ModelMergeSimple

ModelMergeSimple is the "one slider" model merger. You feed it two diffusion models, drag a single ratio, and get a blended MODEL out the other end - no training, no dataset, no GPU hours. Most of the "models" people download off CivitAI are exactly this kind of thing done blind in someone else's script; this is the version you can do yourself, live, in the graph, and actually see what you're making before you commit.

That last part is the whole trick of merging in ComfyUI. You don't have to save anything. The merge node hands you a patched MODEL that you can wire straight into a KSampler and sample from, iterate the ratio, sample again. It's an in-graph creative tool as much as an export pipeline. If you've ever wanted to nudge a checkpoint toward another checkpoint's style by 15% and see what happens, this is the cheapest way to do it.

How it works

Merging is just tensor math on weights. Internally the node clones model1, then applies model2's weights as a "patch" on top, so each weight tensor becomes:

result = ratio · model1 + (1 − ratio) · model2

At ratio = 1.0 you get pure model1; drag it to 0 and you get pure model2; 0.5 is a genuine 50/50 average of every weight in both models. Because it happens as patches rather than a saved file, the blend is applied on the fly when the model runs - which is why it's instant and free.

A word of warning before you get excited, and it's the same warning the whole checkpoint world trips on: merging averages everything, including both models' defects. The checkpoint KB in this area has the receipts - "99% of everything posted, since the beginning, are merges. Homogenization is inevitable" is a direct quote from the community. Blending a realism model with an anime model gives you a model weaker at both. Same-architecture, same-vibe parents merge better than opposites. DreamShaper v3.3, one of the most successful merges ever, blended eight SD 1.5 models of roughly one aesthetic family - that's the winning recipe.

The inputs that matter

• model1 and model2 (MODEL) - the two checkpoints. Wire them straight out of Load Checkpoint nodes.
• ratio (FLOAT, default 1.0, 0–1, step 0.01) - the one knob. It's the fraction of model1 that survives.

Output is a single MODEL - feed it to a KSampler, a second merge, or a Save Checkpoint node.

Both models need the same architecture, and ideally the same-ish latent space. Merging an SD 1.5 checkpoint into an SDXL one produces a shape-mismatch warning and a model that isn't a model.

Where people get burned

The naive multi-LoRA trick is the classic footgun. Because the node adds full weights, chaining several LoRA-injected models together with pure ModelMergeSimple stacking adds rather than averages, and tensor peaks blow up. The community worked around it by merging LoRA-loaded models at ratios that renormalize back to 1.0 - a real, posted workflow for Z-Image. It works, but the moment you stack more than a couple of adapters you're better off with a proper difference merge (see ModelMergeSubtract) or a dedicated LoRA stacker.

The other gotcha: this node is a blend, not a checkpoint. If you want the result as a file, you have to explicitly save it - the graph doesn't do it for you.

How you get it

It ships with ComfyUI core under model/merging - no install, no Manager, nothing to download. It's been there since the early days of the project.

The siblings in this family give you the same idea with more control: ModelMergeBlocks splits the blend per architectural block, and the model-specific merges (ModelMergeSDXL, ModelMergeSD3_2B, and friends) pre-name every block so you don't have to. Start with Simple to learn the feel, then graduate.