Open-source TaoMate-FVSR restores face videos with audio conditioning and visual dubbing
Apprehensive_Sky892 · reddit · 2026-11-04
TaoMate-FVSR is an open-source audio-guided facial video restoration model, released with a paper, project page, and Hugging Face weights.
- Core capability: restores face videos using the original audio, sparse visual references, and generated-frame history; supports restoration with original audio or visual dubbing with replacement speech
- Pipeline: DWPose crop → face restoration → color/mask blend → paste-back; full-video inference with 1×/2× modes (512×512 or 1024×1024), compositing back to the original resolution
- Flexible references: GPEN-enhanced first frame, externally edited first frame, or a finished reference image; default enhancement is one GPEN-BFR-512 pass for 1× and GPEN-BFR-1024 for 2×
The repo contains only the standalone inference runtime; weights download separately and media stays local.
Related event: Alibaba Open-Sources TaoMate-FVSR for Audio-Guided Facial Video Restoration(2 posts)→
More from Multimodal
- Open image model MOSS slammed for photos that look smeared in vaseline — mogged_by_dasha · 2026-11-04
- Reddit user open-sources full MiniMax H3 ComfyUI workflows: TTS, image editing, 3MP upscale — Support_Marmoset · 2026-11-04
- Dev builds Grok-powered game dev bot producing detailed voxel characters and starships — Promptmethus · 2026-11-04
- HeyGen's HyperFrames skill turns text into editable MP4 videos in Perplexity — toolstelegraph · 2026-11-04
- AI video drop: Draconia's Halloween Date, part 2 — R34vspec · 2026-11-04
- AI-Generated Stop-Motion Series: Cat & Chicken Roommates Steal a Gorilla — RealBrainRocket · 2026-11-04