Merge Models

Take two loaded models, set a blend ratio, get one merged model out. That's the node. It's the on-the-fly version of what people do offline in tools like SuperMerger - average two checkpoints' weights to make a hybrid - except it happens live in your graph, so you can wire the ratio to a slider and watch the blend change without ever writing a merged file to disk.

The appeal of merging in-graph is experimentation. Want to see what 30% of a lighting-focused model does to your main checkpoint? Set the ratio, run, look. No 6GB file to save, no reload. The trade-off is that you're re-merging every run, which costs a little time and memory - but for dialing in a blend, that's exactly the loop you want.

A word of grounding on merging itself, because it's easy to over-expect: averaging two models does not cleanly combine their strengths. If the parents pull in different directions, the child is weaker at both, and merged lineages inherit each other's biases and defects. Merge close relatives (same base architecture, similar training) and it works; merge a realism model into an anime model and you mostly get mud. Treat the ratio as a taste dial, not a magic mixer.

The inputs and outputs that matter

• model1 / model2 - the two MODEL inputs to blend, straight from Load Checkpoint or any node that outputs a MODEL.
• ratio - the blend weight, default 1, range 0–1, step 0.01. It interpolates between the two models. There's no universal convention for which end is which, so the reliable approach is to sweep it - try 0.25, 0.5, 0.75 - and compare outputs to find the balance you're after. Small ratio changes can shift the result more than you'd expect near the extremes.

Output is a single merged MODEL, ready for a KSampler.

MergeModels vs AV_CheckpointMerge

Same neighborhood, slightly different framing. AV_CheckpointMerge is marked deprecated (it steers you to ComfyUI's native ModelMergeSimple) and uses a -1–1 weight that allows subtractive merges. MergeModels uses a clean 0–1 ratio for a straight blend. For a normal two-model average, either works; if you're picking fresh, ComfyUI's built-in ModelMergeSimple is the most future-proof and does the same job.

How it works

It computes a weighted average of the two models' tensors by ratio and returns the combined MODEL. Note it operates on the MODEL (the diffusion/UNet part), not the full checkpoint bundle - CLIP and VAE aren't touched here. If your merge needs those combined too, handle them with the appropriate native merge nodes.

How to install it

Part of the Art Venture pack. ComfyUI Manager → search comfyui-art-venture → Install → restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/sipherxyz/comfyui-art-venture

then pip install -r comfyui-art-venture/requirements.txt and restart. No model download - you supply the two models.

Common issues & troubleshooting

The blend went the "wrong" way. ratio interpolates between the two, but which end maps to which model isn't intuitive - just sweep it and read the results. Two quick test runs settle it.

Muddy or degraded output. Classic merge dilution. The two models are probably too dissimilar. Merge within a family (both SDXL realism models, say), not across architectures or wildly different styles.

I want to keep the merged model. This produces a live in-memory MODEL, not a file. Add a save-checkpoint node after it if you want a reusable .safetensors.

Node missing after importing a workflow. The pack isn't installed - Manager → Install Missing Custom Nodes.