RefMod Pilot: LoRA-free consistent characters in MiniMax H3 with 1-8 reference photos
no3us · reddit · 2026-10-10
The author built RefMod Pilot, a workflow package for MiniMax H3 character references that uses Full Reference encoding instead of LoRA training: drop in 1-8 photos, name the character, and it becomes a reusable RefMod; multiple photos are stacked into one video-kind reference.
Two workflows are included: Creator — upload/drag photos, give the character a name, run; folder and image count are auto-filled. Inference — pick RefMods, edit the scene and view output, with prompt, aspect ratio, resolution, seconds, seed, and steps consolidated into one Scene node and the rest in a subgraph. The character loader stays visible so Add RefMod, Refresh, and library buttons remain available. For multi-character scenes, check the Reference map and match its labels in the prompt — slot numbers aren't necessarily picture numbers.
A six-view orange robot RefMod and demo workflow are bundled. Creation is verified but likeness and motion in finished videos haven't been validated. Disclosure: the author makes LoRA Pilot; this is their workflow/UX layer over Luisacaotica's ComfyUI-MiniMaxH3Mod. Available on GitHub and Hugging Face.
Related event: RefMod Workflow Brings Training-Free Character Consistency to MiniMax H3(3 posts)→
More from Multimodal
- "They finally invented tech to get my girlfriend to watch The Sopranos": AI video goes viral — NathanpmYoung · 2026-10-12
- AI-generated Monkey King beats everyone except a pelican in viral video — lucky-plume · 2026-10-12
- Video Prompt Craft: Splitting 30 Seconds into Timed Beats with Camera Notes — lmoroney · 2026-10-12
- JESTER: The Last Laugh — a 9-minute AI sci-fi short film released on Reddit — theodore_70 · 2026-10-12
- Open-Source Long Shot Studio v0.6.13 Adds Multi-Take Batch Renders and Faster RTX Upscaling — thetalkinghawk · 2026-10-12
- Hook viewers with Magnific One: familiar-but-striking starting images — techhalla · 2026-10-12