Nodes/braintacles-nodes/VAE Encode Tiled Pipe
ComfyUI Node

VAE Encode Tiled Pipe

Tiled VAE encode for images too big for your VRAM — VAE included

By braintacles·Created 3 years ago·Updated 2 years ago· 2
VAE Encode Tiled Pipe
  • pixels
  • vae
  • LATENT
  • VAE
tile_size1024

Core VAE Encode on a huge image is where cards start swapping to system RAM and throughput collapses. The tiled variant splits the encode into chunks so you can push much bigger pixels on far less VRAM - and VAE Encode Tiled Pipe adds the pipe pattern on top: VAE in, LATENT and VAE out, so the encode, sampler and decode can all hang off one VAE supply line.

How it works

It crops the image to a multiple of 8 first (centered, so no lopsided edges), then calls ComfyUI's vae.encode_tiled with tile dimensions of tile_size ÷ 8 - because latent space is 8× smaller than pixel space, a 1024 tile means 128×128 tiles of latent. Inputs: pixels (IMAGE), vae (VAE), tile_size (default 1024, range 320–4096 in steps of 64). Outputs: LATENT, VAE.

tile_size is the one input to care about. The 1024 default is a sensible starting point; drop it toward 320–512 for very large images or genuinely small cards, and keep it as high as your VRAM allows for cleaner results.

The tradeoff is real

Tiled VAE is dramatically slower than a single-pass encode - it's doing the same work in more, smaller pieces - but the memory saving is the point. The community numbers on the tiled decode side are the ones to remember: tiling used less than a third of the VRAM of the full pass on the same hardware. If your 2K+ inputs are OOM-ing, this is the fix; if your inputs fit comfortably, tiling just costs you time for nothing.

The honest problem

This node left the pack. It shipped in the early braintacles releases and was removed in May 2024 when the pack shrank to five nodes; the current GitHub and ComfyUI Manager version doesn't include it. If a workflow references it, that's why ComfyUI reports it missing.

Recreate it today with the core VAE Encode (tiled) node plus a second VAE wire to your decode - same behavior, one extra line. Or pin the pack to the last release that had it:

git clone https://github.com/braintacles/braintacles-comfyui-nodes
cd braintacles-comfyui-nodes
git checkout 26732fd

For a missing-node fix, just use the core tiled encode. The pipe is a nicety, not a capability.

Categorybraintacles/latent

Inputs (3)

NameTypeDefaultDescription
pixelsIMAGE
vaeVAE
tile_sizeINT1024320–4096

Outputs (2)

NameTypeDescription
LATENTLATENT
VAEVAE