---
license: apache-2.0
pipeline_tag: any-to-any
tags:
- image-to-video
- image-text-to-video
- image-to-audio-video
- image-text-to-audio-video
- MOVA
- OpenMOSS
- SII
- MOSI
- sglang-diffusion
---
## MOVA: Towards Scalable and Synchronized Video–Audio Generation
We introduce **MOVA** (**MO**SS **V**ideo and **A**udio), a foundation model designed to break the "silent era" of open-source video generation. Unlike cascaded pipelines that generate sound as an afterthought, MOVA synthesizes video and audio simultaneously for perfect alignment.
🌟Key Highlights
- **Native Bimodal Generation**: Moves beyond clunky cascaded pipelines. MOVA generates high-fidelity video and synchronized audio in a single inference pass, eliminating error accumulation.
- **Precise Lip-Sync & Sound FX**: Achieves state-of-the-art performance in multilingual lip-synchronization and environment-aware sound effects.
- **Fully Open-Source**: In a field dominated by closed-source models (Sora 2, Veo 3, Kling), we are releasing model weights, inference code, training pipelines, and LoRA fine-tuning scripts.
- **Asymmetric Dual-Tower Architecture**: Leverages the power of pre-trained video and audio towers, fused via a bidirectional cross-attention mechanism for rich modality interaction.
## Demo
## Model Details
### Model Description
MOVA addresses the limitations of proprietary systems like Sora 2 and Veo 3 by offering a fully open-source framework for Image-to-Video-Audio (IT2VA) and Text-to-Video-Audio (T2VA) tasks. The model employs an asymmetric dual-tower architecture fused via a bidirectional cross-attention mechanism, leveraging a Mixture-of-Experts (MoE) design with 32B total parameters (18B active during inference) to ensure high-quality synthesis with efficient deployment. Alongside the model weights, we provide a fine-grained bimodal data pipeline and support for LoRA fine-tuning, empowering the community to advance research in synchronized cinematic synthesis.
### Model Sources
- **Project Page:** https://mosi.cn/models/mova
- **Github:** https://github.com/OpenMOSS/MOVA
- **Paper:** [MOVA: Towards Scalable and Synchronized Video-Audio Generation](https://huggingface.co/papers/2602.08794)
## Model Usage
Please refer to the [GitHub repository](https://github.com/OpenMOSS/MOVA) for environment setup and detailed instructions.
### Sample Inference
Generate a video of single person speech:
```bash
export CP_SIZE=1
export CKPT_PATH=/path/to/MOVA-360p/
torchrun \
--nproc_per_node=$CP_SIZE \
scripts/inference_single.py \
--ckpt_path $CKPT_PATH \
--cp_size $CP_SIZE \
--height 352 \
--width 640 \
--prompt "A man in a blue blazer and glasses speaks in a formal indoor setting, framed by wooden furniture and a filled bookshelf. Quiet room acoustics underscore his measured tone as he delivers his remarks. At one point, he says, \"I would also say that this election in Germany wasn’t surprising.\"" \
--ref_path "./assets/single_person.jpg" \
--output_path "./data/samples/single_person.mp4" \
--seed 42 \
--offload cpu
```
## Evaluation
We evaluate our model through both objective benchmarks and subjective human evaluations. Below are the Elo scores and win rates comparing MOVA to existing open-source models.
## Citation
```bibtex
@article{yu2026mova,
title={MOVA: Towards Scalable and Synchronized Video-Audio Generation},
author={Donghua Yu and Mingshu Chen and Qi Chen and Qi Luo and Qianyi Wu and Qinyuan Cheng and others},
journal={arXiv preprint arXiv:2602.08794},
year={2026}
}
```
