ModelMergeRatios

Model merging is how a huge share of CivitAI checkpoints are born - combine two models' weights and you get a blend that inherits a bit of both. The fiddly part is dialing in how much of each. ModelMergeRatios is a control panel for exactly that: three labeled sliders - input, middle, out - and three float outputs, ready to feed a model-merge node's ratio inputs. It doesn't merge anything itself; it's the settings node that keeps your ratios visible and adjustable instead of buried in a merge node's widget row.

It's from artyclaw/artyclaw-comfy, the same author who ships ClipMergerSettings - this is the checkpoint-side sibling of that pattern.

How it works

Three inputs, three outputs, straight through. Each value is clamped to 0–1 and rounded to two decimals - so 0.85 stays 0.85, and a stray 1.3 from an upstream wire gets pinned to 1.00. The input names follow the classic SD 1.5 merge block split: input blocks (early layers, structure and composition), middle blocks (the model's mid-layer personality), and out blocks (late layers, fine detail and color). Wire the three outputs into the corresponding ratio inputs of your merge node, and you get a clean visual panel for the three-way blend.

There's nothing more to the mechanism - and that's the design. It's a pass-through with safety clamping, the same honest pattern as the pack's other settings node.

Inputs and outputs

• input - float 0–1, default 0.5.
• middle - float 0–1, default 0.25.
• out - float 0–1, default 0.25.

Three FLOAT outputs in the same order (input, middle, out), rounded to two decimals and clamped. The defaults sum to 1.0, which is the sensible starting point for a weighted merge - adjust from there.

Install

ComfyUI Manager → search ArtyClaw Comfy Nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/artyclaw/artyclaw-comfy

Restart. Standard library only (math, re).

Where people get burned

The classic mistake is treating this as the merge itself. Wire the outputs into an actual model-merge node or you've got three floats doing nothing. Second, the round-to-two-decimals means a slider at 0.005 renders as 0.01 - fine for merging, annoying if you expected precision you were never going to use anyway. Third, remember the naming: input/middle/out is the SD-family block split, not "model A, model B, model C" - if you're blending two whole models with a simple merge, check which ratio set your merge node actually wants before assuming order. And like all pass-through nodes, a stale-looking output usually means graph caching, not a bug - force a re-run. For what it is - a tidy, labeled ratio panel for merges - it's done well.