fix(wip): second pass
Browse files
app.py
CHANGED
@@ -7,6 +7,7 @@ from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepSchedu
|
|
7 |
from diffusers.schedulers.scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
|
8 |
import os
|
9 |
import tempfile
|
|
|
10 |
|
11 |
# Define model options
|
12 |
MODEL_OPTIONS = {
|
@@ -43,7 +44,7 @@ def generate_video(
|
|
43 |
second_pass_flow_shift,
|
44 |
second_pass_cfg,
|
45 |
show_both_outputs
|
46 |
-
):
|
47 |
# Get model ID from selection
|
48 |
model_id = MODEL_OPTIONS[model_choice]
|
49 |
|
@@ -98,23 +99,16 @@ def generate_video(
|
|
98 |
num_frames=num_frames,
|
99 |
guidance_scale=guidance_scale,
|
100 |
num_inference_steps=num_inference_steps,
|
101 |
-
|
|
|
102 |
return_dict=True
|
103 |
)
|
104 |
|
105 |
-
# Get the latents from the first pass output
|
106 |
-
|
107 |
|
108 |
-
#
|
109 |
if not enable_second_pass or (enable_second_pass and show_both_outputs):
|
110 |
-
# Decode the latents to frames with the VAE (only needed if we requested latents)
|
111 |
-
if enable_second_pass:
|
112 |
-
print("Decoding first pass latents...")
|
113 |
-
with torch.no_grad():
|
114 |
-
first_pass_frames = pipe.vae.decode(latents / pipe.vae.config.scaling_factor).sample
|
115 |
-
else:
|
116 |
-
first_pass_frames = latents
|
117 |
-
|
118 |
# Export first pass to video
|
119 |
first_pass_file = "output_first_pass.mp4"
|
120 |
export_to_video(first_pass_frames, first_pass_file, fps=output_fps)
|
@@ -125,6 +119,14 @@ def generate_video(
|
|
125 |
if enable_second_pass:
|
126 |
print("Running second pass with scale factor:", second_pass_scale)
|
127 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
128 |
# Resize latents for second pass (upscale)
|
129 |
new_height = int(height * second_pass_scale)
|
130 |
new_width = int(width * second_pass_scale)
|
@@ -135,10 +137,18 @@ def generate_video(
|
|
135 |
|
136 |
print(f"Upscaling latents from {height}x{width} to {new_height}x{new_width}")
|
137 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
138 |
# Upscale latents using interpolate
|
139 |
upscaled_latents = torch.nn.functional.interpolate(
|
140 |
latents,
|
141 |
-
size=(num_frames,
|
142 |
mode="trilinear",
|
143 |
align_corners=False
|
144 |
)
|
@@ -183,15 +193,18 @@ def generate_video(
|
|
183 |
output_files.append(second_pass_file)
|
184 |
|
185 |
# Return the appropriate video output(s)
|
186 |
-
if enable_second_pass and
|
187 |
-
return
|
188 |
-
elif
|
189 |
-
return [
|
190 |
else:
|
191 |
-
return
|
192 |
|
193 |
-
# Create the Gradio interface
|
194 |
with gr.Blocks() as demo:
|
|
|
|
|
|
|
195 |
gr.HTML("""
|
196 |
<p align="center">
|
197 |
<svg version="1.1" viewBox="0 0 1200 295" xmlns="http://www.w3.org/2000/svg" xmlns:v="https://vecta.io/nano" width="400">
|
@@ -364,34 +377,33 @@ with gr.Blocks() as demo:
