forked from newgenai79/LTXVideo
-
Notifications
You must be signed in to change notification settings - Fork 0
/
inference.py
369 lines (311 loc) · 12.2 KB
/
inference.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
from omegaconf import OmegaConf
import json
import os
import random
from datetime import datetime
from pathlib import Path
from diffusers.utils import logging
import imageio
import numpy as np
import safetensors.torch
import torch
import torch.nn.functional as F
from PIL import Image
from transformers import T5EncoderModel, T5Tokenizer
from torchao.quantization import quantize_, int8_weight_only
from ltx_video.models.autoencoders.causal_video_autoencoder import (
CausalVideoAutoencoder,
)
from ltx_video.models.transformers.symmetric_patchifier import SymmetricPatchifier
from ltx_video.models.transformers.transformer3d import Transformer3DModel
from ltx_video.pipelines.pipeline_ltx_video import LTXVideoPipeline
from ltx_video.schedulers.rf import RectifiedFlowScheduler
from ltx_video.utils.conditioning_method import ConditioningMethod
MAX_HEIGHT = 720
MAX_WIDTH = 1280
MAX_NUM_FRAMES = 257
def load_vae(vae_dir):
vae_ckpt_path = vae_dir / "vae_diffusion_pytorch_model.safetensors"
vae_config_path = vae_dir / "config.json"
with open(vae_config_path, "r") as f:
vae_config = json.load(f)
vae = CausalVideoAutoencoder.from_config(vae_config)
vae_state_dict = safetensors.torch.load_file(vae_ckpt_path)
vae.load_state_dict(vae_state_dict)
if torch.cuda.is_available():
vae = vae.cuda()
if config.int8:
quantize_(vae, int8_weight_only())
torch.cuda.empty_cache()
return vae.to(torch.bfloat16)
def load_unet(unet_dir):
unet_ckpt_path = unet_dir / "unet_diffusion_pytorch_model.safetensors"
unet_config_path = unet_dir / "config.json"
transformer_config = Transformer3DModel.load_config(unet_config_path)
transformer = Transformer3DModel.from_config(transformer_config)
unet_state_dict = safetensors.torch.load_file(unet_ckpt_path)
transformer.load_state_dict(unet_state_dict, strict=True)
if torch.cuda.is_available():
transformer = transformer.cuda()
if config.int8:
quantize_(transformer, int8_weight_only())
torch.cuda.empty_cache()
return transformer
def load_scheduler(scheduler_dir):
scheduler_config_path = scheduler_dir / "scheduler_config.json"
scheduler_config = RectifiedFlowScheduler.load_config(scheduler_config_path)
return RectifiedFlowScheduler.from_config(scheduler_config)
def load_image_to_tensor_with_resize_and_crop(
image_path, target_height=512, target_width=768
):
image = Image.open(image_path).convert("RGB")
input_width, input_height = image.size
aspect_ratio_target = target_width / target_height
aspect_ratio_frame = input_width / input_height
if aspect_ratio_frame > aspect_ratio_target:
new_width = int(input_height * aspect_ratio_target)
new_height = input_height
x_start = (input_width - new_width) // 2
y_start = 0
else:
new_width = input_width
new_height = int(input_width / aspect_ratio_target)
x_start = 0
y_start = (input_height - new_height) // 2
image = image.crop((x_start, y_start, x_start + new_width, y_start + new_height))
image = image.resize((target_width, target_height))
frame_tensor = torch.tensor(np.array(image)).permute(2, 0, 1).float()
frame_tensor = (frame_tensor / 127.5) - 1.0
# Create 5D tensor: (batch_size=1, channels=3, num_frames=1, height, width)
return frame_tensor.unsqueeze(0).unsqueeze(2)
def calculate_padding(
source_height: int, source_width: int, target_height: int, target_width: int
) -> tuple[int, int, int, int]:
# Calculate total padding needed
pad_height = target_height - source_height
pad_width = target_width - source_width
# Calculate padding for each side
pad_top = pad_height // 2
pad_bottom = pad_height - pad_top # Handles odd padding
pad_left = pad_width // 2
pad_right = pad_width - pad_left # Handles odd padding
# Return padded tensor
# Padding format is (left, right, top, bottom)
padding = (pad_left, pad_right, pad_top, pad_bottom)
return padding
def convert_prompt_to_filename(text: str, max_len: int = 20) -> str:
# Remove non-letters and convert to lowercase
clean_text = "".join(
char.lower() for char in text if char.isalpha() or char.isspace()
)
# Split into words
words = clean_text.split()
# Build result string keeping track of length
result = []
current_length = 0
for word in words:
# Add word length plus 1 for underscore (except for first word)
new_length = current_length + len(word)
if new_length <= max_len:
result.append(word)
current_length += len(word)
else:
break
return "-".join(result)
# Generate output video name
def get_unique_filename(
base: str,
ext: str,
prompt: str,
seed: int,
resolution: tuple[int, int, int],
dir: Path,
endswith=None,
index_range=1000,
) -> Path:
base_filename = f"{base}_{convert_prompt_to_filename(prompt, max_len=30)}_{seed}_{resolution[0]}x{resolution[1]}x{resolution[2]}"
for i in range(index_range):
filename = dir / f"{base_filename}_{i}{endswith if endswith else ''}{ext}"
if not os.path.exists(filename):
return filename
raise FileExistsError(
f"Could not find a unique filename after {index_range} attempts."
