Already Loaded Model

Every other model-loading node in this pack reads a file from disk. This one grabs the model you've already loaded in your ComfyUI workflow - the MODEL you got out of a checkpoint loader, an Unet Loader, or even a prior merge - and turns it into a MECHA_RECIPE without touching disk again. It's the pack's "use what's already here" node, and it exists to solve a specific annoyance: you've built a complex workflow, a model is already in memory, and you don't want to reload the same file just to feed a merge branch.

How it works

ComfyUI wraps loaded models in a ModelPatcher with the weights sitting in whatever state it's currently in. This node calls state_dict_for_saving to extract the tensors from the live object, then wraps them as a lazy recipe - the clip and vae you optionally wire in get folded into the recipe too, so the downstream merge sees a complete model rather than a headless UNet. Because it's a recipe, the usual rules apply: nothing merges until a Mecha Merger executes the graph, and the execution streams key-by-key.

Inputs:

• model (required) - any MODEL in your workflow.
• clip and vae (optional) - wire them in if you want the merged output to include a text encoder and VAE. Leave empty and the recipe is UNet-only, which matters for save time.

Output: one recipe (MECHA_RECIPE).

When to use it vs Model Mecha Recipe

Model Mecha Recipe loads from file and is the standard starting point - use that for the parents of a merge. Already Loaded Model is for the special cases: merging with a model that only exists in memory (like the output of another merge node), re-merging the result of a Mecha Merger without saving it to disk first, or folding a LoRA-patched model into a merge. The trade-off is real: the loaded model is already materialized in RAM, so this node is less memory-efficient than the streaming file loader - it's a convenience node, not the memory-optimal path. If your goal is a low-RAM merge of two checkpoints, reach for the file loader instead; if you need to merge the thing you're already working with, this is the one.

The merge_checkpointing toggle is notably useful here - the loaded model doesn't change, so caching its recipe extraction on CPU in fp16 saves real time across runs.

Install & gotchas

ComfyUI Manager → search mecha → "Mecha Merge Node Pack", or:

cd ComfyUI/custom_nodes
git clone https://github.com/ljleb/comfy-mecha.git
pip install -r comfy-mecha/requirements.txt

Restart after. Dependency is sd-mecha==1.1.7.

The easy miss: if you skip clip/vae, the recipe is UNet-only, and a downstream save may silently lack a text encoder. And don't treat it as a magic memory fix - the whole point of the pack's streaming is that you don't hold full models in RAM, and this node starts from one that's already there.