Spaces:
Running
on
Zero
Running
on
Zero
ui
Browse files
app.py
CHANGED
|
@@ -289,7 +289,7 @@ def call_pipe(
|
|
| 289 |
def main():
|
| 290 |
# Define the interface inputs
|
| 291 |
|
| 292 |
-
with gr.Blocks() as demo:
|
| 293 |
with gr.Group():
|
| 294 |
gr.Markdown("""# HunyuanVideo Keyframe Control Lora for Video Generation
|
| 295 |
**Generate videos using the HunyuanVideo model with a prompt and two (or more) frames as conditions. Gradio / HF Spaces implementation demo.**
|
|
@@ -301,66 +301,61 @@ def main():
|
|
| 301 |
I will fill out a request for GPU allocation for the demo with HF soon.
|
| 302 |
|
| 303 |
""")
|
| 304 |
-
gr.Markdown("")
|
| 305 |
-
gr.Markdown("")
|
| 306 |
-
gr.Markdown("---")
|
| 307 |
-
gr.Markdown("")
|
| 308 |
|
| 309 |
-
|
| 310 |
-
with gr.Column():
|
| 311 |
-
prompt_textbox = gr.Textbox(label="Prompt", value="a subject ... ")
|
| 312 |
-
resolution = gr.Dropdown(
|
| 313 |
-
label="Resolution",
|
| 314 |
-
choices=["720x1280", "544x960", "1280x720", "960x544", "720x720"],
|
| 315 |
-
value="544x960"
|
| 316 |
-
)
|
| 317 |
-
with gr.Column():
|
| 318 |
-
guidance_scale = gr.Slider(minimum=0.1, maximum=20, step=0.1, label="Guidance Scale", value=6.0)
|
| 319 |
-
num_frames = gr.Slider(minimum=1, maximum=129, step=1, label="Number of Frames", value=49)
|
| 320 |
|
| 321 |
with gr.Row():
|
| 322 |
with gr.Column():
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 323 |
frame1 = gr.Image(label="Frame 1", type="pil")
|
| 324 |
frame2 = gr.Image(label="Frame 2", type="pil")
|
| 325 |
num_inference_steps = gr.Slider(minimum=1, maximum=100, step=1, label="Number of Inference Steps", value=30)
|
|
|
|
|
|
|
| 326 |
generate_button = gr.Button("Generate Video")
|
| 327 |
with gr.Column():
|
| 328 |
outputs = gr.Video(label="Generated Video")
|
| 329 |
-
|
| 330 |
-
|
| 331 |
-
|
| 332 |
|
| 333 |
-
|
| 334 |
-
|
| 335 |
-
|
| 336 |
-
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
* The model is conditioned on user-defined keyframes, allowing precise control over the generated video's start and end frames. This conditioning ensures that the generated content aligns seamlessly with the specified keyframes, enhancing the coherence and narrative flow of the video.
|
| 340 |
|
| 341 |
-
|
| 342 |
-
|
| 343 |
-
|
|
|
|
| 344 |
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
|
| 349 |
-
|
| 350 |
-
|
|
|
|
| 351 |
|
| 352 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 353 |
|
| 354 |
-
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
3. It is recommended to set frames from 33 upto 97. Can go upto 121 frames as well (but not tested much).
|
| 358 |
-
4. Prompting helps a lot but works even without. The prompt can be as simple as just the name of the object you want to generate or can be detailed.
|
| 359 |
-
5. `num_inference_steps` is recommended to be 50, but for fast results you can use 30 as well. Anything less than 30 is not recommended.
|
| 360 |
|
| 361 |
-
## FINAL THOUGHTS: This ZeroGPU space, while successfully loaded, has its memory packed to the rim. If you're lucky you may be able to sneak in a small demo inference here and there, but you will most definitely not be using the recommended settings listed above. Help, of course, is not only welcome but very much appreciated. Learn more about our non-profit initiative, [AI Without Borders](https://huggingface.co/aiwithoutborders-xyz), by following us on Huggingface or on [X](http://x.com/borderlesstools), where we will be announcing a handful of exciting developments.
|
| 362 |
-
|
| 363 |
-
""", sanitize_html=False, elem_id="md_footer", container=True)
|
| 364 |
generate_button.click(generate_video, inputs=[prompt_textbox, frame1, frame2, resolution, guidance_scale, num_frames, num_inference_steps], outputs=outputs)
|
| 365 |
|
| 366 |
demo.launch(show_error=True)
|
|
|
|
| 289 |
def main():
|
| 290 |
# Define the interface inputs
|
| 291 |
|
| 292 |
+
with gr.Blocks(css=".gradio-container { max-width: 80vw; margin: 0 auto; }, /* Target all media elements (img, video, audio) within table cells */ tr td img, tr td video, tr td audio { max-height: 268px; object-fit: contain; display: block; width: auto; } /* Target all table cells within table rows */ tr td { overflow: hidden; }") as demo:
|
| 293 |
with gr.Group():
|
| 294 |
gr.Markdown("""# HunyuanVideo Keyframe Control Lora for Video Generation
|
| 295 |
**Generate videos using the HunyuanVideo model with a prompt and two (or more) frames as conditions. Gradio / HF Spaces implementation demo.**
|
|
|
|
| 301 |
I will fill out a request for GPU allocation for the demo with HF soon.
