Spaces:
Running
on
A100
Running
on
A100
different tabs for different functionality
Browse files
app.py
CHANGED
@@ -24,12 +24,14 @@ hf_token = os.getenv("HF_TOKEN")
|
|
24 |
# Set model download directory within Hugging Face Spaces
|
25 |
model_path = "asset"
|
26 |
if not os.path.exists(model_path):
|
27 |
-
snapshot_download(
|
|
|
|
|
28 |
|
29 |
# Global variables to load components
|
30 |
-
vae_dir = Path(model_path) /
|
31 |
-
unet_dir = Path(model_path) /
|
32 |
-
scheduler_dir = Path(model_path) /
|
33 |
|
34 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
35 |
|
@@ -37,7 +39,7 @@ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
37 |
def load_vae(vae_dir):
|
38 |
vae_ckpt_path = vae_dir / "vae_diffusion_pytorch_model.safetensors"
|
39 |
vae_config_path = vae_dir / "config.json"
|
40 |
-
with open(vae_config_path,
|
41 |
vae_config = json.load(f)
|
42 |
vae = CausalVideoAutoencoder.from_config(vae_config)
|
43 |
vae_state_dict = safetensors.torch.load_file(vae_ckpt_path)
|
@@ -69,11 +71,11 @@ def center_crop_and_resize(frame, target_height, target_width):
|
|
69 |
if aspect_ratio_frame > aspect_ratio_target:
|
70 |
new_width = int(h * aspect_ratio_target)
|
71 |
x_start = (w - new_width) // 2
|
72 |
-
frame_cropped = frame[:, x_start:x_start + new_width]
|
73 |
else:
|
74 |
new_height = int(w / aspect_ratio_target)
|
75 |
y_start = (h - new_height) // 2
|
76 |
-
frame_cropped = frame[y_start:y_start + new_height, :]
|
77 |
frame_resized = cv2.resize(frame_cropped, (target_width, target_height))
|
78 |
return frame_resized
|
79 |
|
@@ -116,7 +118,7 @@ preset_options = [
|
|
116 |
{"label": "544x320, 241 frames", "width": 544, "height": 320, "num_frames": 241},
|
117 |
{"label": "512x320, 249 frames", "width": 512, "height": 320, "num_frames": 249},
|
118 |
{"label": "512x320, 257 frames", "width": 512, "height": 320, "num_frames": 257},
|
119 |
-
{"label": "Custom", "height": None, "width": None, "num_frames": None}
|
120 |
]
|
121 |
|
122 |
|
@@ -130,10 +132,17 @@ def preset_changed(preset):
|
|
130 |
selected["num_frames"],
|
131 |
gr.update(visible=False),
|
132 |
gr.update(visible=False),
|
133 |
-
gr.update(visible=False)
|
134 |
)
|
135 |
else:
|
136 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
137 |
|
138 |
|
139 |
# Load models
|
@@ -141,8 +150,12 @@ vae = load_vae(vae_dir)
|
|
141 |
unet = load_unet(unet_dir)
|
142 |
scheduler = load_scheduler(scheduler_dir)
|
143 |
patchifier = SymmetricPatchifier(patch_size=1)
|
144 |
-
text_encoder = T5EncoderModel.from_pretrained(
|
145 |
-
|
|
|
|
|
|
|
|
|
146 |
|
147 |
pipeline = XoraVideoPipeline(
|
148 |
transformer=unet,
|
@@ -154,26 +167,108 @@ pipeline = XoraVideoPipeline(
|
|
154 |
).to(device)
|
155 |
|
156 |
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
163 |
if len(prompt.strip()) < 50:
|
164 |
-
raise gr.Error(
|
|
|
|
|
|
|
165 |
|
166 |
-
if image_path:
|
167 |
-
|
168 |
-
media_items=None
|
169 |
|
|
|
170 |
|
171 |
sample = {
|
172 |
"prompt": prompt,
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
}
|
178 |
|
179 |
generator = torch.Generator(device="cpu").manual_seed(seed)
|
@@ -196,14 +291,16 @@ def generate_video(image_path=None, prompt="", negative_prompt="",
|
|
196 |
vae_per_channel_normalize=True,
|
197 |
conditioning_method=ConditioningMethod.FIRST_FRAME,
|
198 |
mixed_precision=True,
|
199 |
-
callback_on_step_end=gradio_progress_callback
|
200 |
).images
|
201 |
|
202 |
output_path = tempfile.mktemp(suffix=".mp4")
|
203 |
video_np = images.squeeze(0).permute(1, 2, 3, 0).cpu().float().numpy()
|
204 |
video_np = (video_np * 255).astype(np.uint8)
|
205 |
height, width = video_np.shape[1:3]
|
206 |
-
out = cv2.VideoWriter(
|
|
|
|
|
207 |
for frame in video_np[..., ::-1]:
|
208 |
out.write(frame)
|
209 |
out.release()
|
@@ -211,55 +308,133 @@ def generate_video(image_path=None, prompt="", negative_prompt="",
|
|
211 |
return output_path
|
212 |
|
213 |
|
214 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
215 |
with gr.Blocks() as iface:
|
216 |
gr.Markdown("# Video Generation with LTX Video")
|
217 |
|
218 |
-
with gr.
