to cuda) with cpu offloading enabled! Otherwise, a second transfer will occur needlessly taking up time.")
|
163 |
+
self.dit.to("cpu")
|
164 |
+
if str(self.vae.device) != "cpu":
|
165 |
+
print("(vae) Warning: Do not preload pipeline components (i.e. to cuda) with cpu offloading enabled! Otherwise, a second transfer will occur needlessly taking up time.")
|
166 |
+
self.vae.to("cpu")
|
167 |
+
|
168 |
+
assert (temp - 1) % self.frame_per_unit == 0, "The frames should be divided by frame_per unit"
|
169 |
+
|
170 |
+
if isinstance(prompt, str):
|
171 |
+
batch_size = 1
|
172 |
+
prompt = prompt + ", hyper quality, Ultra HD, 8K"
|
173 |
+
else:
|
174 |
+
assert isinstance(prompt, list)
|
175 |
+
batch_size = len(prompt)
|
176 |
+
prompt = [_ + ", hyper quality, Ultra HD, 8K" for _ in prompt]
|
177 |
+
|
178 |
+
if isinstance(num_inference_steps, int):
|
179 |
+
num_inference_steps = [num_inference_steps] * len(self.stages)
|
180 |
+
|
181 |
+
if isinstance(video_num_inference_steps, int):
|
182 |
+
video_num_inference_steps = [video_num_inference_steps] * len(self.stages)
|
183 |
+
|
184 |
+
negative_prompt = negative_prompt or ""
|
185 |
+
|
186 |
+
# Get the text embeddings
|
187 |
+
if cpu_offloading:
|
188 |
+
self.text_encoder.to("cuda")
|
189 |
+
prompt_embeds, prompt_attention_mask, pooled_prompt_embeds = self.text_encoder(prompt, device)
|
190 |
+
negative_prompt_embeds, negative_prompt_attention_mask, negative_pooled_prompt_embeds = self.text_encoder(negative_prompt, device)
|
191 |
+
if cpu_offloading:
|
192 |
+
self.text_encoder.to("cpu")
|
193 |
+
self.dit.to("cuda")
|
194 |
+
|
195 |
+
if use_linear_guidance:
|
196 |
+
max_guidance_scale = guidance_scale
|
197 |
+
guidance_scale_list = [max(max_guidance_scale - alpha * t_, min_guidance_scale) for t_ in range(temp)]
|
198 |
+
print(guidance_scale_list)
|
199 |
+
|
200 |
+
self._guidance_scale = guidance_scale
|
201 |
+
self._video_guidance_scale = video_guidance_scale
|
202 |
+
|
203 |
+
if self.do_classifier_free_guidance:
|
204 |
+
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
|
205 |
+
pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
|
206 |
+
prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
|
207 |
+
|
208 |
+
# Create the initial random noise
|
209 |
+
num_channels_latents = self.dit.config.in_channels
|
210 |
+
latents = self.prepare_latents(
|
211 |
+
batch_size * num_images_per_prompt,
|
212 |
+
num_channels_latents,
|
213 |
+
temp,
|
214 |
+
height,
|
215 |
+
width,
|
216 |
+
prompt_embeds.dtype,
|
217 |
+
device,
|
218 |
+
generator,
|
219 |
+
)
|
220 |
+
|
221 |
+
temp, height, width = latents.shape[-3], latents.shape[-2], latents.shape[-1]
|
222 |
+
|
223 |
+
latents = rearrange(latents, 'b c t h w -> (b t) c h w')
|
224 |
+
for _ in range(len(self.stages)-1):
|
225 |
+
height //= 2;width //= 2
|
226 |
+
latents = F.interpolate(latents, size=(height, width), mode='bilinear') * 2
|
227 |
+
|
228 |
+
latents = rearrange(latents, '(b t) c h w -> b c t h w', t=temp)
|
229 |
+
|
230 |
+
num_units = 1 + (temp - 1) // self.frame_per_unit
|
231 |
+
stages = self.stages
|
232 |
+
|
233 |
+
generated_latents_list = []
|
234 |
+
last_generated_latents = None
|
235 |
+
|
236 |
+
for unit_index in tqdm(range(num_units)):
|
237 |
+
if use_linear_guidance:
|
238 |
+
self._guidance_scale = guidance_scale_list[unit_index]
|
239 |
+
self._video_guidance_scale = guidance_scale_list[unit_index]
|
240 |
+
|
241 |
+
if unit_index == 0:
|
242 |
+
past_condition_latents = [[] for _ in range(len(stages))]
|
243 |
+
with torch.no_grad():
|
244 |
+
intermed_latents = self.generate_one_unit(
|
245 |
+
latents[:,:,:1],
|
246 |
+
past_condition_latents,
|
247 |
+
prompt_embeds,
|
248 |
+
prompt_attention_mask,
|
249 |
+
pooled_prompt_embeds,
|
250 |
+
num_inference_steps,