|
|
364 |
output_video = gr.Video(label="Generated Video")
|
365 |
second_output_video = gr.Video(label="Second Pass Video", visible=False)
|
366 |
|
367 |
-
#
|
368 |
-
def
|
369 |
-
return
|
370 |
|
|
|
371 |
enable_second_pass.change(
|
372 |
-
fn=
|
373 |
inputs=[enable_second_pass, show_both_outputs],
|
374 |
outputs=[second_output_video]
|
375 |
)
|
376 |
|
377 |
show_both_outputs.change(
|
378 |
-
fn=
|
379 |
inputs=[enable_second_pass, show_both_outputs],
|
380 |
outputs=[second_output_video]
|
381 |
)
|
382 |
|
383 |
-
#
|
384 |
-
def
|
385 |
-
|
386 |
-
|
387 |
-
return [result[0], result[1], {"visible": True}]
|
388 |
-
elif isinstance(result, list) and len(result) == 1:
|
389 |
-
return [result[0], None, {"visible": False}]
|
390 |
else:
|
391 |
-
return
|
392 |
-
|
|
|
393 |
generate_btn.click(
|
394 |
-
fn=
|
395 |
inputs=[
|
396 |
model_choice,
|
397 |
prompt,
|
@@ -416,12 +428,23 @@ with gr.Blocks() as demo:
|
|
416 |
show_both_outputs
|
417 |
],
|
418 |
outputs=[
|
419 |
-
output_video,
|
420 |
-
second_output_video,
|
421 |
-
second_output_video # Update visibility
|
422 |
]
|
423 |
)
|
424 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
425 |
gr.Markdown("""
|
426 |
## Tips for best results:
|
427 |
- For smaller resolution videos, try lower values of flow shift (2.0-5.0)
|
|
|
7 |
from diffusers.schedulers.scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
|
8 |
import os
|
9 |
import tempfile
|
10 |
+
from typing import List, Union, Optional
|
11 |
|
12 |
# Define model options
|
13 |
MODEL_OPTIONS = {
|
|
|
44 |
second_pass_flow_shift,
|
45 |
second_pass_cfg,
|
46 |
show_both_outputs
|
47 |
+
) -> Union[str, List[str]]:
|
48 |
# Get model ID from selection
|
49 |
model_id = MODEL_OPTIONS[model_choice]
|
50 |
|
|
|
99 |
num_frames=num_frames,
|
100 |
guidance_scale=guidance_scale,
|
101 |
num_inference_steps=num_inference_steps,
|
102 |
+
# For Wan, we may need to approach this differently for the latents
|
103 |
+
output_type="pt", # Always get PyTorch tensors for the first pass
|
104 |
return_dict=True
|
105 |
)
|
106 |
|
107 |
+
# Get the frames or latents from the first pass output
|
108 |
+
first_pass_frames = first_pass.frames[0]
|
109 |
|
110 |
+
# Output the first pass video if needed
|
111 |
if not enable_second_pass or (enable_second_pass and show_both_outputs):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
# Export first pass to video
|
113 |
first_pass_file = "output_first_pass.mp4"
|
114 |
export_to_video(first_pass_frames, first_pass_file, fps=output_fps)
|
|
|
119 |
if enable_second_pass:
|
120 |
print("Running second pass with scale factor:", second_pass_scale)
|
121 |
|
122 |
+
# For second pass, we need to first encode the frames to get latents
|
123 |
+
print("Encoding first pass frames to latents...")