|
219 |
-
with gr.
|
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 |
fn=preset_changed,
|
254 |
-
inputs=[
|
255 |
-
outputs=[
|
256 |
)
|
257 |
|
258 |
-
|
259 |
-
fn=
|
260 |
-
inputs=[
|
261 |
-
|
262 |
-
|
|
|
|
|
|
|
|
|
263 |
)
|
264 |
|
265 |
iface.launch(share=True)
|
|
|
24 |
# Set model download directory within Hugging Face Spaces
|
25 |
model_path = "asset"
|
26 |
if not os.path.exists(model_path):
|
27 |
+
snapshot_download(
|
28 |
+
"Lightricks/LTX-Video", local_dir=model_path, repo_type="model", token=hf_token
|
29 |
+
)
|
30 |
|
31 |
# Global variables to load components
|
32 |
+
vae_dir = Path(model_path) / "vae"
|
33 |
+
unet_dir = Path(model_path) / "unet"
|
34 |
+
scheduler_dir = Path(model_path) / "scheduler"
|
35 |
|
36 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
37 |
|
|
|
39 |
def load_vae(vae_dir):
|
40 |
vae_ckpt_path = vae_dir / "vae_diffusion_pytorch_model.safetensors"
|
41 |
vae_config_path = vae_dir / "config.json"
|
42 |
+
with open(vae_config_path, "r") as f:
|
43 |
vae_config = json.load(f)
|
44 |
vae = CausalVideoAutoencoder.from_config(vae_config)
|
45 |
vae_state_dict = safetensors.torch.load_file(vae_ckpt_path)
|
|
|
71 |
if aspect_ratio_frame > aspect_ratio_target:
|
72 |
new_width = int(h * aspect_ratio_target)
|
73 |
x_start = (w - new_width) // 2
|
74 |
+
frame_cropped = frame[:, x_start : x_start + new_width]
|
75 |
else:
|
76 |
new_height = int(w / aspect_ratio_target)
|
77 |
y_start = (h - new_height) // 2
|
78 |
+
frame_cropped = frame[y_start : y_start + new_height, :]
|
79 |
frame_resized = cv2.resize(frame_cropped, (target_width, target_height))
|
80 |
return frame_resized
|
81 |
|
|
|
118 |
{"label": "544x320, 241 frames", "width": 544, "height": 320, "num_frames": 241},
|
119 |
{"label": "512x320, 249 frames", "width": 512, "height": 320, "num_frames": 249},
|
120 |
{"label": "512x320, 257 frames", "width": 512, "height": 320, "num_frames": 257},
|
121 |
+
{"label": "Custom", "height": None, "width": None, "num_frames": None},
|
122 |
]
|
123 |
|
124 |
|
|
|
132 |
selected["num_frames"],
|
133 |
gr.update(visible=False),
|
134 |
gr.update(visible=False),
|
135 |
+
gr.update(visible=False),
|
136 |
)
|
137 |
else:
|
138 |
+
return (
|
139 |
+
None,
|
140 |
+
None,
|
141 |
+
None,
|
142 |
+
gr.update(visible=True),
|
143 |
+
gr.update(visible=True),
|
144 |
+
gr.update(visible=True),
|
145 |
+
)
|
146 |
|
147 |
|
148 |
# Load models
|
|
|
150 |
unet = load_unet(unet_dir)
|
151 |
scheduler = load_scheduler(scheduler_dir)
|
152 |
patchifier = SymmetricPatchifier(patch_size=1)
|
153 |
+
text_encoder = T5EncoderModel.from_pretrained(
|
154 |
+
"PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="text_encoder"
|
155 |
+
).to(device)
|
156 |
+
tokenizer = T5Tokenizer.from_pretrained(
|
157 |
+
"PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="tokenizer"
|
158 |
+
)
|
159 |
|
160 |
pipeline = XoraVideoPipeline(
|
161 |
transformer=unet,
|
|
|
167 |
).to(device)
