VAE Merge

Checkpoint merging is a whole cottage industry in Stable Diffusion; VAE merging is the neglected cousin. VAEMergeMEC is the node that finally gives it the same treatment: merge two or three VAEs with thirteen blend strategies, per-block weight sliders, a data-driven auto-alpha, a latent-space probe that reports reconstruction quality, and a recipe system so a good merge is reproducible. It's the heavyweight of the ComfyUI-CustomNodePacks pack's VAE Tools corner.

Before you get excited, the honest framing: VAE merging is a narrow tool. The KB's VAE panel notes that modern checkpoints bake their VAE in, and swapping or mixing decoders is mostly a relic of the SD 1.5 era (where the 840k decoder was a real upgrade) and a small set of specialist cases since. Where this genuinely earns its keep: you have two fine-tuned decoders - one renders better skin, one better text - and you want a blend, or you've found a VAE that decodes colors beautifully but smears texture, and want to keep the good part. That's exactly what per-block merging is for.

How it works

vae_a acts as the base (a deep-copied clone is returned, so your originals stay untouched), and vae_b (plus optional vae_c) is blended in. The merge_mode dropdown is where the strategies live - weighted_sum, add_difference, tensor_sum, triple_sum, slerp, sigmoid, geometric, max_abs, smooth_add_diff, distribution_xover, dare_ties (sparse delta + sign election), block_swap, and clamp_interp. add_difference, triple_sum, and smooth_add_diff need the third VAE.

Two features make it more than a glorified lerp:

• use_blocks - per-block sliders (block_conv_in, block_0…block_3, block_mid, etc.) so you can, say, keep encoder blocks from A and decoder blocks from B. The block names correspond to the encoder/decoder structure the pack's Block Inspector reports on.
• auto_alpha - computes per-block cosine similarity between A and B, then biases weights toward A wherever the blocks disagree. It's a sensible default for "I don't know which blocks matter."

You also get brightness/contrast post-shifts on the decoder output (a quick way to fix a merged VAE that decodes slightly dark), and device (cpu default - safe; cuda is faster but eats VRAM).

The QA loop that makes it usable

The genuinely nice part is the reporting:

• dry_run - skip the actual merge and get only the recipe + similarity report. Run this first: it's fast, and it tells you whether your two VAEs are even compatible before you commit.
• reference_image - connect one and the node encodes/decodes it through A, B, and the merged VAE, reporting MSE/PSNR per VAE in probe_report. That's an actual number telling you which merge settings are hurting reconstruction, instead of guessing by eye.
• recipe_json / recipe_in - a merge's full settings serialize to a recipe JSON that you can feed back in to reproduce the exact merge. Screenshot-free reproducibility.

Inputs and outputs that matter

• vae_a / vae_b (+ optional vae_c) - the inputs.
• merge_mode, alpha, beta - the blend. alpha is the primary weight (0 = all A, 1 = all B in weighted_sum).
• use_blocks, auto_alpha, and the block_* sliders - per-block control.
• dry_run, reference_image, recipe_in - the QA loop.

Outputs: vae (the merged result), info, recipe_json, and probe_report.

Installing it

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git

restart ComfyUI or use Manager → "CustomNodePacks". No extra dependencies - pure weight math on already-loaded VAEs.

The trap

A "merge" that preserves the good parts is not automatic - weighted_sum with a naive alpha is the fastest way to average two things into mediocrity. Do the dry run, look at the similarity report, and let auto_alpha pick where B actually contributes. And keep expectations calibrated: this node gives you tools and measurement, not magic. If the base VAE was already bad, merging doesn't fix that.