sd-mecha for Comfyui

comfy-mecha is a complete model merging node pack for ComfyUI with a focus on low memory footprint.

compose complex recipes without needing to save dozens of intermediate merges to disk
merge loras to models
support for block weights
and a bunch of other stuff. For more info, see the nodes listing below. See also the readme of the underlying library sd-mecha

Workflows

Basic weighted sum

resources/weighted_sum.png

Clipped add difference

resources/clipped_add_difference.png

Ties merging

resources/ties_merging.png

Recipe workflows can get much, much more complex than this.
If you are familiar with writing python code, you might be interested in using the sd-mecha library directly for experiments as an alternative to ComfyUI: https://github.com/ljleb/sd-mecha

Install

Install with ComfyUI-Manager

Assuming you have ComfyUI-Manager installed:

Open a browser tab on ComfyUI
Click on the "Manager" button
Click on "Install Custom Nodes"
Search for "mecha"
Install "Mecha Merge Node Pack"

Install manually

You can also install the node pack manually:

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

Nodes listing

Merge nodes

Nodes used for merging. They all have Recipe in their name except for Mecha Merger.

nodes ending in ... Mecha Recipe return a merge recipe
Mecha Merger takes a MECHA_RECIPE as input, and returns a unet and a text encoder
Serializer takes a MECHA_RECIPE as input, and returns the recipe instructions using the mecha format
Deserializer takes a mecha recipe string as input, and returns the deserialized MECHA_RECIPE (this is the inverse operation of Serializer)
Mecha Model Recipe loads a model as a recipe to be used as input to other recipe nodes. model architecture auto-detection is not yet supported, so please make sure the right model_arch parameter is selected!
Mecha Lora Recipe loads a lora model as a recipe to be used as input to other recipe nodes. again, model architecture auto-detection is not yet supported, so make sure to use the right model_arch
Mecha Recipe List takes an arbitrary number of recipes and returns a MECHA_RECIPE_LIST. It is intended to be used as input to recipe nodes that accept an arbitrary number of recipes as input, i.e. the bounds input of Clip Mecha Recipe

Hyper nodes

Nodes used to specify hyper(parameters) to merge methods. For example, Weighted Sum Mecha Recipe has a hyper input alpha with a default value of 0.5.

Blocks Mecha Hyper can specify a different hyper for each block of the models to be merged (AKA "merge block weighted" or "MBW")
Float Mecha Hyper specifies the same float for all blocks of the models to be merged

Extensions

To add custom merge nodes, you can add python scripts that make use of the mecha extension API under the mecha_extensions directory. The node pack will run all scripts placed there before creating the comfy nodes.

Currently, the documentation for the mecha extension API is under construction. For now, to get more information, you can either take a look at the custom merge method example, open a discussion post to ask questions, or join the discord server.

ljleb / comfy-mecha

readme