|
168 |
|
169 |
|
170 |
+
import gradio as gr
|
171 |
+
import torch
|
172 |
+
from huggingface_hub import snapshot_download
|
173 |
+
|
174 |
+
# [Previous imports remain the same...]
|
175 |
+
|
176 |
+
|
177 |
+
def generate_video_from_text(
|
178 |
+
prompt="",
|
179 |
+
negative_prompt="",
|
180 |
+
seed=171198,
|
181 |
+
num_inference_steps=40,
|
182 |
+
num_images_per_prompt=1,
|
183 |
+
guidance_scale=3,
|
184 |
+
height=512,
|
185 |
+
width=768,
|
186 |
+
num_frames=121,
|
187 |
+
frame_rate=25,
|
188 |
+
progress=gr.Progress(),
|
189 |
+
):
|
190 |
+
if len(prompt.strip()) < 50:
|
191 |
+
raise gr.Error(
|
192 |
+
"Prompt must be at least 50 characters long. Please provide more details for the best results.",
|
193 |
+
duration=5,
|
194 |
+
)
|
195 |
+
|
196 |
+
sample = {
|
197 |
+
"prompt": prompt,
|
198 |
+
"prompt_attention_mask": None,
|
199 |
+
"negative_prompt": negative_prompt,
|
200 |
+
"negative_prompt_attention_mask": None,
|
201 |
+
"media_items": None,
|
202 |
+
}
|
203 |
+
|
204 |
+
generator = torch.Generator(device="cpu").manual_seed(seed)
|
205 |
+
|
206 |
+
def gradio_progress_callback(self, step, timestep, kwargs):
|
207 |
+
progress((step + 1) / num_inference_steps)
|
208 |
+
|
209 |
+
images = pipeline(
|
210 |
+
num_inference_steps=num_inference_steps,
|
211 |
+
num_images_per_prompt=num_images_per_prompt,
|
212 |
+
guidance_scale=guidance_scale,
|
213 |
+
generator=generator,
|
214 |
+
output_type="pt",
|
215 |
+
height=height,
|
216 |
+
width=width,
|
217 |
+
num_frames=num_frames,
|
218 |
+
frame_rate=frame_rate,
|
219 |
+
**sample,
|
220 |
+
is_video=True,
|
221 |
+
vae_per_channel_normalize=True,
|
222 |
+
conditioning_method=ConditioningMethod.FIRST_FRAME,
|
223 |
+
mixed_precision=True,
|
224 |
+
callback_on_step_end=gradio_progress_callback,
|
225 |
+
).images
|
226 |
+
|
227 |
+
output_path = tempfile.mktemp(suffix=".mp4")
|
228 |
+
video_np = images.squeeze(0).permute(1, 2, 3, 0).cpu().float().numpy()
|
229 |
+
video_np = (video_np * 255).astype(np.uint8)
|
230 |
+
height, width = video_np.shape[1:3]
|
231 |
+
out = cv2.VideoWriter(
|
232 |
+
output_path, cv2.VideoWriter_fourcc(*"mp4v"), frame_rate, (width, height)
|
233 |
+
)
|
234 |
+
for frame in video_np[..., ::-1]:
|
235 |
+
out.write(frame)
|
236 |
+
out.release()
|
237 |
+
|
238 |
+
return output_path
|
239 |
+
|
240 |
+
|
241 |
+
def generate_video_from_image(
|
242 |
+
image_path,
|
243 |
+
prompt="",
|
244 |
+
negative_prompt="",
|
245 |
+
seed=171198,
|
246 |
+
num_inference_steps=40,
|
247 |
+
num_images_per_prompt=1,
|
248 |
+
guidance_scale=3,
|
249 |
+
height=512,
|
250 |
+
width=768,
|
251 |
+
num_frames=121,
|
252 |
+
frame_rate=25,
|
253 |
+
progress=gr.Progress(),
|
254 |
+
):
|
255 |
if len(prompt.strip()) < 50:
|
256 |
+
raise gr.Error(
|
257 |
+
"Prompt must be at least 50 characters long. Please provide more details for the best results.",
|
258 |
+
duration=5,
|
259 |
+
)
|
260 |
|
261 |
+
if not image_path:
|
262 |
+
raise gr.Error("Please provide an input image.", duration=5)
