Kijai

/

MiniMax-H3-experimental

Like 448

Thank you, it’s a great model: I’m using it with the latest COMFI.
The only thing that’s hard is choosing the sigmas — here are my sigmas for (personally for me):

4 step: 0.9999166, 0.9000, 0.7000, 0.4500, 0.0000

6 steps: 0.9999166, 0.9729730, 0.9000000, 0.7397260, 0.4500000, 0.1558442, 0.0000000

Step 8: 0.9999166, 0.9552000, 0.9000000, 0.8303000, 0.7397260, 0.6183000, 0.4500000, 0.2160000, 0.0000000

Step 10: 0.9999166, 0.9650000, 0.9236000, 0.8741000, 0.8141000, 0.7397260, 0.6458000, 0.5243000, 0.3649000, 0.1622000, 0.0000000

Step 15: 0.9999166, 0.9772400, 0.9519100, 0.9235600, 0.8916500, 0.8555500, 0.8140700, 0.7663900, 0.7109100, 0.6457400, 0.5684200, 0.4758900, 0.3648600, 0.2336400, 0.0910600, 0.0000000

20 steps: 0.9999166, 0.9831600, 0.9650000, 0.9451000, 0.9236000, 0.9000000, 0.8741000, 0.8457000, 0.8141000, 0.7790000, 0.7397260, 0.6956000, 0.6458000, 0.5891000, 0.5243000, 0.4500000, 0.3649000, 0.2588000, 0.1622000, 0.0582000, 0.0000000

25 steps: 0.9999166, 0.9862200, 0.9729700, 0.9602000, 0.9478400, 0.9358000, 0.9240000, 0.9123000, 0.9006000, 0.8888000, 0.8768000, 0.8645000, 0.8517000, 0.8383000, 0.8241000, 0.8089000, 0.7925000, 0.7745000, 0.7545000, 0.7320000, 0.7063000, 0.6765000, 0.6414000, 0.5992000, 0.5467000, 0.0000000

I calculated the sigmas using the formula:

Formula for Generating Sigma Schedules

σ(x) = 1 / (1 + 0.324074 * x/(1-x) + 0.027778 * (x/(1-x))²)

where
x = i / (N-1), i = 0, 1, …, N-1,
N = number of sigma points (one more than the number of denoising steps).

What It’s About

This is a universal formula for creating a smooth sequence of noise levels (sigmas) during the reverse diffusion process.
It provides:

• a slow decrease at the beginning (first steps),
• acceleration in the middle,
• and a micro‑step at the very end (the last values are very close to each other) – which is crucial for clear audio without metallic artefacts.

The formula was empirically tuned for the MiniMax H3 FastVideo model and works excellently for 6, 8, 10, 15, 20, and 25 steps.
For 4 steps, a manually tested set is recommended:
0.9999166, 0.9000, 0.7000, 0.4500, 0.0000

Just plug in your desired number of steps, get a ready‑to‑use sigma list, paste it into ManualSigmas – and enjoy clean sound.