Workflows / MiniMax H3
Free workflow · new
Flow Long — 10 to 30 seconds in one job
One job, one file, continuous sound.
MiniMax H3 with ComfyUI's native context windows: sampling runs in overlapping 124-frame windows on H3's own chunk grid, so a 12-second clip renders in one pass with its audio intact — no chunk files, no joins. An optional refine pass re-samples the finished clip at 1.5× with the core H3 latent upscaler, inside the same windows.
Drag the .json onto ComfyUI. Runs best on the Gmanski RunPod template.
You get
- Flow (Long) workflow with the bypassed 1.5× refine pass
- A Sol-H3-VAE fast-decoder slot to try
You need
- ComfyUI 0.44+
- 48 GB+ VRAM (96 GB for refined 20 s+ clips)
Model packs · Video
minimax-h3-flowminimax-h3-refineOn the template: gmanski-packs install minimax-h3-flow minimax-h3-refine
Pairs well with.
MiniMax H3newCharacter Video Swap — one photo, any performance
A guide video gives the moves. Your photo gives the person.
LTX 2.5Long Video Upscale — LTX 2.5
The whole clip, not the first three seconds. 2× and sharper, audio kept.
MiniMax H3Character Ensemble Video
The same cast, moving — with sound.