|
251 |
+
height,
|
252 |
+
width,
|
253 |
+
1,
|
254 |
+
device,
|
255 |
+
dtype,
|
256 |
+
generator,
|
257 |
+
is_first_frame=True,
|
258 |
+
)
|
259 |
+
else:
|
260 |
+
past_condition_latents = []
|
261 |
+
clean_latents_list = self.get_pyramid_latent(torch.cat(generated_latents_list, dim=2), len(stages) - 1)
|
262 |
+
|
263 |
+
for i_s in range(len(stages)):
|
264 |
+
last_cond_latent = clean_latents_list[i_s][:,:,-(self.frame_per_unit):]
|
265 |
+
stage_input = [torch.cat([last_cond_latent] * 2) if self.do_classifier_free_guidance else last_cond_latent]
|
266 |
+
|
267 |
+
cur_unit_num = unit_index
|
268 |
+
cur_stage = i_s
|
269 |
+
cur_unit_ptx = 1
|
270 |
+
|
271 |
+
while cur_unit_ptx < cur_unit_num:
|
272 |
+
cur_stage = max(cur_stage - 1, 0)
|
273 |
+
if cur_stage == 0:
|
274 |
+
break
|
275 |
+
cur_unit_ptx += 1
|
276 |
+
cond_latents = clean_latents_list[cur_stage][:, :, -(cur_unit_ptx * self.frame_per_unit) : -((cur_unit_ptx - 1) * self.frame_per_unit)]
|
277 |
+
stage_input.append(torch.cat([cond_latents] * 2) if self.do_classifier_free_guidance else cond_latents)
|
278 |
+
|
279 |
+
if cur_stage == 0 and cur_unit_ptx < cur_unit_num:
|
280 |
+
cond_latents = clean_latents_list[0][:, :, :-(cur_unit_ptx * self.frame_per_unit)]
|
281 |
+
stage_input.append(torch.cat([cond_latents] * 2) if self.do_classifier_free_guidance else cond_latents)
|
282 |
+
|
283 |
+
stage_input = list(reversed(stage_input))
|
284 |
+
past_condition_latents.append(stage_input)
|
285 |
+
|
286 |
+
with torch.no_grad():
|
287 |
+
intermed_latents = self.generate_one_unit(
|
288 |
+
latents[:,:, 1 + (unit_index - 1) * self.frame_per_unit:1 + unit_index * self.frame_per_unit],
|
289 |
+
past_condition_latents,
|
290 |
+
prompt_embeds,
|
291 |
+
prompt_attention_mask,
|
292 |
+
pooled_prompt_embeds,
|
293 |
+
video_num_inference_steps,
|
294 |
+
height,
|
295 |
+
width,
|
296 |
+
self.frame_per_unit,
|
297 |
+
device,
|
298 |
+
dtype,
|
299 |
+
generator,
|
300 |
+
is_first_frame=False,
|
301 |
+
)
|
302 |
+
|
303 |
+
generated_latents_list.append(intermed_latents[-1])
|
304 |
+
last_generated_latents = intermed_latents
|
305 |
+
|
306 |
+
torch.cuda.empty_cache()
|
307 |
+
|
308 |
+
generated_latents = torch.cat(generated_latents_list, dim=2)
|
309 |
+
|
310 |
+
if output_type == "latent":
|
311 |
+
image = generated_latents
|
312 |
+
else:
|
313 |
+
if cpu_offloading:
|
314 |
+
self.dit.to("cpu")
|
315 |
+
self.vae.to("cuda")
|
316 |
+
image = self.decode_latent(generated_latents, save_memory=save_memory)
|
317 |
+
if cpu_offloading:
|
318 |
+
self.vae.to("cpu")
|
319 |
+
|
320 |
+
return image
|
321 |
+
|
322 |
+
def decode_latent(self, latents, save_memory=True):
|
323 |
+
if latents.shape[2] == 1:
|
324 |
+
latents = (latents / self.vae_scale_factor) + self.vae_shift_factor
|
325 |
+
else:
|
326 |
+
latents[:, :, :1] = (latents[:, :, :1] / self.vae_scale_factor) + self.vae_shift_factor
|
327 |
+
latents[:, :, 1:] = (latents[:, :, 1:] / self.vae_video_scale_factor) + self.vae_video_shift_factor
|
328 |
+
|
329 |
+
with torch.no_grad(), torch.cuda.amp.autocast():
|
330 |
+
if save_memory:
|
331 |
+
image = self.vae.decode(latents, temporal_chunk=True, window_size=1, tile_sample_min_size=128).sample
|
332 |
+
else:
|
333 |
+
image = self.vae.decode(latents, temporal_chunk=True, window_size=2, tile_sample_min_size=256).sample
|
334 |
+
|
335 |
+
image = image.float()
|
336 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
337 |
+
image = rearrange(image, "B C T H W -> (B T) C H W")
|
338 |
+
image = image.cpu().permute(0, 2, 3, 1).numpy()
|
339 |
+
image = self.numpy_to_pil(image)
|
340 |
+
return image
|
341 |
+
|
342 |
+
@staticmethod
|
343 |
+
def numpy_to_pil(images):
|
344 |
+
if images.ndim == 3:
|
345 |
+
images = images[None, ...]
