← sources

MiniMax H3 video: how to keep faces consistent during dynamic camera motion?

source: reddit (reddit:StableDiffusion)published 2026-11-13 11:34ingested 2026-11-14 02:38

retained text

**Question**: I can't make my rendering pipeline produce dynamic motion videos without glitches. On the attached video you see the face/eyes are slightly 'glitching/wobbling', how to fix that?

**Config**: MiniMax H3 video (ref2va DiT, **w6a8** quant) via ComfyUI, 1344×768 @ 24 fps, 192 frames (8 s), single link with exact latent carry (no seams, no re-conditioning)

**Prompt** was: "*She stands centered on the foggy harbour promenade at dawn in a crimson wool coat with a knitted scarf, still, facing the camera; the camera orbits one slow full revolution around her, from directly in front, around her right side, behind her, around her left side, and back to the front, holding her full figure at a constant size in frame, about two fifths of the frame height, the rail, fishing boats and dock lamps swinging smoothly through the mist behind her. Foggy dawn light, soft grey-blue grading, no added contrast*"

**Conditioning**: prompt + start\_image (exact first frame) + 2 refs (wide still + tight portrait of the same woman, Qwen3-VL encoder). Sampler: res\_multistep, 32 steps, cfg 1, fixed seed (so no turbo LoRA, no seams, refs present; the wobble persists anyway)

document fetched 2026-11-14 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xv5jrj/minimax_h3_video_how_to_keep_faces_consistent