|
|
|
263 |
|
264 |
+
media_items = load_image_to_tensor_with_resize(image_path, height, width).to(device)
|
265 |
|
266 |
sample = {
|
267 |
"prompt": prompt,
|
268 |
+
"prompt_attention_mask": None,
|
269 |
+
"negative_prompt": negative_prompt,
|
270 |
+
"negative_prompt_attention_mask": None,
|
271 |
+
"media_items": media_items,
|
272 |
}
|
273 |
|
274 |
generator = torch.Generator(device="cpu").manual_seed(seed)
|
|
|
291 |
vae_per_channel_normalize=True,
|
292 |
conditioning_method=ConditioningMethod.FIRST_FRAME,
|
293 |
mixed_precision=True,
|
294 |
+
callback_on_step_end=gradio_progress_callback,
|
295 |
).images
|
296 |
|
297 |
output_path = tempfile.mktemp(suffix=".mp4")
|
298 |
video_np = images.squeeze(0).permute(1, 2, 3, 0).cpu().float().numpy()
|
299 |
video_np = (video_np * 255).astype(np.uint8)
|
300 |
height, width = video_np.shape[1:3]
|
301 |
+
out = cv2.VideoWriter(
|
302 |
+
output_path, cv2.VideoWriter_fourcc(*"mp4v"), frame_rate, (width, height)
|
303 |
+
)
|
304 |
for frame in video_np[..., ::-1]:
|
305 |
out.write(frame)
|
306 |
out.release()
|
|
|
308 |
return output_path
|
309 |
|
310 |
|
311 |
+
def create_advanced_options():
|
312 |
+
with gr.Accordion("Advanced Options", open=False):
|
313 |
+
seed = gr.Slider(label="Seed", minimum=0, maximum=1000000, step=1, value=171198)
|
314 |
+
inference_steps = gr.Slider(
|
315 |
+
label="Inference Steps", minimum=1, maximum=100, step=1, value=40
|
316 |
+
)
|
317 |
+
images_per_prompt = gr.Slider(
|
318 |
+
label="Images per Prompt", minimum=1, maximum=10, step=1, value=1
|
319 |
+
)
|
320 |
+
guidance_scale = gr.Slider(
|
321 |
+
label="Guidance Scale", minimum=1.0, maximum=20.0, step=0.1, value=3.0
|
322 |
+
)
|
323 |
+
|
324 |
+
height_slider = gr.Slider(
|
325 |
+
label="Height", minimum=256, maximum=1024, step=64, value=704, visible=False
|
326 |
+
)
|
327 |
+
width_slider = gr.Slider(
|
328 |
+
label="Width", minimum=256, maximum=1024, step=64, value=1216, visible=False
|
329 |
+
)
|
330 |
+
num_frames_slider = gr.Slider(
|
331 |
+
label="Number of Frames",
|
332 |
+
minimum=1,
|
333 |
+
maximum=200,
|
334 |
+
step=1,
|
335 |
+
value=41,
|
336 |
+
visible=False,
|
337 |
+
)
|
338 |
+
frame_rate = gr.Slider(
|
339 |
+
label="Frame Rate", minimum=1, maximum=60, step=1, value=25, visible=False
|
340 |
+
)
|
341 |
+
|
342 |
+
return [
|
343 |
+
seed,
|
344 |
+
inference_steps,
|
345 |
+
images_per_prompt,
|
346 |
+
guidance_scale,
|
347 |
+
height_slider,
|
348 |
+
width_slider,
|
349 |
+
num_frames_slider,
|
350 |
+
frame_rate,
|
351 |
+
]
|
352 |
+
|
353 |
+
|
354 |
+
# Define the Gradio interface with tabs
|
355 |
with gr.Blocks() as iface:
|
356 |
gr.Markdown("# Video Generation with LTX Video")
|
357 |
|
358 |
+
with gr.Tabs():
|
359 |
+
with gr.TabItem("Text to Video"):
|
360 |
+
with gr.Row():
|
361 |
+
with gr.Column():
|
362 |
+
txt2vid_prompt = gr.Textbox(
|
363 |
+
label="Prompt",
|
364 |
+
