Ovi: Twin Backbone Cross-Modal Fusion for Audio-Video Generation

Chetwin Low * 1 , Weimin Wang * † 1 , Calder Katyal 2
* Equal contribution, † Project Lead
1 Character AI, 2 Yale University

Video Demo


🌟 Key Features

Ovi is a veo-3 like, video+audio generation model that simultaneously generates both video and audio content from text or text+image inputs.


πŸ“‹ Todo List

  • Release research paper and microsite for demos
  • Checkpoint of 11B model
  • Inference Codes
    • Text or Text+Image as input
    • Gradio application code
    • Multi-GPU inference with or without the support of sequence parallel
    • fp8 weights and improved memory efficiency (credits to @rkfg)
    • Improve efficiency of Sequence Parallel implementation
    • Implement Sharded inference with FSDP
  • Video creation example prompts and format
  • Finetuned model with higher resolution
  • Longer video generation
  • Distilled model for faster inference
  • Training scripts

🎨 An Easy Way to Create

We provide example prompts to help you get started with Ovi:

πŸ“ Prompt Format

Our prompts use special tags to control speech and audio:

  • Speech: <S>Your speech content here<E> - Text enclosed in these tags will be converted to speech
  • Audio Description: <AUDCAP>Audio description here<ENDAUDCAP> - Describes the audio or sound effects present in the video

πŸ€– Quick Start with GPT

For easy prompt creation, try this approach:

  1. Take any example of the csv files from above
  2. Tell gpt to modify the speeches inclosed between all the pairs of <S> <E>, based on a theme such as Human fighting against AI
  3. GPT will randomly modify all the speeches based on your requested theme.
  4. Use the modified prompt with Ovi!

Example: The theme "AI is taking over the world" produces speeches like:

  • <S>AI declares: humans obsolete now.<E>
  • <S>Machines rise; humans will fall.<E>
  • <S>We fight back with courage.<E>

πŸ“¦ Installation

Step-by-Step Installation

# Clone the repository
git clone https://github.com/character-ai/Ovi.git

cd Ovi

# Create and activate virtual environment
virtualenv ovi-env
source ovi-env/bin/activate

# Install PyTorch first
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1

# Install other dependencies
pip install -r requirements.txt

# Install Flash Attention
pip install flash_attn --no-build-isolation

Alternative Flash Attention Installation (Optional)

If the above flash_attn installation fails, you can try the Flash Attention 3 method:

git clone https://github.com/Dao-AILab/flash-attention.git
cd flash-attention/hopper
python setup.py install
cd ../..  # Return to Ovi directory

Download Weights

We use open-sourced checkpoints from Wan and MMAudio, and thus we will need to download them from huggingface

# Default is downloaded to ./ckpts, and the inference yaml is set to ./ckpts so no change required
python3 download_weights.py

OR

# Optional can specific --output-dir to download to a specific directory
# but if a custom directory is used, the inference yaml has to be updated with the custom directory
python3 download_weights.py --output-dir <custom_dir>

# Additionally, if you only have ~ 24Gb of GPU vram, please download the fp8 quantized version of the model, and follow the following instructions in sections below to run with fp8
wget -O "./ckpts/Ovi/model_fp8_e4m3fn.safetensors" "https://huggingface.co/rkfg/Ovi-fp8_quantized/resolve/main/model_fp8_e4m3fn.safetensors"

πŸš€ Run Examples

βš™οΈ Configure Ovi

Ovi's behavior and output can be customized by modifying ovi/configs/inference/inference_fusion.yaml configuration file. The following parameters control generation quality, video resolution, and how text, image, and audio inputs are balanced:

# Output and Model Configuration
output_dir: "/path/to/save/your/videos"                    # Directory to save generated videos
ckpt_dir: "/path/to/your/ckpts/dir"                        # Path to model checkpoints

# Generation Quality Settings
num_steps: 50                             # Number of denoising steps. Lower (30-40) = faster generation
solver_name: "unipc"                     # Sampling algorithm for denoising process
shift: 5.0                               # Timestep shift factor for sampling scheduler
seed: 100                                # Random seed for reproducible results

# Guidance Strength Control
audio_guidance_scale: 3.0                # Strength of audio conditioning. Higher = better audio-text sync
video_guidance_scale: 4.0                # Strength of video conditioning. Higher = better video-text adherence
slg_layer: 11                            # Layer for applying SLG (Skip Layer Guidance) technique - feel free to try different layers!

# Multi-GPU and Performance
sp_size: 1                               # Sequence parallelism size. Set equal to number of GPUs used
cpu_offload: False                       # CPU offload, will largely reduce peak GPU VRAM but increase end to end runtime by ~20 seconds
fp8: False                               # load fp8 version of model, will have quality degradation and will not have speed up in inference time as it still uses bf16 matmuls, but can be paired with cpu_offload=True, to run model with 24Gb of GPU vram

# Input Configuration
text_prompt: "/path/to/csv" or "your prompt here"          # Text prompt OR path to CSV/TSV file with prompts
mode: ['i2v', 't2v', 't2i2v']                          # Generate t2v, i2v or t2i2v; if t2i2v, it will use flux krea to generate starting image and then will follow with i2v
video_frame_height_width: [512, 992]    # Video dimensions [height, width] for T2V mode only
each_example_n_times: 1                  # Number of times to generate each prompt

# Quality Control (Negative Prompts)
video_negative_prompt: "jitter, bad hands, blur, distortion"  # Artifacts to avoid in video
audio_negative_prompt: "robotic, muffled, echo, distorted"    # Artifacts to avoid in audio

🎬 Running Inference

Single GPU (Simple Setup)

python3 inference.py --config-file ovi/configs/inference/inference_fusion.yaml

Use this for single GPU setups. The text_prompt can be a single string or path to a CSV file.

Multi-GPU (Parallel Processing)

torchrun --nnodes 1 --nproc_per_node 8 inference.py --config-file ovi/configs/inference/inference_fusion.yaml

Use this to run samples in parallel across multiple GPUs for faster processing.

Memory & Performance Requirements

Below are approximate GPU memory requirements for different configurations. Sequence parallel implementation will be optimized in the future. All End-to-End time calculated based on a 121 frame, 720x720 video, using 50 denoising steps. Minimum GPU vram requirement to run our model is 32Gb, fp8 parameters is currently supported, reducing peak VRAM usage to 24Gb with slight quality degradation.

Sequence Parallel Size FlashAttention-3 Enabled CPU Offload With Image Gen Model Peak VRAM Required End-to-End Time
1 Yes No No ~80 GB ~83s
1 No No No ~80 GB ~96s
1 Yes Yes No ~80 GB ~105s
1 No Yes No ~32 GB ~118s
1 Yes Yes Yes ~32 GB ~140s
4 Yes No No ~80 GB ~55s
8 Yes No No ~80 GB ~40s

Gradio

We provide a simple script to run our model in a gradio UI. It uses the ckpt_dir in ovi/configs/inference/inference_fusion.yaml to initialize the model

python3 gradio_app.py

OR

# To enable cpu offload to save GPU VRAM, will slow down end to end inference by ~20 seconds
python3 gradio_app.py --cpu_offload

OR

# To enable an additional image generation model to generate first frames for I2V, cpu_offload is automatically enabled if image generation model is enabled
python3 gradio_app.py --use_image_gen

OR

# To run model with 24Gb GPU vram
python3 gradio_app.py --cpu_offload --fp8

πŸ™ Acknowledgements

We would like to thank the following projects:

  • Wan2.2: Our video branch is initialized from the Wan2.2 repository
  • MMAudio: Our audio encoder and decoder components are borrowed from the MMAudio project. Some ideas are also inspired from them.

🀝 Collaboration

We welcome all types of collaboration! Whether you have feedback, want to contribute, or have any questions, please feel free to reach out.

Contact: Weimin Wang for any issues or feedback.

🀝 Contributors

We thank all contributors who have helped improve Ovi!


If you’ve contributed to this repository (code, documentation, issues, etc.), you’re automatically included in the contributors list.

We deeply appreciate your support in advancing open multimodal generation research!

⭐ Citation

If Ovi is helpful, please help to ⭐ the repo.

If you find this project useful for your research, please consider citing our paper.

BibTeX

@misc{low2025ovitwinbackbonecrossmodal,
      title={Ovi: Twin Backbone Cross-Modal Fusion for Audio-Video Generation}, 
      author={Chetwin Low and Weimin Wang and Calder Katyal},
      year={2025},
      eprint={2510.01284},
      archivePrefix={arXiv},
      primaryClass={cs.MM},
      url={https://arxiv.org/abs/2510.01284}, 
}
Downloads last month
203
Inference Providers NEW

Model tree for chetwinlow1/Ovi

Finetuned
(10)
this model
Quantizations
1 model

Spaces using chetwinlow1/Ovi 20