Model Merge Weighted 🔀

Checkpoint merges are how half the popular models on CivitAI were born, and until fairly recently you had to leave ComfyUI entirely to do it - boot up a separate merge tool, blend weights, save a new file, come back. Model Merge Weighted 🔀 keeps the whole thing inside the graph: two MODEL inputs in, one merged MODEL out, blended on the fly. No file written, no merge tool, and you can tweak the ratio and re-queue as many times as you like.

It's not going to replace a serious merge suite for fine-grained block-level control - but for the everyday case of "I want a bit more of model B's character in model A," it's exactly the right size.

How it works

It clones model A and applies patches from model B's keys, with the blend determined by the strategy you pick:

• weighted_sum - the classic: out = A * (1 - ratio) + B * ratio. ratio is a single float, default 0.5, from 0 (pure A) to 1 (pure B).
• sigmoid_blend - applies a smooth S-curve to the ratio across the layer stack, so layers near the ratio boundary blend softly instead of cutting over sharply. Good for avoiding the "hard switch" artifacts a straight sum can produce.
• layer_select - encoder-ish layers from A, later layers from B, split at the ratio point. It's the "torso from one, face from the other" move: keep early features from A, let B's later layers drive the output.

The inputs that matter

model_a, model_b, ratio, and strategy. That's the whole surface. Output is a single merged_model, which you wire into your sampler exactly like any other model. Since it's a true graph-time merge, you can also chain it - merge two checkpoints, feed the result into a second merge with a third - which is how people build "50/50 of the 50/50" blends.

The community lesson worth keeping in mind from all those merge-driven checkpoint wars: a merge is only as reproducible as its inputs and ratio. Lock the ratio in the node and note both model names in a comment or metadata, or you'll never be able to rebuild the exact blend that produced a good image.

Installing it

It's in the OmniNodes pack:

cd ComfyUI/custom_nodes/
git clone https://github.com/TensorVizion/OmniNodes

Restart ComfyUI (ComfyUI Manager: search "OmniNodes"). No extra dependencies - it uses ComfyUI's own model-patching internals. Registers under TensorVizion/Model Utilities. If it doesn't appear after install, restart completely and watch the terminal for [OmniNodes] ✅ Loaded lines.

The honest caveat: this merges the whole model uniformly. It doesn't give you per-block multipliers or anything resembling a full merge UI, and on models with very different architectures the result can be mush. For same-family checkpoints at modest ratios it's a genuine time-saver; for serious merge craft, reach for the dedicated tools.