[Deprecated] Checkpoint Merge
AV_CheckpointMerge — the deprecated model-blend node (use ModelMergeSimple)
- model1
- model2
- MODEL
Straight up: this one's deprecated, and the pack tells you to use ComfyUI's native ModelMergeSimple instead. It blends two loaded models into one by weight - the classic "average two checkpoints and see what you get" operation - but ComfyUI ships that capability natively now, so there's no reason to depend on the Art Venture version for new work.
What it does is worth understanding regardless, because model merging is everywhere on CivitAI. Merging averages the weights of two models to produce a hybrid - in minutes, with no training and no dataset. That's why the bulk of community "models" are merges rather than fine-tunes. The upside is you can dial a blend between, say, two styles. The well-known downside, and it's worth internalizing: averaging two incompatible models gives you something weaker at both, and merge lineages inherit each other's defects (the same faces, the same color biases, the same VAE quirks propagating down generations). A merge is a quick experiment, not a free lunch.
How it works
Despite the "Checkpoint Merge" name, its inputs are MODEL types - the diffusion model (UNet) part, not the whole checkpoint bundle of UNet + CLIP + VAE. It does a weighted average of the two models' tensors according to model1_weight and outputs a single merged MODEL. So this blends the generative half; if you want CLIP and VAE handled too, that's a job for the fuller native merge nodes.
The inputs and outputs that matter
model1/model2- the two models to blend, from Load Checkpoint (or any node that outputs aMODEL).model1_weight- how much ofmodel1goes into the mix, default1, range-1to1, step0.01. Higher leans towardmodel1, lower towardmodel2. The negative range lets you do difference-style (subtractive) merges, which is an advanced move - for a normal blend, stay between0and1and sweep it to find the balance you like.
Output is a single merged MODEL you feed into a KSampler like any other.
How to install it
It's in the Art Venture pack (ComfyUI Manager → search comfyui-art-venture → Install → restart; or clone https://github.com/sipherxyz/comfyui-art-venture into ComfyUI/custom_nodes and restart). But for merging, you don't need it - see below.
Common issues & troubleshooting
Should I use this? No, not for new graphs. Use ComfyUI's native ModelMergeSimple - the pack itself points you there. It's built in, maintained, and does the same weighted blend. Keep AV_CheckpointMerge only to avoid breaking an old workflow that references it.
The merged model looks worse than either parent. That's the merge trap, not a node bug. Averaging two models that pull in different directions (a realism model and an anime model, say) dilutes both. Merges work best between close relatives sharing a base. If the result is muddy, your parents were probably too far apart.
Where's the VAE / CLIP? This node blends MODEL only. If your merge needs the text encoder and VAE combined too, reach for the native merge nodes that handle the full checkpoint, or load CLIP/VAE separately.
It's not merging my .safetensors files directly. Right - it merges loaded models, not files on disk. Put a Load Checkpoint in front of each input. If you want to save the result as a new checkpoint file, add a save-checkpoint node after it.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| model1 | MODEL | — | |
| model2 | MODEL | — | |
| model1_weight | FLOAT | 1.00-1–1 | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |