MiniMax H3 Singularity: split-sigma sampling in ComfyUI

This tutorial compares two sampling workflows for MiniMax H3 Singularity 1.3 in ComfyUI. The presenter keeps the model, prompts and character references consistent while testing conventional two-pass sampling with latent upscale against a split-sigma approach. The model version stays unchanged; the focus is sampling structure and LoRA placement for fast action.

The split workflow starts with two sampling steps that produce a noisy, low-resolution latent, then passes it into refinement at a higher resolution. The demonstration uses a 0.5-megapixel first pass and a 1.5-megapixel second pass. The presenter describes a 14-step split-sigma schedule and combines Turbo, Combat V2, Motion Repair and LMS LoRAs across the workflow.

In the examples, the presenter reports sharper hair, clearer water details and more coherent movement than with the older method, including a fight scene with two reference images. These are observations from the demonstrated outputs, rather than a guarantee for every prompt. The higher-resolution refinement also takes longer to generate.

For plasticky textures, the presenter suggests adding a realism LoRA or lowering the second-pass LMS strength, with 0.4 and 0.7 as starting values. The tutorial gives local AI video users a sampling comparison built around native ComfyUI nodes, with optional low-VRAM patches.