Nodes/ComfyUI-GlifNodes/Patch Consistency VAE Decoder
ComfyUI Node

Patch Consistency VAE Decoder

Swap your SD 1.5 VAE decoder for DALL·E 3's, in tiles so it fits

By glifxyz·Created 3 years ago·Updated about a year ago· 65
Patch Consistency VAE Decoder
  • vae
  • VAE

This is the sibling of the Consistency VAE Decoder, and it's the one you actually want if you're serious about it. Instead of decoding latents in a standalone node, Patch Consistency VAE Decoder pokes a hole in your existing VAE and drops OpenAI's consistency decoder in as its decoder - then adds tiled decoding so large latents don't blow up your VRAM. You keep all your normal VAE wiring, and the swap is invisible downstream.

How the patch works

You feed it a vae - the same one you'd normally wire into VAE Decode - and it returns a patched vae. Under the hood it deletes the stock first_stage_model.decoder, reassigns the VAE's decode method to the consistency decoder's feedforward pass, and wraps vae.decode in a tiled decode (decode_tiled_ with 512px tiles and 64px overlap) that reassembles the image and moves channels around. The result behaves like a normal VAE, so your graph doesn't change at all - the VAE output feeds into whatever it always fed into.

That tiled part is the real feature. The consistency decoder is a big, memory-hungry feedforward network, and standalone it OOMs on anything beyond moderate resolutions. Tiling is exactly how this ecosystem handles "big image, small GPU" - split, decode, overlap, blend - which is the same trick tiled diffusion upscalers use. On a 6-8GB card, that's the difference between a working decode and an out-of-memory crash.

The honest caveats

Everything that's true of the standalone consistency decoder is true here: it only understands SD 1.x/2.x latent space, so patching an SDXL or Flux VAE gives you garbage, and the community verdict on quality was genuinely mixed (better text and lines for some, worse faces for others, and a decode that's slower than the stock VAE). This is a 2023 novelty with a fan, not a universal upgrade. If your image comes out as noise or flat color, the first thing to check is which model's latent you fed it - that's the failure mode, not a bug in the node.

Install

Ships in ComfyUI-GlifNodes from the glif.app team. ComfyUI Manager → "ComfyUI-GlifNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/glifxyz/ComfyUI-GlifNodes
pip install -r requirements.txt   # diffusers>=0.28.0

Restart ComfyUI. The openai/consistency-decoder model (~2.5GB, fp16) downloads on first use into ComfyUI/models/huggingface_cache/ - not at install, so the clone itself is light.

Reach for this one when: you're on SD 1.5, you've tested the consistency decoder standalone and liked it, and your resolution outgrew what the non-tiled node could decode. Otherwise the stock VAE is still the everyday choice.

Categorysd

Inputs (1)

NameTypeDefaultDescription
vaeVAE

Outputs (1)

NameTypeDescription
VAEVAE