Open-source avatar model built upon LongCat-Video (meigen-ai.github.io)

The introduction of LongCat-Video-Avatar marks a significant advancement in audio-driven video synthesis, allowing for the generation of super-realistic, lip-synchronized videos that maintain consistent identity and natural dynamics over extended durations. This model supports various generation modes, including text-audio-to-video and image-text-audio-to-video, tackling persistent challenges in long-duration video generation, such as error accumulation and identity drift, which have plagued existing methods. Key innovations include the Disentangled Unconditional Guidance strategy, which separates audio signals from body motion to ensure more fluid movements during silent segments, and a Reference Skip Attention mechanism that balances visual fidelity with dynamic motion, avoiding rigid artifacts while preserving identity. Additionally, the Cross-Chunk Latent Stitching approach minimizes pixel degradation caused by repetitive neural network cycles, enhancing the overall quality of the generated output. This research not only enhances the quality of long-duration video generation but also opens avenues for multi-person scenarios and richer animation experiences, making it a vital contribution to the AI/ML community.