MoDA

Overview

Talking head videos where speech matches face motion are important for avatars and virtual uses. But earlier methods often don’t perform well. They can be slow and produce visual glitches or poorly synced mouth movement.

MoDA a new model released July 4 2025 by Xinyang Li and team tries to fix that with a few smart changes.

Instead of splitting the process into separate steps for how a face moves and how it looks MoDA brings them together in a shared setup. That makes training and running it more efficient.

It also uses flow matching to help guide the learning so it runs more smoothly.

MoDA listens to speech tracks and tracks noisy face movement then mixes that with other hints. This happens inside a multi-modal diffusion model which helps create more detailed facial expressions and better synced lip movement.

Rather than combining all data at once MoDA adds things in stages starting broad then moving to fine details. That makes the final result clearer and more natural.

MoDA outperforms JoyVASA which uses two separate stages. MoDA creates more expressive facial motion more accurate lip sync and head movement that looks less robotic.

It also runs faster and needs less computing power which means it’s practical for everyday use.