VAE Merge Subtract

VAE Merge Subtract does one thing: takes every weight in vae1 and subtracts multiplier × the matching weight in vae2. It's the subtraction half of the delta-merge pattern, and it's the node you reach for when you want to remove something - not blend two VAEs, but peel one apart.

What it's actually for

The one pattern this node exists to serve is isolating a "delta." If you have a base VAE and a fine-tuned variant, subtracting the base from the variant gives you just the change - the fingerprint of that fine-tune. Once you have the delta, you can add it onto a different base (the pack's VAE Merge Add is the natural partner), or scale it up or down with the multiplier to control how strongly the change lands.

A concrete framing: the Qwen-Image VAE is the Wan 2.1 architecture with the encoder frozen and the decoder retuned for text legibility, which is what causes the "too smooth, weak contrast" complaints. If you had a decoder you thought carried the right look, subtracting the Qwen flavor from a Wan-based decode is the same mental model - identify the contribution, remove it, keep the rest.

Inputs and output

• vae1 - the VAE you're subtracting from.
• vae2 - the VAE being removed.
• multiplier - how much of vae2 to remove, from −10 to 10, default 1.0. A negative multiplier flips this into addition, which is handy when you want to add a contribution without switching nodes.

Keys that exist only in vae1 pass through untouched; keys only in vae2 are dropped entirely (there's nothing in vae1 to subtract from). One VAE comes out, fully materialized in memory - wire it into VAEDecode, or save it with a VAE-save node if you want to keep it. Nothing is written to disk by this node itself.

Where it bites

All the usual VAE-merge rules apply, and they're sharper here because subtraction is less forgiving than blending. Only subtract VAEs of the same architecture and channel count - SDXL is 4-channel, Flux/Wan/Qwen are 16-channel, and crossing them gives noise, not nuance. And keep multiplier modest: you're operating directly on weights, and pushing them too far can drive activations out of range, which is how you end up with black frames or NaN on decode (the SDXL VAE's fp16 NaN history is the classic example of "weights got too big").

This is a surgical tool for people who already know they want to remove a VAE's contribution. If you just want to nudge one VAE toward another, VAE Merge Simple is the friendlier dial. Subtract is for when you know exactly what you're carving out.