MixMod Options Scheduler

Diffusion sampling isn't one operation - the early steps decide the overall structure and composition, the late steps paint in the details. MixMod's Scheduler option is how you exploit that split: you attach a start_step/end_step window to a component, and that component's model only contributes to the denoise predictions inside that window.

The obvious use is the two-phase mix: model A (say, a base model that's great at composition) handles steps 0–60%, then hands off to model B (a finetune that's great at texture) for the rest. The pack's README literally pitches "schedule models to activate at different sampling steps for creative control," and it's one of the less exotic ideas in this pack - it's the model-mixing version of a two-pass sampler, done inside one denoise loop instead of across two KSamplers.

Inputs and output

• start_step (default 0, 0–100) - the first step this component's model participates in.
• end_step (default -1, -1–100) - the last step it participates in. The -1 default is special: it means "until the end." Set it to a real number to cut the model off.
• prev_options (optional) - chain other options to layer a mask or scale factor onto the same component.
• Output: OPTIONS, into the options pin of a MixMod Guider Component or Guider Component Pipeline node.

The window is checked per step against the component's position in the chain: inside the window the model predicts and its weights/CFG count; outside it, that model sits out entirely. If no model is active at a step, the guider falls back to the first component's prediction and logs a warning - so make sure your windows overlap or cover the whole range.

Installing

Part of the ComfyUI-MixMod pack. ComfyUI Manager → search "ComfyUI-MixMod" → install and restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/kantsche/ComfyUI-MixMod

No requirements.txt, no model downloads. Companion packs (GGUF + ExtraModels for Pixart Sigma, ControlNet Aux for depth) are only needed for those specific mixes.

Troubleshooting

The classic screw-up is leaving a gap: model A on 0–5 and model B on 6–20 with step 5 covered by nobody - you'll see the fallback warning in the console and a step where only the default model drove. Overlap is safer than precision here. Second gotcha: the step numbers are denoise steps, not percentages, so they depend on your total step count - a 0–15 window is half of a 30-step sample but barely a slice of a 50-step one. And if a model you scheduled out seems to still affect things, check you attached the Scheduler to that component's options pin and not a neighboring one; cross-wired options is the easiest thing to miss in a big mix graph.