Mecha Merge Node Pack

Add only what's actually similar

The more forgiving similarity merge — Add Cosine A, plus magnitude

The merge formula most CivitAI checkpoints are made of

Get the mask itself — 'where do these models disagree?' as a param recipe

Add a delta only where the parents actually disagree

Only add where the models are *truly* opposite — the strict mask

Merge with the model already in your workflow — no reload, no refile

Load *any* safetensors as a merge ingredient — checkpoints, LoRAs, even diffusers-style files

Permute attention layers so two models actually line up before you merge

Per-block merge weights — the SuperMerger MBW dial for ComfyUI

A true/false switch as a recipe — the pack's on/off dials are params too

Force a merge to run in fp64 or on CPU — the precision/VRAM escape hatch

Add a difference but refuse to leave the [A, B] band — the 'safe transplant' merge

Squash any merge into the box you define — bounds as a recipe list, stiffness as a dial

Mixing models by frequency — keep the texture of one, the structure of the other

DARE-style pruning for deltas — the n-supermario merge that beats plain averaging

Swap in the EMA weights before you merge — they're usually the better half

Extract a LoKr (the LoRA that kept its rank) from a merge delta

Turn a merge delta into an actual LoRA file

Extract a LyCORIS-format LoKr, for when the layout matters

Extract a LoRA the LyCORIS way — same SVD, different key names

Use A, and only fall back to default where A is missing

A plain number, wrapped

Weight FLUX's four sub-models separately — diffuser, CLIP-L, T5, and VAE as one param recipe

The merge that doesn't get dragged by outliers — a median instead of a mean

Interpolate in log space — the geometric mean between two models

Ask a recipe where it lives — and hand that answer back to the merge

Read a recipe's precision as a param — wire it into cast-style merges or just debug

Size a low-rank truncation automatically — rank = ratio × min(tensor dims)

Integer params as recipes — how you feed rank, maxiter, and seed into a merge

Merge a LoRA into a checkpoint with ComfyUI, no extra RAM

Translate a merge recipe between key layouts — kohya, diffusers, supermerger, and back

Paste a .mecha recipe string back into the graph — the load half of recipe bookkeeping

Stop re-SVDing the same merge — a persistent cache for the expensive methods

The terminal node — turn a recipe graph into a usable MODEL, CLIP, and VAE

Bundle any number of recipes into one wire — the input every multi-model merge needs

Per-key merge weights by pattern — block weights, but with regex instead of architecture

Freeze your merge in plain text (and share it without the gigabytes)

Turn a pile of models into deltas in one step

Turn any checkpoint into a merge ingredient

The mergekit stock recipe, now streaming on your GPU (or your CPU)

The weird interpolation node nobody tells you about (and why it exists)

The honest average for merging more than two models

Merge everything except the VAE (or whatever you don't want touched)

Quietly delete the NaN weights so your merge doesn't fall apart

Remove the part of B that points along A, and merge what's actually new

Keep only the VAE, the CLIP, or the UNet — merge pieces on purpose

Shuffle your model's neurons on purpose (research-grade merging)

Align models by permuting their weights

Align two models by rotation, not by averaging

Turn strength into a merge parameter

Weigh each part of an SD1 model separately — CLIP, VAE, UNet, EMA, scheduler

Per-part weights for SD1 in kohya's key layout — the LoRA merger's param node

Component weights for kohya-format LoRAs on SD1

Per-component weights for LyCORIS networks on SD1

SuperMerger-style block weights for SD1 — now in ComfyUI

One slider for all of SD1's blocks — the uniform block-weight shortcut

Weigh the five parts of an SD3 merge — two CLIPs, T5, UNet, and VAE

Merge the UNet at full strength and the text encoders at half

The one-slider param node for SDXL UNets that live in a diffusers folder

The kohya node for SDXL, except the UNet keys are diffusers-style

The node for the messiest common setup

Kohya checkpoint, diffusers UNet, LyCORIS LoRA

Per-component merge weights for kohya-style SDXL checkpoints

Per-component weights for kohya checkpoints with kohya-trained LoRAs

Per-component weights for kohya checkpoints with LyCORIS LoRAs

26 sliders for the UNet

The SuperMerger blocks param node that keeps the old muscle memory

Merge anything, keys be damned

The merge that goes around, not through

The plumbing node you'll probably never wire by hand

When a merge recipe needs to be a text value

How you build a delta, the move every good merge starts with

Swap a slice of one model's weights with another's

The full version with geometric median, model stock, and every knob

Merge a pile of LoRAs without them fighting to a standstill

The LoRA-packing node for big style stacks

Keep the important weights, graft in the other model's values

The training trick that says 'change here, not there' — as a merge mask

The mask that keeps only what a fine-tune actually learned

The LoKr path

SVD your deltas

The weighted sum merge, without the 20 GB RAM spike