|
--- |
|
license: apache-2.0 |
|
language: |
|
- en |
|
library_name: diffusers |
|
--- |
|
<p align="center"> |
|
<img src="https://huggingface.co/rhymes-ai/Allegro/resolve/main/banner_white.gif" width="500" height="400"/> |
|
</p> |
|
<p align="center"> |
|
<a href="https://rhymes.ai/" target="_blank"> Gallery</a> 路 <a href="https://github.com/rhymes-ai/Aria" target="_blank">GitHub</a> 路 <a href="https://www.rhymes.ai/blog-details/" target="_blank">Blog</a> 路 <a href="https://arxiv.org/pdf/2410.05993" target="_blank">Paper</a> 路 <a href="https://discord" target="_blank">Discord</a> |
|
|
|
</p> |
|
|
|
# Gallery |
|
<img src="https://huggingface.co/rhymes-ai/Allegro/resolve/main/gallery.gif" width="1000" height="800"/>For more demos and corresponding prompts, see the [Allegro Gallery](TBD). |
|
|
|
|
|
# Key Feature |
|
|
|
- **Open Source**: [Full model weights](https://huggingface.co/rhymes-ai/Allegro) and [code](https://github.com/rhymes-ai/Allegro) available to the community, Apache 2.0! |
|
- **Versatile Content Creation**: Capable of generating a wide range of content, from close-ups of humans and animals to diverse dynamic scenes. |
|
- **High-Quality Output**: Generate detailed 6-second videos at 15 FPS with 720x1280 resolution, can be interpolated to 30 FPS with [EMA-VFI](https://github.com/MCG-NJU/EMA-VFI). |
|
- **Small and Efficient**: Features a 175M parameter VideoVAE and a 2.8B parameter VideoDiT model. Supports multiple precisions (FP32, BF16, FP16) and uses 9.3 GB of GPU memory in BF16 mode with CPU offloading. Context length is 79.2k, equivalent to 88 frames. |
|
|
|
# Model info |
|
|
|
<table> |
|
<tr> |
|
<th>Model</th> |
|
<td>Allegro</td> |
|
</tr> |
|
<tr> |
|
<th>Description</th> |
|
<td>Text-to-Video Generation Model</td> |
|
</tr> |
|
<tr> |
|
<th>Download</th> |
|
<td><HF link - TBD></td> |
|
</tr> |
|
<tr> |
|
<th rowspan="2">Parameter</th> |
|
<td>VAE: 175M</td> |
|
</tr> |
|
<tr> |
|
<td>DiT: 2.8B</td> |
|
</tr> |
|
<tr> |
|
<th rowspan="2">Inference Precision</th> |
|
<td>VAE: FP32/TF32/BF16/FP16 (best in FP32/TF32)</td> |
|
</tr> |
|
<tr> |
|
<td>DiT/T5: BF16/FP32/TF32</td> |
|
</tr> |
|
<tr> |
|
<th>Context Length</th> |
|
<td>79.2k</td> |
|
</tr> |
|
<tr> |
|
<th>Resolution</th> |
|
<td>720 x 1280</td> |
|
</tr> |
|
<tr> |
|
<th>Frames</th> |
|
<td>88</td> |
|
</tr> |
|
<tr> |
|
<th>Video Length</th> |
|
<td>6 seconds @ 15 fps</td> |
|
</tr> |
|
<tr> |
|
<th>Single GPU Memory Usage</th> |
|
<td>9.3G BF16 (with cpu_offload)</td> |
|
</tr> |
|
</table> |
|
|
|
|
|
# Quick start |
|
You can quickly get started with Allegro using the Hugging Face Diffusers library. |
|
For more tutorials, see Allegro GitHub (link-tbd). |
|
|
|
1. Install necessary requirements. Please refer to [requirements.txt](https://github.com/rhymes-ai) on Allegro GitHub. |
|
2. Perform inference on a single GPU. |
|
```python |
|
from diffusers import DiffusionPipeline |
|
import torch |
|
|
|
allegro_pipeline = DiffusionPipeline.from_pretrained( |
|
"rhymes-ai/Allegro", trust_remote_code=True, torch_dtype=torch.bfloat16 |
|
).to("cuda") |
|
|
|
allegro_pipeline.vae = allegro_pipeline.vae.to(torch.float32) |
|
|
|
prompt = "a video of an astronaut riding a horse on mars" |
|
|
|
positive_prompt = """ |
|
(masterpiece), (best quality), (ultra-detailed), (unwatermarked), |
|
{} |
|
emotional, harmonious, vignette, 4k epic detailed, shot on kodak, 35mm photo, |
|
sharp focus, high budget, cinemascope, moody, epic, gorgeous |
|
""" |
|
|
|
negative_prompt = """ |
|
nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, |
|
low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry. |
|
""" |
|
|
|
num_sampling_steps, guidance_scale, seed = 100, 7.5, 42 |
|
|
|
user_prompt = positive_prompt.format(args.user_prompt.lower().strip()) |
|
out_video = allegro_pipeline( |
|
user_prompt, |
|
negative_prompt=negative_prompt, |
|
num_frames=88, |
|
height=720, |
|
width=1280, |
|
num_inference_steps=num_sampling_steps, |
|
guidance_scale=guidance_scale, |
|
max_sequence_length=512, |
|
generator = torch.Generator(device="cuda:0").manual_seed(seed) |
|
).video[0] |
|
|
|
imageio.mimwrite("test_video.mp4", out_video, fps=15, quality=8) |
|
``` |
|
Tip: |
|
- It is highly recommended to use a video frame interpolation model (such as EMA-VFI) to enhance the result to 30 FPS. |
|
- For more tutorials, see [Allegro GitHub](https://github.com/rhymes-ai). |
|
|
|
# License |
|
This repo is released under the Apache 2.0 License. |
|
|