|
346 |
+
images = (images * 255).round().astype("uint8")
|
347 |
+
if images.shape[-1] == 1:
|
348 |
+
pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
|
349 |
+
else:
|
350 |
+
pil_images = [Image.fromarray(image) for image in images]
|
351 |
+
return pil_images
|
352 |
+
|
353 |
+
@property
|
354 |
+
def device(self):
|
355 |
+
return next(self.dit.parameters()).device
|
356 |
+
|
357 |
+
@property
|
358 |
+
def dtype(self):
|
359 |
+
return next(self.dit.parameters()).dtype
|
360 |
+
|
361 |
+
@property
|
362 |
+
def guidance_scale(self):
|
363 |
+
return self._guidance_scale
|
364 |
+
|
365 |
+
@property
|
366 |
+
def video_guidance_scale(self):
|
367 |
+
return self._video_guidance_scale
|
368 |
+
|
369 |
+
@property
|
370 |
+
def do_classifier_free_guidance(self):
|
371 |
+
return self._guidance_scale > 0
|
372 |
+
|
373 |
+
def prepare_latents(
|
374 |
+
self,
|
375 |
+
batch_size,
|
376 |
+
num_channels_latents,
|
377 |
+
temp,
|
378 |
+
height,
|
379 |
+
width,
|
380 |
+
dtype,
|
381 |
+
device,
|
382 |
+
generator,
|
383 |
+
):
|
384 |
+
shape = (
|
385 |
+
batch_size,
|
386 |
+
num_channels_latents,
|
387 |
+
int(temp),
|
388 |
+
int(height) // self.downsample,
|
389 |
+
int(width) // self.downsample,
|
390 |
+
)
|
391 |
+
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
392 |
+
return latents
|
393 |
+
|
394 |
+
def sample_block_noise(self, bs, ch, temp, height, width):
|
395 |
+
gamma = self.scheduler.config.gamma
|
396 |
+
dist = torch.distributions.multivariate_normal.MultivariateNormal(torch.zeros(4), torch.eye(4) * (1 + gamma) - torch.ones(4, 4) * gamma)
|
397 |
+
block_number = bs * ch * temp * (height // 2) * (width // 2)
|
398 |
+
noise = torch.stack([dist.sample() for _ in range(block_number)])
|
399 |
+
noise = rearrange(noise, '(b c t h w) (p q) -> b c t (h p) (w q)',b=bs,c=ch,t=temp,h=height//2,w=width//2,p=2,q=2)
|
400 |
+
return noise
|
401 |
+
|
402 |
+
@torch.no_grad()
|
403 |
+
def generate_one_unit(
|
404 |
+
self,
|
405 |
+
latents,
|
406 |
+
past_conditions,
|
407 |
+
prompt_embeds,
|
408 |
+
prompt_attention_mask,
|
409 |
+
pooled_prompt_embeds,
|
410 |
+
num_inference_steps,
|
411 |
+
height,
|
412 |
+
width,
|
413 |
+
temp,
|
414 |
+
device,
|
415 |
+
dtype,
|
416 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
417 |
+
is_first_frame: bool = False,
|
418 |
+
):
|
419 |
+
stages = self.stages
|
420 |
+
intermed_latents = []
|
421 |
+
|
422 |
+
for i_s in range(len(stages)):
|
423 |
+
self.scheduler.set_timesteps(num_inference_steps[i_s], i_s, device=device)
|
424 |
+
timesteps = self.scheduler.timesteps
|
425 |
+
|
426 |
+
if i_s > 0:
|
427 |
+
height *= 2; width *= 2
|
428 |
+
latents = rearrange(latents, 'b c t h w -> (b t) c h w')
|
429 |
+
latents = F.interpolate(latents, size=(height, width), mode='nearest')
|
430 |
+
latents = rearrange(latents, '(b t) c h w -> b c t h w', t=temp)
|
431 |
+
ori_sigma = 1 - self.scheduler.ori_start_sigmas[i_s]
|
432 |
+
gamma = self.scheduler.config.gamma
|
433 |
+
alpha = 1 / (math.sqrt(1 + (1 / gamma)) * (1 - ori_sigma) + ori_sigma)
|
434 |
+
beta = alpha * (1 - ori_sigma) / math.sqrt(gamma)
|
435 |
+
|
436 |
+
bs, ch, temp, height, width = latents.shape
|
437 |
+
noise = self.sample_block_noise(bs, ch, temp, height, width)
|
438 |
+
noise = noise.to(device=device, dtype=dtype)
