🅛🅣🅧 STG Guider Node

STG is one of those quietly load-bearing tricks in LTX workflows. Short for Spatiotemporal Skip Guidance (the paper calls it Spatiotemporal Guidance, arXiv 2411.18664), it's a training-free way to make transformer-based video models produce cleaner, more coherent motion without hurting variety. STGGuiderNode packages it as a GUIDER you drop into a custom-sampler graph. If your LTX output looks a little mushy or the motion falls apart, this is one of the first knobs people reach for.

The idea is a cousin of classifier-free guidance, but instead of steering with a negative prompt, it steers away from a deliberately degraded version of the model. STG runs a second forward pass with certain transformer layers skipped - that produces a worse, blurrier prediction - and then pushes the real prediction away from it. The model ends up sharpening the parts a weakened version of itself would have gotten wrong. Net effect: better perceptual quality and motion fidelity, at the cost of an extra pass per step.

How it works

The node builds a guider from your model plus positive and negative conditioning, mixing standard CFG with the STG perturbation. You then feed that GUIDER into a custom sampler (SamplerCustomAdvanced-style) rather than using a plain KSampler. During sampling it does the normal denoise, the layer-skipped denoise, and combines them under your cfg, stg, and rescale settings.

The inputs and outputs that matter

• model, positive, negative - your LTX model and conditioning. Standard wiring.
• cfg (FLOAT, default 1) - classic guidance scale. Distilled LTX checkpoints run at CFG 1, which is why the default is 1; on dev checkpoints you'll push it higher.
• stg (FLOAT, default 1) - the STG strength, the knob that actually matters here. It sets how hard the model steers away from its skip-layer self. More is not always better (see below).
• rescale (FLOAT, default 0.7, range 0–1) - a guidance-rescale factor that reins in over-saturation and blown-out contrast from strong guidance. The 0.7 default is a sane starting point.

Output is a single GUIDER. Wire it into your custom sampler's guider input, alongside your noise, sigmas, and sampler nodes.

How to install it

The pack's own recommendation is ComfyUI Manager: Ctrl+M → Install Custom Nodes → search LTXVideo → Install → restart. Manual:

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

then restart ComfyUI. It shows up under the LTXV menu. STG itself adds no downloads, but it runs on top of the LTX-2 / LTX-2.3 model, which Lightricks lists at 32GB+ VRAM and 100GB+ disk. Models fetch on first use.

Common issues & troubleshooting

Overcooked hair and skin. This is the classic STG failure and it's well documented in the community - crank STG too high and you get "an exaggerated and unnatural appearance," especially on hair and faces. If your subjects look plasticky or over-etched, bring stg down. It's a subtle enhancement, not a slider to max.

Blown-out highlights or oversaturation. That's guidance overshooting. Keep rescale around 0.7, and if lights still blow out, lower it or ease cfg back. LTX-2.3 users specifically report tuning CFG in the 1.1–2.0 range when highlights overcook on single-stage T2V.

Distilled models: leave CFG at 1. Distilled LTX doesn't need classifier-free guidance - that's the point of distillation. Fighting it with high CFG mostly makes things worse. STG still helps on distilled, just keep the guidance scale where the checkpoint expects it.

It's slower than a plain sampler. Expected. STG does an extra denoise per step. Given LTX's absurd base speed you usually won't care, but if you're chasing every second on a big batch, that's the trade you're making.