ID-V2V weights land on Hugging Face: Wan 2.1 + VACE, two checkpoints including normal/depth control
minchoi · x · 2026-09-08
The ID-V2V model card is now public on Hugging Face (Eyeline-Labs/ID-V2V, Apache 2.0):
- Paper accepted at SIGGRAPH Asia 2026, code open-sourced on GitHub
- Architecture: two finetuned checkpoints on Wan 2.1 image-to-video with VACE control:
- idv2v.pth (recommended): single condition — segmented subject on gray pixels via SAM3, relit and preserved while the rest of the frame is regenerated from the prompt
- idv2vwithnormaldepth.pth: adds surface normals (DAViD) and depth (DepthAnything-V2) for tighter geometric control
- Supports "shoot first, restyle later": preserves identity, expressions, gaze, and motion from the source video
More from Multimodal
- Spatial-first AI video: map the scene with GPT-6 Astra before rendering — HeyZoyaKhan · 2026-09-09
- GPT-6 Astra codes geometry, Blender + Clay plugin renders, Dreamina finalizes footage — thetripathi58 · 2026-09-09
- Claude skips the training set: HTML/CSS rendered in a headless browser makes his site images — shashib · 2026-09-09
- A SpaceX homage built with Grok + Intangible shows AI 3D video creation in action — bilawalsidhu · 2026-09-09
- Marigold V2 launches at SIGGRAPH Asia 2026: sharp diffusion-transformer depth estimation — AntonObukhov1 · 2026-09-09
- AI-generated POV: riding a dragon through a medieval town — Brave-Wishbone-3650 · 2026-09-08