Real-time streaming audio-video generation with parallel adapters
Real-Time Joint Audio-Video Generation by Parallel Adapter Composition
Computer Vision and Pattern Recognition
Summary
Generating videos with matching sound in real time is hard because models need to produce frames as they go and do so quickly. The authors found a way to combine two improvements in parallel—one allowing the model to work frame-by-frame and another making each step faster—without retraining the combined model. By adding two special modules that handle these parts separately but don’t interfere, their system can make short videos with sound at about 26 frames per second, keeping good quality. This method avoids the usual slow step-by-step training and works smoothly for longer clips.
What this means in practice
- •For video streaming developers: Build systems that generate synchronized video and audio clips interactively at real-time frame rates without long retraining cycles.
- •For real-time media production teams: Produce joint audio-video streams efficiently for interactive media applications using adapters combined in parallel rather than sequential retraining.
Authors
Jingyu Li, Xiaoxiao Xiang, Yiwen Guo
Abstract
Deploying a joint audio-video diffusion transformer for real-time, interactive generation normally requires two essential modifications: block-autoregressive attention, so frames can be emitted before the whole clip is finished, and few-step sampling, so each block is cheap. Conventionally, the streaming video literature obtains both capabilities from a chained pipeline. It first distills a bidirectional teacher into a causal student, then into a few-step one, or proceeds in reverse order. Each stage of such a chain fine-tunes the weights the previous one produced, so a later objective can undo an earlier capability. Following the idea of model merging, we show that on a packed audio-video backbone the two capabilities can be acquired in parallel. A causal adapter is trained against the frozen backbone, and an off-the-shelf few-step adapter provides the few-step capability. As the two edit different functional axes, we predict, and then verify, that their weight-update directions are near-orthogonal, without any explicit orthogonality constraint during training. Orthogonal updates should combine without interfering, so parallel composition is a direct sum. The two adapters are simply added at inference, with no joint training, yielding few-step, streaming audio-video whose image quality tracks the bidirectional teacher. Compared to the chained baselines, the composed model matches or beats them on most metrics, making parallel composition a practical approach. The resulting streaming system generates joint audio-video in real time, $\approx$26 fps at $480\times832$ without quantization, and sustains 30 s of continuous generation with stable image quality.