Pusa: Open-Source Video Diffusion Model Fine-Tuned from Mochi with Text/Image/Video-to-Video Support 50

What is Pusa?

Pusa is a new AI video generation model released by Raphael Liu in April 2025. Its core innovation lies in redefining temporal modeling in video diffusion through a novel diffusion paradigm with frame-level noise control. The Pusa model supports multiple generation tasks including text-to-video, image-to-video, and video-to-video, while maintaining motion fidelity and prompt adherence, and offers the advantages of low cost and high performance.

Features of the Pusa Model

• Frame-Level Noise Control and Vectorized Timesteps: Pusa introduces frame-level noise control and vectorized timesteps, an innovation originally proposed in the FVDM paper. Compared with traditional methods, Pusa increases the number of timesteps from the conventional 1,000 to several thousand, greatly enhancing the model's flexibility and scalability. This frame-level noise control enables Pusa to handle the content of each frame more precisely during the video generation process, thereby achieving higher-quality video generation results.

• Non-Destructive Modification: Pusa's adaptation to the base model preserves its original text-to-video generation capabilities. With only light fine-tuning, Pusa can adapt to various video generation tasks without requiring large-scale modifications to the base model. This non-destructive modification not only saves development time and cost but also ensures the stability and reliability of the model.

• Broad Applicability: Pusa's approach is not limited to the Mochi model and can be easily applied to other leading video diffusion models, such as Hunyuan Video and Wan2.1.

Pusa Model Architecture

• Frame-Level Noise Control: Pusa implements frame-level noise control and vectorized timesteps, providing unprecedented flexibility and scalability.

• Non-Destructive Modification: The modifications to the base model preserve its original text-to-video generation capabilities, requiring only light fine-tuning.

• Broad Applicability: The approach can be easily applied to other leading video diffusion models, such as Hunyuan Video and Wan2.1.

Pusa Application Scenarios

• Suitable for film and television special effects production, advertising creative generation, and personalized video content production.

• Combined with multimodal inputs (such as text + image) to achieve complex scene simulation, enhancing the flexibility of video generation.

Pusa Installation and Usage

Installation:

git clone https://github.com/genmoai/models
cd models
pip install uv
uv venv .venv
source .venv/bin/activate
uv pip install setuptools
uv pip install -e . --no-build-isolation

If you need to install Flash Attention, you can use:

uv pip install -e .[flash] --no-build-isolation

Downloading Weights:

• Using the Hugging Face CLI:

pip install huggingface_hub
huggingface-cli download RaphaelLiu/Pusa-V0.5 --local-dir

Or download directly from Hugging Face to your local machine.

Basic Usage:

• Text-to-Video Generation:

python ./demos/cli_test_ti2v_release.py \
  --model_dir "/path/to/Pusa-V0.5" \
  --dit_path "/path/to/Pusa-V0.5/pusa_v0_dit.safetensors" \
  --prompt "A man is playing basketball" \
  --num_steps 30

• Image-to-Video Generation:

python ./demos/cli_test_ti2v_release.py \
  --model_dir "/path/to/Pusa-V0.5" \
  --dit_path "/path/to/Pusa-V0.5/pusa_v0_dit.safetensors" \
  --prompt "Your_prompt_here" \
  --image_dir "/path/to/input/image.jpg" \
  --cond_position 1 \
  --num_steps 30

Training:

Complete Pusa training code and detailed information are provided, supporting training on the original model.

Code Repository:https://github.com/Yaofang-Liu/Pusa-VidGen

HuggingFace:https://huggingface.co/RaphaelLiu/Pusa-V0.5