What we work on
Audio editing by latent inpainting
A two-stage system — a continuous neural codec plus a flow-matching transformer — that regenerates a masked span of speech conditioned on surrounding audio and target text, zero-shot, with inaudible boundaries.
Anchored tree sampling beats autoregressive drift
A training-free, inference-time scheduler that replaces left-to-right rollout with anchor-bounded tree imputation — converting horizon-compounding drift into bounded drift, and running 5.3× faster than autoregressive baselines.
Audio-driven lip-sync by masked frame regeneration
A flow-matching transformer that regenerates the lower face from new audio. Whisper-encoded speech and reference frames condition the model; masked velocity prediction during training keeps identity, lighting, and the boundary to untouched video in place.
Invisible jumpcuts by regenerating the bridge
Trim a talking-head clip and the head jumps to a new pose. Jumpcut Smoothing masks the frames across the cut and generates a short bridge, then runs Video Regenerate over it to re-sync the lips — so the join plays like a continuous take.
