VAELoaderDecode

VAELoaderDecode fuses two stock nodes into one: it loads a VAE from a dropdown and decodes a latent with it, returning the IMAGE. Inputs are samples (LATENT) and vae_name (the dropdown of your models/vae folder); output is one IMAGE. That's the entire interface, and it's all it does.

Honestly, it's a convenience node with a narrow audience. The stock ComfyUI VAELoader + VAEDecode pair is two nodes and two clicks, so you're saving maybe one connection. Where it earns its keep is graphs that want to swap VAEs fast - testing a different VAE against the same latent for that color/contrast shift people chase - because there's exactly one widget to change instead of two nodes to rewire. It's also the decode half of this pack's SimpleSamplerVAE, which is the part that actually pulls its weight.

How it works

The mechanism is textbook: it resolves vae_name against ComfyUI's vae folder paths, constructs a VAE from that file (comfy.sd.VAE(ckpt_path=...)), then calls vae.decode(samples["samples"]). No scaling tricks, no latent pre-processing - it expects a latent that came out of a sampler using the same VAE's latent space. That last point is the real trap: decode a latent with a mismatched VAE and you get noise or garbage colors, which is true of the stock decoder too.

Install

Clone and restart; ComfyUI Manager finds it by searching "Lilly":

cd ComfyUI/custom_nodes
git clone https://github.com/lilly1987/ComfyUI_node_Lilly

Restart after cloning. No model files to download; the pack auto-installs rich and chardet on first load. It also has a habit of logging the full VAE list to the console on every decode - harmless, mildly annoying.

The verdict: if you're already running this pack for its wildcards, use this node to cut one wire off your decode chain. If you're starting fresh, the stock nodes do the same job, and this pack is unmaintained - the author's own README now points to Impact Pack. Nothing here will surprise you either way; it's about as simple as a ComfyUI node gets.