Nodes/ComfyUI-LTXVideo/πŸ…›πŸ…£πŸ…§ LTXV Patcher VAE
ComfyUI Node Runs on cloud

πŸ…›πŸ…£πŸ…§ LTXV Patcher VAE

Wire your VAE through it so LTX decode behaves

By LightricksΒ·Created 2 years agoΒ·Updated about a month agoΒ· 3,956
πŸ…›πŸ…£πŸ…§ LTXV Patcher VAE
  • vae
  • VAE

Some nodes are a wire with a job. LTXVPatcherVAE takes a VAE in and hands a patched VAE back out - no settings, no knobs, nothing to tune. You route your VAE through it before the decode, and it applies the LTX-specific behavior the pack's pipeline expects. If your workflow came with it inserted, leave it there; if you're building from scratch and following an official LTX example, this is one of those "just keep it in the chain" nodes.

Why does the VAE need patching at all? Because LTX's VAE is not a normal image VAE. It's a high-compression spatiotemporal decoder - that combined spatial-and-temporal compression is exactly what makes LTX fast - and it has version-specific quirks. The community has felt those quirks directly: LTX-2.3 rebuilt the VAE and latent space, which even broke the old TinyVAE preview trick because the preview decoder was built for the previous latent layout. A decoder this particular benefits from a wrapper that makes sure it's driven correctly.

How it works

It's a patcher node, the same pattern as other "patch the model/VAE" nodes in ComfyUI: the object passes through and comes out with modified behavior attached. In this case the LTX-specific decode handling is applied to your VAE so downstream decoding lines up with how the pack samples and stores latents. Because it exposes no parameters, there's nothing to configure - the patch is fixed.

The inputs and outputs that matter

This is about as minimal as a node gets:

  • vae (VAE) - the VAE to patch. Typically the LTX VAE you loaded for the workflow.

Output is a single patched VAE, which you feed into your VAE Decode (or wherever the VAE is consumed).

There are no other inputs. Insert it between your VAE loader and your decode, and move on.

How to install it

Grab the pack from ComfyUI Manager the way the README recommends: Ctrl+M β†’ Install Custom Nodes β†’ search LTXVideo β†’ Install β†’ restart. Or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Lightricks/ComfyUI-LTXVideo

then restart ComfyUI; the node is under the LTXV menu. It downloads nothing itself, but it's meaningless outside an LTX pipeline, and that pipeline runs the LTX-2/2.3 model - 32GB+ VRAM and 100GB+ disk per Lightricks, models fetched on first use.

Common issues & troubleshooting

Not sure whether you need it. The safe rule: if the official example workflow you started from includes it, keep it. If you removed it and decoding got weird, put it back. It's a compatibility wrapper, so its absence tends to show up as decode oddities rather than a hard error.

Decode artifacts or wrong colors. First, make sure your VAE matches your checkpoint's version. LTX-2.3 changed the latent space, so a mismatched or old VAE will decode poorly regardless of this node. Patching the correct VAE is fine; patching the wrong VAE won't fix a version mismatch.

Previews stopped working after upgrading to 2.3. That's the known TinyVAE preview breakage from 2.3's new latent space, not this node - you're effectively generating blind on 2.3 until the preview tooling catches up. Decode the full clip to check results.

Nothing seems different with it in vs out. That can be normal - it's a behind-the-scenes patch, not a visible effect. When your pipeline already handles the VAE correctly, its contribution is invisible; its value shows up in the cases where it's preventing a subtle decode mismatch.

Categorylightricks/LTXV

Inputs (1)

NameTypeDefaultDescription
vaeVAEβ€”

Outputs (1)

NameTypeDescription
VAEVAEβ€”