STG Guider (Ray)

LTX Video is fast - that's its whole identity - but speed came with a price: without spatio-temporal guidance it produces soft, structurally lazy video. STG is Lightricks' answer, and if you've used LTXV 0.9.x in ComfyUI, you've seen the STGGuider node in every working workflow. RaySTGGuider is the same guider, built to run on Raylight's distributed workers, and it's genuinely the one to reach for when your Raylight LTX output looks undercooked.

STG works by perturbing where the model is forced to pay attention: it runs a second pass where the video's spatial and temporal self-attention blocks are disabled, so the model can't use its usual shortcuts to reconstruct the frame sequence, then extrapolates the clean prediction away from that degraded one. The result is sharper spatial detail and stronger temporal coherence - less swimming, less mush - and it's the technique that turned LTXV from "impressive speed demo" into a usable model.

The three inputs are the same trio the original node uses, and the defaults are the community consensus:

• cfg (default 1.0) - the classifier-free guidance strength. LTX runs CFG 1; don't raise this expecting more adherence, it mostly just doubles the time.
• stg (default 1.0) - STG strength. The official guidance recommends values around 1.0 for 0.9.x, and 1.0 is where people actually run it.
• rescale (default 0.7) - rescales the STG-extrapolated result's magnitude. Same spirit as RescaleCFG: prevents the push from blowing out the output. 0.7 is the known-good default.

It takes positive and negative CONDITIONING inputs (for LTX the negative is usually the empty conditioning) and outputs a RAY_GUIDER - not a ray_actors stream. That guider feeds the Raylight KSampler's guider input, which is a different wiring than the patch-style nodes in this pack.

Where it fits

The Raylight LTXV graph: Ray Init Actor → LTXV model loading → this guider → the Ray KSampler. If you're coming from a single-GPU LTX workflow, think of this as the drop-in replacement for the stock STGGuider.

Install

Part of raylight:

cd ComfyUI/custom_nodes
git clone https://github.com/komikndr/raylight
cd raylight
pip install -r requirements.txt

Restart, or install "raylight" via ComfyUI Manager.

Gotchas

• LTXV's distilled models (0.9.6+) explicitly don't need STG or CFG - the KB's LTX guide quotes the release notes saying exactly that. On distilled weights this guider is wasted work; the non-distilled 0.9.x models are where it earns its keep.
• It's a guider, not a patch: wire its guider output into the sampler, don't chain it as another ray_actors node.