|
124 |
+
with torch.no_grad():
|
125 |
+
# Move frames to the same device as the VAE
|
126 |
+
first_pass_frames = first_pass_frames.to(pipe.vae.device)
|
127 |
+
# Encode to get latents
|
128 |
+
latents = pipe.vae.encode(first_pass_frames).latent_dist.sample()
|
129 |
+
|
130 |
# Resize latents for second pass (upscale)
|
131 |
new_height = int(height * second_pass_scale)
|
132 |
new_width = int(width * second_pass_scale)
|
|
|
137 |
|
138 |
print(f"Upscaling latents from {height}x{width} to {new_height}x{new_width}")
|
139 |
|
140 |
+
# Get latent dimensions
|
141 |
+
latent_height = latents.shape[2] # Should be height//8
|
142 |
+
latent_width = latents.shape[3] # Should be width//8
|
143 |
+
|
144 |
+
# Calculate new latent dimensions
|
145 |
+
new_latent_height = new_height // 8
|
146 |
+
new_latent_width = new_width // 8
|
147 |
+
|
148 |
# Upscale latents using interpolate
|
149 |
upscaled_latents = torch.nn.functional.interpolate(
|
150 |
latents,
|
151 |
+
size=(num_frames, new_latent_height, new_latent_width),
|
152 |
mode="trilinear",
|
153 |
align_corners=False
|
154 |
)
|
|
|
193 |
output_files.append(second_pass_file)
|
194 |
|
195 |
# Return the appropriate video output(s)
|
196 |
+
if enable_second_pass and show_both_outputs and len(output_files) > 1:
|
197 |
+
return output_files
|
198 |
+
elif len(output_files) > 0:
|
199 |
+
return output_files[-1] # Return the last generated output (either first or second pass)
|
200 |
else:
|
201 |
+
return "No video was generated. Please check the logs for errors."
|
202 |
|
203 |
+
# Create the Gradio interface
|
204 |
with gr.Blocks() as demo:
|
205 |
+
# Import gr.update for visibility control
|
206 |
+
from gradio import update
|
207 |
+
|
208 |
gr.HTML("""
|
209 |
<p align="center">
|
210 |
<svg version="1.1" viewBox="0 0 1200 295" xmlns="http://www.w3.org/2000/svg" xmlns:v="https://vecta.io/nano" width="400">
|
|
|
377 |
output_video = gr.Video(label="Generated Video")
|
378 |
second_output_video = gr.Video(label="Second Pass Video", visible=False)
|
379 |
|
380 |
+
# Control visibility through the UI changes directly
|
381 |
+
def toggle_second_video(enable_pass, show_both):
|
382 |
+
return gr.update(visible=enable_pass and show_both)
|
383 |
|
384 |
+
# Update visibility when checkboxes change
|
385 |
enable_second_pass.change(
|
386 |
+
fn=toggle_second_video,
|
387 |
inputs=[enable_second_pass, show_both_outputs],
|
388 |
outputs=[second_output_video]
|
389 |
)
|
390 |
|
391 |
show_both_outputs.change(
|
392 |
+
fn=toggle_second_video,
|
393 |
inputs=[enable_second_pass, show_both_outputs],
|
394 |
outputs=[second_output_video]
|
395 |
)
|
396 |
|
397 |
+
# Define a visibility update function separately
|
398 |
+
def update_second_video_visibility(enable_pass, show_both):
|
399 |
+
if enable_pass and show_both:
|
400 |
+
return gr.update(visible=True)
|
|
|
|
|
|
|
401 |
else:
|
402 |
+
return gr.update(visible=False)
|
403 |
+
|
404 |
+
# Process generation without trying to update visibility in the same function
|
405 |
generate_btn.click(
|
406 |
+
fn=generate_video,
|
407 |
inputs=[
|
408 |
model_choice,
|
409 |
prompt,
|
|
|
428 |
show_both_outputs
|
429 |
],
|
430 |
outputs=[
|
431 |
+
output_video if not show_both_outputs else [output_video, second_output_video]
|
|
|
|
|
432 |
]
|
433 |
)
|
434 |
|
435 |
+
# Update visibility when options change
|
436 |
+
enable_second_pass.change(
|
437 |
+
fn=update_second_video_visibility,
|
438 |
+
inputs=[enable_second_pass, show_both_outputs],
|
439 |
+
outputs=[second_output_video]
|
440 |
+
)
|
441 |
+
|
442 |
+
show_both_outputs.change(
|
443 |
+
fn=update_second_video_visibility,
|
444 |
+
inputs=[enable_second_pass, show_both_outputs],
|
445 |
+
outputs=[second_output_video]
|
446 |
+
)
|
447 |
+
|
448 |
gr.Markdown("""
|
449 |
## Tips for best results:
|
450 |
- For smaller resolution videos, try lower values of flow shift (2.0-5.0)
|