Two free ComfyUI workflows for MiniMax H3. Each makes a 5 second clip with its own sound from one prompt, in 4 steps instead of 20. The 2-step file is about twice as fast again and is good for trying prompts.
We tested six speed files on the same four prompts, with the same seed, and compared every clip with the slow 20-step version. The video shows which one stayed closest and how long each one takes.
Get the setup notes, the model links and more free workflows.
How to use them
- Open the 2-step workflow, type your prompt and run it.
- Change the prompt until the clip does what you want.
- Open the 4-step workflow, use the same prompt and seed, and render the clip you keep.
Settings that belong to these files
- 4 steps (or 2 with the 2-step file), euler, simple
- Shift 12 for video, 3 for sound
- Basic Guider, no CFG
- 1344 x 768, 124 frames for 5 seconds
Core ComfyUI nodes only. The download includes QUICKSTART.md with every file and folder, and the four test prompts from the video.
PDMD speed LoRAs by pdmd2026 (Apache-2.0), ComfyUI copies by Iwannapose. MiniMax H3 model by MiniMax. ComfyUI packaging by Comfy-Org. The people in these clips were made by AI.
Description
FAQ
Comments (2)
Good stuff!
Ran both 2 and 4 step. Both decent image quality, 4 step better, as expected. But 2 step was total audio gibberish, and with 4 step she spoke something unintelligible before picking up and drinking from the cup.
UPDATE: after a few gens, even the 4 step puts out audio gibberish throughout.
Thanks for testing both and for the update, that’s really useful. When I tested the speed versions I was mostly checking the picture, so the speech side is something I haven’t checked properly yet. The 2 step one is really meant for quick drafts. For clips where she has to talk, the full step version is the safer pick for now. One thing that would help me: if you restart ComfyUI and run the same prompt at 4 steps, is the first clip clean again? That tells me if something builds up between runs or if it’s just the steps.