ComfyUI-Omnichar ships .char format for consistent characters across Flux Klein and MiniMax H3
ashishsanu · reddit · 2026-10-03
A developer released ComfyUI-Omnichar, a set of custom ComfyUI nodes that encode a character—face, body, and wardrobe—into a portable .char file reusable across workflows without identity drift, demoed with Flux 2 Klein 9B for images and MiniMax H3 for video.
Key nodes: Encode Character builds the .char from references; Save Character stores it in models/characters; Decode Character outputs prompt, reference images, a numbered character sheet, and conditioning when wired to CLIP; plus nodes for splitting references and attaching them to conditioning. Decoded characters can even become reference sheets or LoRA adapters.
Prompting tips: name your character and trigger it by name ("sia walking on the beach"), describe features only at encode time, and keep each reference single-element (face-only, body-only) to avoid conflicts—face images are auto-crossed by Sface. GPLv3 open source, with workflow JSONs and a community character library; official node submission pending.
Related event: ComfyUI Open-Source .char Node Locks Character Identity in One File(2 posts)→
More from Multimodal
- Creator uses H3 Max video model on fal to add sign language to movie scenes — PurzBeats · 2026-10-03
- Video Genie: free Mac AI video editor with 1,000+ skills and parallel agents — realmeetjames · 2026-10-03
- Local video models make commercial video generators obsolete for one Redditor — CupQuakeBE · 2026-10-03
- Midjourney v8.2 duotone Xenomorph: neon-noir prompt and parameters shared — michaelrabone · 2026-10-03
- Seedance 2.5 x Runway demo shows a race car speeding as the road vanishes behind it — azed_ai · 2026-10-03
- Geometry Leak: a GPT Image 2.5 prompt trick for premium product ads — aziz4ai · 2026-10-03