Nodes/Comfyroll Studio/⛏️ CR Apply Model Merge
ComfyUI Node Runs on cloud

⛏️ CR Apply Model Merge

Blend a whole stack of checkpoints at once

By Suzie1Β·Created 3 years agoΒ·Updated 2 years agoΒ· 1,296
⛏️ CR Apply Model Merge
  • model_stack
  • MODEL
  • CLIP
  • model_mix_info
  • show_help
β—„merge_methodβ–Ύβ–Ί
β—„normalise_ratiosβ–Ύβ–Ί
β—„weight_factor1.00β–Ί

Model merging is the old community sport of averaging two or more checkpoints together to get a new one that borrows a bit from each - a realism model's skin plus an anime model's linework, that kind of thing. ComfyUI's native merge nodes handle two models at a time and chain awkwardly for more. This node takes a whole MODEL_STACK and merges the lot in one go.

It's the "apply" end of Comfyroll Studio's model-merge system (Suzie1 / RockOfFire), paired with CR Model Stack, which is where you list the checkpoints and their ratios. Merging in-graph like this is one of those things ComfyUI can do that a form-based UI simply can't - you can merge, generate, and compare without ever writing a checkpoint to disk.

How it works

You build a list of models with weights in a CR Model Stack node, then feed that MODEL_STACK here. The node walks the stack and blends the weights together according to your chosen method, producing one merged model (and a matching CLIP) you can sample with immediately. The result lives in memory - it's a temporary merge for this run, not a saved .safetensors, unless you separately add a save node.

The inputs that matter

  • model_stack - the list of checkpoints and their ratios, from CR Model Stack. This is the substance of the merge.
  • merge_method - Recursive or Weighted. Recursive blends them in sequence (each merge folded into the next); Weighted averages them by their assigned weights. Weighted is the more intuitive "50/50, 70/30" mental model; Recursive changes how influence compounds across more than two models.
  • normalise_ratios (Yes / No) - rescale the stack's weights so they sum sensibly. Leave it on Yes unless you know you want raw, unnormalized weights; off, it's easy to blow past 1.0 total and get a muddy result.
  • weight_factor (0–1) - a global scaling dial over the blend.

Outputs: MODEL, CLIP, a model_mix_info string that reports what got merged (handy to route into a text/preview node so you remember the recipe), and the show_help link.

How to install it

  • ComfyUI Manager: search Comfyroll Studio, install, restart.
  • Manual: cd ComfyUI/custom_nodes && git clone https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes.git, then restart.

Bring your own checkpoints in ComfyUI/models/checkpoints.

Common issues

Rule zero of merging: the models have to be the same architecture. You cannot merge an SD 1.5 checkpoint with an SDXL one - different shapes, it'll error or produce garbage. Keep the whole stack in one family.

If your merge comes out washed-out, low-contrast, or muddy, that's usually unnormalized weights summing past 1.0 - flip normalise_ratios to Yes and pull the individual ratios down. Merging is also memory-hungry: holding several full checkpoints to blend them can spike VRAM/RAM, so a stack of large models on a small card may OOM. And keep that model_mix_info output wired to a preview somewhere - merges are easy to stumble into a good one and then have no idea what you did. If the node is missing entirely, the Comfyroll pack failed to load, which is a pack-level problem rather than this node.

Category🧩 Comfyroll Studio/✨ Essential/⛏️ Model Merge

Inputs (4)

NameTypeDefaultDescription
model_stackMODEL_STACKβ€”
merge_methodCOMBO2 options: Recursive, Weighted
normalise_ratiosCOMBO2 options: Yes, No
weight_factorFLOAT1.000–1β€”

Outputs (4)

NameTypeDescription
MODELMODELβ€”
CLIPCLIPβ€”
model_mix_infoSTRINGβ€”
show_helpSTRINGβ€”