MiniMax H3 in ComfyUI: 4 speed-ups tested on the same seed
MiniMax H3 in ComfyUI: four ways to make it faster, tested on the same seed
One MiniMax H3 clip took about 85 seconds of drawing time on my PC. I wanted it faster without losing the scene, the voice or the lip sync. So I ran the same prompt with the same seed five times: once normal and once with each of four speed-ups.
What I tested
Three sets: a woman pouring tea (seed 7), and hands shaping a clay bowl while a cat watches (seed 7 and seed 11). Each clip is about 5 seconds with sound.
Four speed-ups: EasyCache, the H3 Forecast Speed Up node from Stable Yogi Nodes (mild and strong), TeaCache, and a 3-step speed add-on.
Same prompt, same seed, same settings. Only the speed-up changes.
Drawing time (mean of the three sets)
Normal: 84.5 s
EasyCache: 66.5 s
Speed Up node, mild: 59.1 s
Speed Up node, strong: 42.3 s
TeaCache: 40.0 s
3-step add-on: 14.5 s
What changes in the clip
Faster is not free. The video shows each result next to the normal clip, with the sound on, so you can judge the scene, the lips and the voice yourself. With the same seed, mild keeps the scene; some small details can change (a prop, a hand position). Use the normal run when you need the exact result.
How to use the free workflow
Install Stable Yogi Nodes in ComfyUI (from the GitHub page).
Open SY - H3 Video v1 from the pack's workflows/h3 folder. The steps are also in the attached quick-start note.
In group 12 Render, the Speed up node starts on mild. mild is the safest; off gives the exact normal result.
Press Run. Compare with your own prompt and seed before you trust any speed-up.
Honest limits
A small test: three sets on one PC, judged by watching and listening. Your prompt and your card may give other numbers.
More
The model files and their folders are listed in the workflows/h3 folder of the same GitHub page. Questions and results are welcome on our Discord. The full write-up, with our other H3 workflows: read it on Stable Yogi.