Ref2VA face swap keeps the original face: four identity photos lose to the driving video
Motion16AI · reddit · 2026-09-20
Reddit user Motion16AI reports a face-swap failure with MiniMax H3's Ref2VA: rendering a 10-second video with four target-identity photos, the output keeps the original woman almost unchanged.
Setup: minimaxh3ref2vaprunedint8convrot, Heretic INT8 text encoder, 736×736, 240 frames at 24fps, 8-step MiniMaxH3TurboSampler, max-size reference images, video shift 12 / audio shift 3, memory-sparse optimization, no SAM/mask, face-swap adapter, or character LoRA.
The author suspects the full reference video is being treated as appearance conditioning rather than motion-only, overpowering the identity images, and is weighing pose extraction, fewer video reference frames, a corrected first frame, or a separate tracked face-swap stage. Asking the community for help.
More from Multimodal
- Experimental video mixes visual styles, generated with Seedance 2.5 on Adobe Firefly — umesh_ai · 2026-11-04
- Hailuo AI claims H3 scored perfect on language tests; Japanese creator praises its Japanese signage fidelity — Hailuo_AI · 2026-11-04
- YANAMI: 400K-dialog multimodal dataset shows models still fail at stateful multiturn interaction — Shengyu Wang · 2026-11-04
- Open-source Idol model turns a single image into 3D in 45s on 8GB VRAM — Promptmethus · 2026-11-04
- ComfyUI v0.44.0 released with new Patina texture nodes — Fcking_Chuck · 2026-11-04
- Krea 3 technical report released: a deep dive into data, architecture and training — bdsqlsz · 2026-11-04