value="A man riding a motorcycle down a winding road, surrounded by lush, green scenery and distant mountains. The sky is clear with a few wispy clouds, and the sunlight glistens on the motorcycle as it speeds along. The rider is dressed in a black leather jacket and helmet, leaning slightly forward as the wind rustles through nearby trees. The wheels kick up dust, creating a slight trail behind the motorcycle, adding a sense of speed and excitement to the scene.",
|
365 |
+
)
|
366 |
+
txt2vid_negative_prompt = gr.Textbox(
|
367 |
+
label="Negative Prompt",
|
368 |
+
value="worst quality, inconsistent motion...",
|
369 |
+
)
|
370 |
+
|
371 |
+
# Preset dropdown for resolution and frame settings
|
372 |
+
txt2vid_preset = gr.Dropdown(
|
373 |
+
choices=[p["label"] for p in preset_options],
|
374 |
+
value="1216x704, 41 frames",
|
375 |
+
label="Resolution Preset",
|
376 |
+
)
|
377 |
+
|
378 |
+
txt2vid_advanced = create_advanced_options()
|
379 |
+
txt2vid_generate = gr.Button("Generate Video")
|
380 |
+
|
381 |
+
with gr.Column():
|
382 |
+
txt2vid_output = gr.Video(label="Generated Video")
|
383 |
+
|
384 |
+
with gr.TabItem("Image to Video"):
|
385 |
+
with gr.Row():
|
386 |
+
with gr.Column():
|
387 |
+
img2vid_image = gr.Image(type="filepath", label="Input Image")
|
388 |
+
img2vid_prompt = gr.Textbox(
|
389 |
+
label="Prompt",
|
390 |
+
value="A man riding a motorcycle down a winding road, surrounded by lush, green scenery and distant mountains...",
|
391 |
+
)
|
392 |
+
img2vid_negative_prompt = gr.Textbox(
|
393 |
+
label="Negative Prompt",
|
394 |
+
value="worst quality, inconsistent motion...",
|
395 |
+
)
|
396 |
+
|
397 |
+
img2vid_preset = gr.Dropdown(
|
398 |
+
choices=[p["label"] for p in preset_options],
|
399 |
+
value="1216x704, 41 frames",
|
400 |
+
label="Resolution Preset",
|
401 |
+
)
|
402 |
+
|
403 |
+
img2vid_advanced = create_advanced_options()
|
404 |
+
img2vid_generate = gr.Button("Generate Video")
|
405 |
+
|
406 |
+
with gr.Column():
|
407 |
+
img2vid_output = gr.Video(label="Generated Video")
|
408 |
+
|
409 |
+
# Event handlers for text-to-video tab
|
410 |
+
txt2vid_preset.change(
|
411 |
+
fn=preset_changed,
|
412 |
+
inputs=[txt2vid_preset],
|
413 |
+
outputs=txt2vid_advanced[4:], # height, width, num_frames, and their visibility
|
414 |
+
)
|
415 |
+
|
416 |
+
txt2vid_generate.click(
|
417 |
+
fn=generate_video_from_text,
|
418 |
+
inputs=[txt2vid_prompt, txt2vid_negative_prompt, *txt2vid_advanced],
|
419 |
+
outputs=txt2vid_output,
|
420 |
+
)
|
421 |
+
|
422 |
+
# Event handlers for image-to-video tab
|
423 |
+
img2vid_preset.change(
|
424 |
fn=preset_changed,
|
425 |
+
inputs=[img2vid_preset],
|
426 |
+
outputs=img2vid_advanced[4:], # height, width, num_frames, and their visibility
|
427 |
)
|
428 |
|
429 |
+
img2vid_generate.click(
|
430 |
+
fn=generate_video_from_image,
|
431 |
+
inputs=[
|
432 |
+
img2vid_image,
|
433 |
+
img2vid_prompt,
|
434 |
+
img2vid_negative_prompt,
|
435 |
+
*img2vid_advanced,
|
436 |
+
],
|
437 |
+
outputs=img2vid_output,
|
438 |
)
|
439 |
|
440 |
iface.launch(share=True)
|