MixMod Guider

All the MixMod component and options nodes are just plumbing. This is the node that actually turns your stacked components into something a sampler can run: it takes the COMPONENT chain, wraps it in the pack's MixModGuider, and emits a GUIDER in the pack's default "team" mode.

"Team" is the plain-vanilla mixing mode, and it's the best place to start because everything else is an experiment on top of it. At every sampling step the guider runs the denoise prediction through every active model, then blends them: the unconditional predictions are summed weighted by each model's base_weight, and each model's CFG is applied to its own conditional-minus-unconditional difference. That's the classic Composable Diffusion formula - weighted prompts AND'd together across checkpoints - and it's why mixing a detail model with a composition model actually works instead of just averaging them into mud.

The input and output

Just one input: component, the COMPONENT from a MixMod Guider Component or Guider Component Pipeline node (optionally chained through prev_component). One output: GUIDER.

Here's where beginners trip: a GUIDER will not plug into a plain KSampler. KSampler builds its own internal CFG guider and never looks at yours. You need a sampler node with a guider input - SamplerCustomAdvanced is the classic one in core ComfyUI. So the wiring is: MixMod Guider → guider; a RandomNoise node → noise; a KSamplerSelect → sampler; a scheduler (with sigmas) → sigmas; an Empty Latent → latent_image. You've essentially replaced KSampler's model/positive/negative trio with a pre-built guider.

What you can push through the chain

The component chain is where the control lives, and it's worth understanding what the guider reads off it:

• Weights normalize across the chain, so 1.0 + 0.5 means a 66/34 split.
• CFG is per-model - you can run one model at 4 and another at 9 in the same pass.
• Options attached to any component (mask, step range, scale factor) gate whether and where that model is active. The plain Guider honors them in team mode - the Mask option is what triggers masked (region) blending instead of global blending.

The MixModGuider is a subclass of ComfyUI's own CFGGuider, so it slots into anything that accepts a guider and respects standard sampler callbacks. It loads every component's checkpoint and keeps them resident for the whole pass, which is the real cost of this pack: mixing two SDXL checkpoints means holding two SDXL UNets in VRAM. The README's answer is GGUF - the author reports running an SDXL plus an SDXL-GGUF plus Pixart Sigma with a GGUF text encoder on 8GB.

Installing

ComfyUI Manager → search "ComfyUI-MixMod" → install, restart. Or manually:

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

No requirements.txt; the pack runs on plain torch and ComfyUI's samplers. Companion packs are optional and only for specific mixes: ComfyUI-GGUF + ComfyUI_ExtraModels for Pixart Sigma, ControlNet Aux for the Depth Guider.

Troubleshooting

If sampling "succeeds" but the image only reflects one model, one component's normalized weight probably crushed the others - check your chain and weights. If you get an error about a missing guider input, you've wired it into the wrong sampler node. And if the first run is mysteriously slow, that's not a bug: it's two models being denoised every step. That's the point, and the tax.