Reverse-engineered prompt schema unlocks reliable character swap in MiniMax H3
photobombolo · reddit · 2026-10-18
The open-weights MiniMax H3 (Ref2VA) model was trained on a rigid six-section prompt schema (subjectdefinitions, summary, retentionanalysis, detaileddescription, overallsoundscape, nondiegeticmusic) with angle-bracketed asset tags and entity aliases. Casual prose or mismatched tags make the Qwen3-VL text encoder fail to bind reference slots, breaking identity swap.
Key rules for 3-image + 1-video character replacement: unify <Picture 1-3> under one alias <Subject 1>, isolate the performer to replace as <Subject 2>, and use exact trigger keywords fullypreserved / attributetransfer / partiallypreserved in retentionanalysis to strip the original actor's look while locking choreography and scene geometry. A copy-paste template is provided.
More from Multimodal
- Dev builds Mumbai-inspired 3D manga background generator from OpenStreetMap data — CurieuxExplorer · 2026-10-18
- Omni 1'1 flash video demo generates atmospheric foggy Victorian haunted house — Dense-Instruction210 · 2026-10-18
- Flux 3 klein distilled tested: int8 workflow delivers surprisingly good 2MP images — Ok_Tale7582 · 2026-10-18
- AI-generated clip with cinematic truck-driver shot fuels 'feature film soon' talk — 1ili0 · 2026-10-18
- AI Videos Are Getting So Good That Real Content Now Gets Doubted — gouterz · 2026-10-18
- Qwen-Image-2.1 LoRA Unlocks Nearly Any Pose, Works for Both T2I and Editing — metalfans · 2026-10-18