rhymes-ai/Allegro

Gallery · GitHub · Blog · Paper · Discord · Join Waitlist (Try it on Discord!)

Gallery

For more demos and corresponding prompts, see the Allegro Gallery.

Key Feature

• Open Source: Full model weights and code 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, which can be interpolated to 30 FPS with 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

Model	Allegro
Description	Text-to-Video Generation Model
Download	Hugging Face
Parameter	VAE: 175M
	DiT: 2.8B
Inference Precision	VAE: FP32/TF32/BF16/FP16 (best in FP32/TF32)
	DiT/T5: BF16/FP32/TF32
Context Length	79.2K
Resolution	720 x 1280
Frames	88
Video Length	6 seconds @ 15 FPS
Single GPU Memory Usage	9.3G BF16 (with cpu_offload)

Quick start

1. Install the necessary requirements.

   • Ensure Python >= 3.10, PyTorch >= 2.4, CUDA >= 12.4.

   • It is recommended to use Anaconda to create a new environment (Python >= 3.10) conda create -n rllegro python=3.10 -y to run the following example.

   • run pip install git+https://github.com/huggingface/diffusers.git@9214f4a3782a74e510eff7e09b59457fe8b63511 torch==2.4.1 transformers==4.40.1 accelerate sentencepiece imageio imageio-ffmpeg beautifulsoup4

2. Run inference.

   import torch
   from diffusers import AutoencoderKLAllegro, AllegroPipeline
   from diffusers.utils import export_to_video
   vae = AutoencoderKLAllegro.from_pretrained("rhymes-ai/Allegro", subfolder="vae", torch_dtype=torch.float32)
   pipe = AllegroPipeline.from_pretrained(
       "rhymes-ai/Allegro", vae=vae, torch_dtype=torch.bfloat16
   )
   pipe.to("cuda")
   pipe.vae.enable_tiling()
   prompt = "A seaside harbor with bright sunlight and sparkling seawater, with many boats in the water. From an aerial view, the boats vary in size and color, some moving and some stationary. Fishing boats in the water suggest that this location might be a popular spot for docking fishing boats."
   
   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.
   """
   
   prompt = prompt.format(prompt.lower().strip())
   
   video = pipe(prompt, negative_prompt=negative_prompt, guidance_scale=7.5, max_sequence_length=512, num_inference_steps=100, generator = torch.Generator(device="cuda:0").manual_seed(42)).frames[0]
   export_to_video(video, "output.mp4", fps=15)
   

   Use pipe.enable_sequential_cpu_offload() to offload the model into CPU for less GPU memory cost (about 9.3G, compared to 27.5G if CPU offload is not enabled), but the inference time will increase significantly.

3. (Optional) Interpolate the video to 30 FPS.

   It is recommended to use EMA-VFI to interpolate the video from 15 FPS to 30 FPS.

   For better visual quality, please use imageio to save the video.

4. For faster inference such Context Parallel, PAB, please refer to our github repo.

License

This repo is released under the Apache 2.0 License.