|
439 |
+
latents = alpha * latents + beta * noise
|
440 |
+
|
441 |
+
for idx, t in enumerate(timesteps):
|
442 |
+
latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
|
443 |
+
timestep = t.expand(latent_model_input.shape[0]).to(latent_model_input.dtype)
|
444 |
+
latent_model_input = past_conditions[i_s] + [latent_model_input]
|
445 |
+
|
446 |
+
noise_pred = self.dit(
|
447 |
+
sample=[latent_model_input],
|
448 |
+
timestep_ratio=timestep,
|
449 |
+
encoder_hidden_states=prompt_embeds,
|
450 |
+
encoder_attention_mask=prompt_attention_mask,
|
451 |
+
pooled_projections=pooled_prompt_embeds,
|
452 |
+
)
|
453 |
+
|
454 |
+
noise_pred = noise_pred[0]
|
455 |
+
|
456 |
+
if self.do_classifier_free_guidance:
|
457 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
458 |
+
if is_first_frame:
|
459 |
+
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
|
460 |
+
else:
|
461 |
+
noise_pred = noise_pred_uncond + self.video_guidance_scale * (noise_pred_text - noise_pred_uncond)
|
462 |
+
|
463 |
+
latents = self.scheduler.step(
|
464 |
+
model_output=noise_pred,
|
465 |
+
timestep=timestep,
|
466 |
+
sample=latents,
|
467 |
+
generator=generator,
|
468 |
+
).prev_sample
|
469 |
+
|
470 |
+
intermed_latents.append(latents)
|
471 |
+
|
472 |
+
return intermed_latents
|
473 |
+
|
474 |
+
def get_pyramid_latent(self, x, stage_num):
|
475 |
+
vae_latent_list = []
|
476 |
+
vae_latent_list.append(x)
|
477 |
+
|
478 |
+
temp, height, width = x.shape[-3], x.shape[-2], x.shape[-1]
|
479 |
+
for _ in range(stage_num):
|
480 |
+
height //= 2
|
481 |
+
width //= 2
|
482 |
+
x = rearrange(x, 'b c t h w -> (b t) c h w')
|
483 |
+
x = torch.nn.functional.interpolate(x, size=(height, width), mode='bilinear')
|
484 |
+
x = rearrange(x, '(b t) c h w -> b c t h w', t=temp)
|
485 |
+
vae_latent_list.append(x)
|
486 |
+
|
487 |
+
vae_latent_list = list(reversed(vae_latent_list))
|
488 |
+
return vae_latent_list
|
489 |
+
|
490 |
+
def load_checkpoint(self, checkpoint_path, model_key='model', **kwargs):
|
491 |
+
checkpoint = torch.load(checkpoint_path, map_location='cpu')
|
492 |
+
dit_checkpoint = OrderedDict()
|
493 |
+
for key in checkpoint:
|
494 |
+
if key.startswith('vae') or key.startswith('text_encoder'):
|
495 |
+
continue
|
496 |
+
if key.startswith('dit'):
|
497 |
+
new_key = key.split('.')
|
498 |
+
new_key = '.'.join(new_key[1:])
|
499 |
+
dit_checkpoint[new_key] = checkpoint[key]
|
500 |
+
else:
|
501 |
+
dit_checkpoint[key] = checkpoint[key]
|
502 |
+
|
503 |
+
load_result = self.dit.load_state_dict(dit_checkpoint, strict=True)
|
504 |
+
print(f"Load checkpoint from {checkpoint_path}, load result: {load_result}")
|
505 |
+
|
506 |
+
def load_vae_checkpoint(self, vae_checkpoint_path, model_key='model'):
|
507 |
+
checkpoint = torch.load(vae_checkpoint_path, map_location='cpu')
|
508 |
+
checkpoint = checkpoint[model_key]
|
509 |
+
loaded_checkpoint = OrderedDict()
|
510 |
+
|
511 |
+
for key in checkpoint.keys():
|
512 |
+
if key.startswith('vae.'):
|
513 |
+
new_key = key.split('.')
|
514 |
+
new_key = '.'.join(new_key[1:])
|
515 |
+
loaded_checkpoint[new_key] = checkpoint[key]
|
516 |
+
|
517 |
+
load_result = self.vae.load_state_dict(loaded_checkpoint)
|
518 |
+
print(f"Load the VAE from {vae_checkpoint_path}, load result: {load_result}")
|