RefMods workflow keeps characters and voices consistent in MiniMax H3 on a 6 GB GPU

EzkaProductions · reddit · 2026-11-05

A hands-on test of RefMods with MiniMax H3 in ComfyUI solves AI video's biggest pain point: consistent characters and voices across clips.

Extract a reference once (character from 8 images, voice from a short clip), save as a reusable "mod", load into any future generation — no retraining, no LoRA
6 GB VRAM setup: H3 SLA Attention + Low VRAM Attention + Chunk FeedForward + Turbo LoRA; visual and audio strengths at 0.9
Results: very consistent visual identity across scenes; voices stay consistent clip to clip
Tip for two characters: keep separate mod files and describe each subject clearly in the prompt
Required nodes: ComfyUI-MiniMaxH3Mod, ComfyUI-PlagueKind-Nodes, ComfyUI-KJNodes; full video walkthrough and free workflow linked

More from Multimodal