MOVA

Why we included this project

Most open-source video generators stop at silent frames, leaving you to sync audio yourself. MOVA is one of the few fully open projects that produces the video and its soundtrack in a single pass, so speech lands on the lips and sound effects track the action without drift. Under the hood it couples a pretrained video model with an audio tower using cross-attention, which keeps things familiar if you've worked with diffusion tooling. The repo goes beyond weights: there's inference code, a CPU-offload path that runs on a single 4090, ComfyUI integration, and LoRA fine-tuning scripts, so you can actually put it to work. And the bundled evaluation metrics for lip-sync and audio-visual alignment give you a concrete way to judge output, which helps when you're comparing models for talking-head or multilingual narration work.