MiniMax H3 character consistency
retained text
I'm probably quite late to the party, but my little video editor project has reached a level where it can actually be used, so perhaps someone might find it useful. The focus is on creating reusable character reels, environments and props, and make it easy to feed them into the ref2va variant of H3. From my poorly executed demo video above, it can be noted that environment consistency can be improved. But it seems this can often be improved by generating a few different takes at a low resolution and picking the best seed. It connects to a running instance of ComfyUI to queue custom created workflows and keep track of the outputs. Prompt assistance is also a main part, using either text models through ComfyUI or cloud based LLMs. Finally, the projects are made to be shared easily - so you can download the source for the above video and make a better version! Source: [radiatingreverberations/lumibelle](https://github.com/radiatingreverberations/lumibelle) Manual: [Personal video studio | Lumibelle](https://lumibelle.ai/)
document fetched 2026-09-30 00:35 · canonical https://reddit.com/r/StableDiffusion/comments/1wtfwmp/minimax_h3_character_consistency