)
def seed_everething(seed: int):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
def main():
global config
config = OmegaConf.load("config.yaml")
logger = logging.get_logger(__name__)
logger.warning(f"Running generation with arguments: {config}")
seed_everething(config.seed)
output_dir = (
Path(config.output_path)
if config.output_path
else Path(f"outputs/{datetime.today().strftime('%Y-%m-%d')}")
)
output_dir.mkdir(parents=True, exist_ok=True)
# Load image
if config.input_image_path:
media_items_prepad = load_image_to_tensor_with_resize_and_crop(
config.input_image_path, config.height, config.width
)
else:
media_items_prepad = None
height = config.height if config.height else media_items_prepad.shape[-2]
width = config.width if config.width else media_items_prepad.shape[-1]
num_frames = config.num_frames
if height > MAX_HEIGHT or width > MAX_WIDTH or num_frames > MAX_NUM_FRAMES:
logger.warning(
f"Input resolution or number of frames {height}x{width}x{num_frames} is too big, it is suggested to use the resolution below {MAX_HEIGHT}x{MAX_WIDTH}x{MAX_NUM_FRAMES}."
)
# Adjust dimensions to be divisible by 32 and num_frames to be (N * 8 + 1)
height_padded = ((height - 1) // 32 + 1) * 32
width_padded = ((width - 1) // 32 + 1) * 32
num_frames_padded = ((num_frames - 2) // 8 + 1) * 8 + 1
padding = calculate_padding(height, width, height_padded, width_padded)
logger.warning(
f"Padded dimensions: {height_padded}x{width_padded}x{num_frames_padded}"
)
if media_items_prepad is not None:
media_items = F.pad(
media_items_prepad, padding, mode="constant", value=-1
) # -1 is the value for padding since the image is normalized to -1, 1
else:
media_items = None
# Paths for the separate mode directories
ckpt_dir = Path(config.ckpt_dir)
unet_dir = ckpt_dir / "unet"
vae_dir = ckpt_dir / "vae"
scheduler_dir = ckpt_dir / "scheduler"
# Load models
vae = load_vae(vae_dir)
unet = load_unet(unet_dir)
scheduler = load_scheduler(scheduler_dir)
patchifier = SymmetricPatchifier(patch_size=1)
text_encoder = T5EncoderModel.from_pretrained(
"PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="text_encoder"
).to(torch.bfloat16)
tokenizer = T5Tokenizer.from_pretrained(
"PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="tokenizer"
)
if config.bfloat16 and unet.dtype != torch.bfloat16:
unet = unet.to(torch.bfloat16)
# Use submodels for the pipeline
submodel_dict = {
"transformer": unet,
"patchifier": patchifier,
"text_encoder": text_encoder,
"tokenizer": tokenizer,
"scheduler": scheduler,
"vae": vae,
}
pipeline = LTXVideoPipeline(**submodel_dict)
if torch.cuda.is_available() and config.disable_load_needed_only:
pipeline = pipeline.to("cuda")
# Prepare input for the pipeline
sample = {
"prompt": config.prompt,
"prompt_attention_mask": None,
"negative_prompt": config.negative_prompt,
"negative_prompt_attention_mask": None,
"media_items": media_items,
}
generator = torch.Generator(
device="cuda" if torch.cuda.is_available() else "cpu"
).manual_seed(config.seed)
images = pipeline(
num_inference_steps=config.num_inference_steps,
num_images_per_prompt=config.num_images_per_prompt,
guidance_scale=config.guidance_scale,
generator=generator,
output_type="pt",
callback_on_step_end=None,
height=height_padded,
width=width_padded,
num_frames=num_frames_padded,
frame_rate=config.frame_rate,
**sample,
is_video=True,
vae_per_channel_normalize=True,
conditioning_method=(
ConditioningMethod.FIRST_FRAME
if media_items is not None
else ConditioningMethod.UNCONDITIONAL
),
mixed_precision=not config.bfloat16,
load_needed_only=not config.disable_load_needed_only
).images
# Crop the padded images to the desired resolution and number of frames
(pad_left, pad_right, pad_top, pad_bottom) = padding
pad_bottom = -pad_bottom
pad_right = -pad_right
if pad_bottom == 0:
pad_bottom = images.shape[3]
if pad_right == 0:
pad_right = images.shape[4]
images = images[:, :, :num_frames, pad_top:pad_bottom, pad_left:pad_right]
for i in range(images.shape[0]):
# Gathering from B, C, F, H, W to C, F, H, W and then permuting to F, H, W, C
video_np = images[i].permute(1, 2, 3, 0).cpu().float().numpy()
# Unnormalizing images to [0, 255] range
video_np = (video_np * 255).astype(np.uint8)
fps = config.frame_rate
height, width = video_np.shape[1:3]
# In case a single image is generated
if video_np.shape[0] == 1:
output_filename = get_unique_filename(
f"image_output_{i}",
".png",
prompt=config.prompt,
seed=config.seed,
resolution=(height, width, num_frames),
dir=output_dir,
)
imageio.imwrite(output_filename, video_np[0])
else:
if config.input_image_path:
base_filename = f"img_to_vid_{i}"
else:
base_filename = f"text_to_vid_{i}"
output_filename = get_unique_filename(
base_filename,
".mp4",
prompt=config.prompt,
seed=config.seed,
resolution=(height, width, num_frames),
dir=output_dir,
)
# Write video
with imageio.get_writer(output_filename, fps=fps) as video:
for frame in video_np:
video.append_data(frame)
# Write condition image
if config.input_image_path:
reference_image = (
(
media_items_prepad[0, :, 0].permute(1, 2, 0).cpu().data.numpy()
+ 1.0
)
/ 2.0
* 255
)
imageio.imwrite(
get_unique_filename(
base_filename,
".png",
prompt=config.prompt,
seed=config.seed,
resolution=(height, width, num_frames),
dir=output_dir,
endswith="_condition",
),
reference_image.astype(np.uint8),
)
logger.warning(f"Output saved to {output_dir}")
if __name__ == "__main__":
main()