MixMod Options Scale

MixMod's Scale option is a way to stop one of your blended models from dictating fine detail. You attach it to a component and set a scale_factor, and that component's model runs at a lower resolution for its entire denoise pass - downscaling the latent, running the UNet, then upscaling the prediction back before it joins the mix.

Why would you want that? Because low frequencies are structure and high frequencies are texture. A model forced down to half resolution contributes composition, lighting, and layout - the broad strokes - while staying out of the pixel-level detail fight. That's the same reasoning behind the pack's FFT modes (different models owning different frequency bands), just implemented crudely and cheaply with a resize instead of a Fourier transform. The README's tip about "mixing a detail-focused model with a composition-focused one" is this exact idea: let the composition model be the low-res one.

Inputs and output

• scale_factor (default 0.5, 0–1) - the fraction of the original resolution the model runs at. 0.5 means half size; lower is more aggressive. 0 effectively disables the wrapper for that component.
• prev_options (optional) - chain options to layer this onto a mask or step range.
• Output: OPTIONS, into the options pin of a MixMod Guider Component or Guider Component Pipeline node.

Under the hood it's a model_function_wrapper that bislerp-downscales the input latent, runs the model, and bislerp-upscales the output back - a cheaper cousin of how ComfyUI's own low-res sampling tricks work. Because the component's prediction is resized back to full res before blending, the final mix still comes out at full resolution; you're only degrading that one model's opinion, not the output.

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 models to download - plain torch plus ComfyUI's samplers. The optional companion packs (GGUF, ControlNet Aux) are only for specific mixes, not for this node.

Troubleshooting

If the scaled component seems to vanish entirely, your scale_factor is probably too low - below ~0.25 the model is working on a latent so small it's mostly guessing, and its contribution turns to mush. A soft failure mode, not an error. Also expect the scaled model to be faster than its siblings (it is - it's working on fewer pixels), so don't read speed differences as a bug. And remember this is a per-component dial: attach it to the model you want to be the broad-strokes artist, not the detail one.