SepGen: one model generates video and per-source audio stems in a single pass
Aviad Dahan · hf · 2026-10-12
A Tel Aviv University and Google paper presents SepGen, which extends a pretrained audio-video generator to emit the video, the mixed soundtrack, and one waveform per captioned source in a single sampling run. Generation mode assigns stems by caption (e.g., one per speaker in order); separation mode extracts sources described in free text from a clean mix. It outperforms language-conditioned separators, especially on speech. Code, checkpoints, and datasets are open-sourced.
Related event: SepGen: One Model Generates Video with Separate Audio Stems(3 posts)→
More from Multimodal
- ThinkDrag: mmlab-ntu adds visual reasoning to drag-based image editing, claims SOTA — mmlab-ntu · 2026-10-12
- Edit videos with Claude Code: 22-min HyperFrames tutorial goes free — toolstelegraph · 2026-10-12
- Midjourney recipe: sref + raw mode yields Kodak film-style golden hour portraits — michaelrabone · 2026-10-12
- Seedance 2.5 Renders Nostalgic 'Visiting Grandparents' Clip, Prompt Included — umesh_ai · 2026-10-12
- Runway Agent on iOS: film-style Japan travel imagery channeling Elliott Erwitt — jongranskog · 2026-10-12
- From Beethoven to Suno: full breakdown of how the AI duet WANTED was made — tetsuoai · 2026-10-12