🅛🅣🅧 Set VAE Decoder Noise

This is a small, easy-to-miss node that does something subtle but real: it tells the LTX VAE to add a controlled dose of noise while it decodes your latent back into pixels. The payoff is texture. LTX output can lean smooth and slightly plasticky, and a touch of decoder noise breaks that up - you get finer grain and more believable detail instead of the airbrushed look.

How it works

The LTX VAE was trained with noise augmentation baked into its decoder, which means it knows how to turn a noised latent into a plausible, detailed frame. This node configures exactly how much of that noise gets injected at decode time. You wire your VAE through it, and it hands back a patched VAE that carries these settings - so every decode downstream applies them. It doesn't change how the video is generated, only how the final latent is rendered to pixels. That makes it a cheap, purely-cosmetic pass you can toggle without re-running the sampler.

The inputs that matter

Three numbers control the effect, and the defaults are deliberately gentle:

• timestep (default 0.05) - how far up the noise schedule the decoder pretends the latent is. Small values mean a light touch; push it and the detail-synthesis gets more aggressive.
• scale (default 0.025) - how much noise actually gets mixed in. This is your main dial for "how much grain."
• seed (default 42) - makes the noise reproducible, so a result you liked is repeatable.

The single output is the wrapped VAE. Plug it into your VAE Decode in place of the raw one.

Installing it

• ComfyUI Manager - search LTXVideo, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo, then restart.

No extra model download of its own - it just patches whichever LTX VAE you've already loaded.

Common issues

Output looks noisy or grainy in a bad way. You've overdone it. Both timestep and scale default low for a reason; bring them back down. This is a seasoning, not a main ingredient.

You see no difference at all. At the defaults the effect is intentionally subtle, especially on already-detailed high-res output. If you're hunting for it, raise scale a little and compare the same seed side by side - the difference shows up most on smooth surfaces like skin and sky.

It's fighting your upscaler. If you run a two-stage pipeline with a latent upscaler, decide where the detail should come from. Stacking heavy decoder noise under an aggressive generative upscale can double up the texture and look busy. Pick one to carry the grain.