Open-source .char file locks face, body and clothing consistency in Flux and Minimax H3
ashishsanu · reddit · 2026-09-27
The author open-sourced OmniChar (GPLv3): a portable .char file that packs one character's face, body and clothing references to keep identity consistent across Flux 2 Klein 9b stills and Minimax H3 video.
How it works:
- Building a .char: up to 9 references, ideally a 2:2:1 face:clothing:body ratio; YuNet finds faces, SFace takes per-reference face signatures, DINOv2 takes a subject signature, then everything is cleaned and packed into one file (face-only required)
- At generation, the .char feeds references into Flux's native multi-reference channel and prepends a locked description to the prompt
Prompting notes:
- Name your character and just say "sia walking on the beach"; don't re-describe features in the generation prompt
- Only add specific overrides (e.g. half sleeves) there; clothing/body references can interfere and cause drift
- Use cropped, single-subject references; conflicting faces across references cause generation conflicts
Requirements: 24GB+ VRAM Nvidia GPU and 64GB RAM locally; Minimax H3 now has an int8 variant. Workflows and GitHub repo linked in the post.
More from Multimodal
- One prompt makes Claude Opus produce a polished app launch video — BLCNYY · 2026-09-28
- Recreating Friends Entirely Locally with MiniMax H3, Upscaling and VFI Pipeline — alisitskii · 2026-09-28
- How a Pro Musician Uses Stable Audio and Moises in a Real AI Music Workflow — jordiponsdotme · 2026-09-28
- AI-Made 'Akira: The Awakening' Cinematic Anime Trailer Shows Video Generation Progress — AbstractCodex12 · 2026-09-28
- Opus 5.5 dug through 3.7GB of logs to produce a 15-minute documentary of its Unciv match — Angaisb_ · 2026-09-28
- Homemade Lemmings Recreated with Gemini 3.8 Flash — Artistic_Solution117 · 2026-09-28