|
| 302 |
|
| 303 |
""")
|
|
|
|
|
|
|
|
|
|
|
|
|
| 304 |
|
| 305 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 306 |
|
| 307 |
with gr.Row():
|
| 308 |
with gr.Column():
|
| 309 |
+
with gr.Row():
|
| 310 |
+
prompt_textbox = gr.Textbox(label="Prompt", value="a subject ...", scale=2)
|
| 311 |
+
resolution = gr.Dropdown(
|
| 312 |
+
label="Resolution",
|
| 313 |
+
choices=["720x1280", "544x960", "1280x720", "960x544", "720x720"],
|
| 314 |
+
value="544x960"
|
| 315 |
+
)
|
| 316 |
frame1 = gr.Image(label="Frame 1", type="pil")
|
| 317 |
frame2 = gr.Image(label="Frame 2", type="pil")
|
| 318 |
num_inference_steps = gr.Slider(minimum=1, maximum=100, step=1, label="Number of Inference Steps", value=30)
|
| 319 |
+
guidance_scale = gr.Slider(minimum=0.1, maximum=20, step=0.1, label="Guidance Scale", value=6.0)
|
| 320 |
+
num_frames = gr.Slider(minimum=1, maximum=129, step=1, label="Number of Frames", value=49)
|
| 321 |
generate_button = gr.Button("Generate Video")
|
| 322 |
with gr.Column():
|
| 323 |
outputs = gr.Video(label="Generated Video")
|
| 324 |
+
with gr.Accordion(label="Examples"):
|
| 325 |
+
gr.Markdown("""
|
| 326 |
+
<div>
|
| 327 |
|
| 328 |
+
| Starting Frame | End Frame | Generated Video |
|
| 329 |
+
|---------|---------|-----------------|
|
| 330 |
+
| <img src="https://content.dashtoon.ai/stability-images/41aeca63-064a-4003-8c8b-bfe2cc80d275.png" > |  | <video controls autoplay src="https://content.dashtoon.ai/stability-images/14b7dd1a-1f46-4c4c-b4ec-9d0f948712af.mp4"></video> |
|
| 331 |
+
|  |  | <video controls autoplay src="https://content.dashtoon.ai/stability-images/b00ba193-b3b7-41a1-9bc1-9fdaceba6efa.mp4"></video> |
|
| 332 |
+
|  |  | <video controls autoplay src="https://content.dashtoon.ai/stability-images/0cb84780-4fdf-4ecc-ab48-12e7e1055a39.mp4"></video> |
|
| 333 |
+
|  |  | <video controls autoplay src="https://content.dashtoon.ai/stability-images/ce12156f-0ac2-4d16-b489-37e85c61b5b2.mp4"></video> |
|
|
|
|
| 334 |
|
| 335 |
+
</div>
|
| 336 |
+
""", sanitize_html=False)
|
| 337 |
+
with gr.Accordion():
|
| 338 |
+
gr.Markdown("""
|
| 339 |
|
| 340 |
+
## HunyuanVideo Keyframe Control Lora is an adapter for HunyuanVideo T2V model for keyframe-based video generation.
|
| 341 |
+
---
|
| 342 |
+
**Our architecture builds upon existing models, introducing key enhancements to optimize keyframe-based video generation**:
|
| 343 |
+
|
| 344 |
+
* We modify the input patch embedding projection layer to effectively incorporate keyframe information. By adjusting the convolutional input parameters, we enable the model to process image inputs within the Diffusion Transformer (DiT) framework.
|
| 345 |
+
* We apply Low-Rank Adaptation (LoRA) across all linear layers and the convolutional input layer. This approach facilitates efficient fine-tuning by introducing low-rank matrices that approximate the weight updates, thereby preserving the base model's foundational capabilities while reducing the number of trainable parameters.
|
| 346 |
+
* The model is conditioned on user-defined keyframes, allowing precise control over the generated video's start and end frames. This conditioning ensures that the generated content aligns seamlessly with the specified keyframes, enhancing the coherence and narrative flow of the video.
|
| 347 |
|
| 348 |
+
## Recommended Settings
|
| 349 |
+
1. The model works best on human subjects. Single subject images work slightly better.
|
| 350 |
+
2. It is recommended to use the following image generation resolutions `720x1280`, `544x960`, `1280x720`, `960x544`.
|
| 351 |
+
3. It is recommended to set frames from 33 upto 97. Can go upto 121 frames as well (but not tested much).
|
| 352 |
+
4. Prompting helps a lot but works even without. The prompt can be as simple as just the name of the object you want to generate or can be detailed.
|
| 353 |
+
5. `num_inference_steps` is recommended to be 50, but for fast results you can use 30 as well. Anything less than 30 is not recommended.
|
| 354 |
|
| 355 |
+
## FINAL THOUGHTS: This ZeroGPU space, while successfully loaded, has its memory packed to the rim. If you're lucky you may be able to sneak in a small demo inference here and there, but you will most definitely not be using the recommended settings listed above. Help, of course, is not only welcome but very much appreciated. Learn more about our non-profit initiative, [AI Without Borders](https://huggingface.co/aiwithoutborders-xyz), by following us on Huggingface or on [X](http://x.com/borderlesstools), where we will be announcing a handful of exciting developments.
|
| 356 |
+
|
| 357 |
+
""", sanitize_html=False, elem_id="md_footer", container=True)
|
|
|
|
|
|
|
|
|
|
| 358 |
|
|
|
|
|
|
|
|
|
|
| 359 |
generate_button.click(generate_video, inputs=[prompt_textbox, frame1, frame2, resolution, guidance_scale, num_frames, num_inference_steps], outputs=outputs)
|
| 360 |
|
| 361 |
demo.launch(show_error=True)
|