A new paper on arXiv introduces UniAVC, a unified model for audio-video creation. The model is designed to handle several related tasks in one framework: generating both audio and video from a text prompt, converting from one modality to the other in either direction, and editing existing content based on instructions.
The authors note that existing methods rely heavily on paired audio-video data, which can limit their flexibility. UniAVC instead uses progressive training and asymmetric scheduling, though the abstract does not spell out the exact mechanics of these techniques.
If successful, this unified approach could simplify workflows that currently require separate models for generation and editing. The paper is available on arXiv under the identifier 2610.24791.