LTXV Apply STG (Ray)

LTXV Apply STG (Ray) is the raylight version of the spatio-temporal guidance patch, the thing that did more than almost anything else to make early LTX output usable. If you've loaded an LTX workflow and seen a node sitting between the model and sampler that you were told "just make sure it's there" - this is that node.

Here's why it matters. Classifier-free guidance (CFG) pushes a generation toward your prompt, but on video models it has a well-known failure mode: it doesn't discriminate between "this improves the image" and "this introduces artifacts," and cranking it up tends to over-apply to the spatial dimension while leaving temporal coherence alone. Spatio-temporal guidance is a targeted fix - it applies the guidance to some transformer blocks and not others, which is how it sharpens subject detail and structure without nuking the video's motion coherence. It's the mechanism behind LTX's recommendation that distilled models can run with STG instead of heavy CFG.

The node's job is deceptively simple: it tells the model which transformer blocks get skipped for guidance. The block_indices input (default 14, 19) is a comma-separated list of block numbers. That default pair is the community-settled sweet spot for the LTX architecture - enough blocks patched to boost detail, few enough that motion stays stable.

What you set

• block_indices - the only input beyond ray_actors. Change the set of blocks the STG patch applies to.

That's it. No strength slider here - STG strength lives on the guider side (see the pack's Guider Parameters (Ray) and the STG guider nodes). This node is just the "which blocks" switch, patched into the transformer options on every Ray worker.

Things worth knowing

• The tooltip pattern 14, 19 - comma and space both accepted, so don't sweat formatting.
• This node composes with other transformer patches in the same graph: the code merges its block list with any existing skip_block_list, so you can chain it with LTX Perturbed Attention (Ray) or the Feta enhancer without them clobbering each other.
• It's an "apply to the model" node - you wire ray_actors through it and into the Ray sampler. If the workflow's guider already applies STG, this is where the block selection for that guider gets set.

The honest advice

Leave 14, 19 alone until you have a specific complaint. If your LTX output has weak subject structure, adding blocks to the list (or raising STG strength on the guider) pushes back. If motion starts feeling unstable or "jelly," that's the sign you went too far and should trim blocks or lower strength. STG over-application is a real, documented failure mode - it produces that exaggerated, waxy look on faces and skin - so resist the urge to max everything.

Install: ComfyUI Manager → "raylight", or clone + pip install -r requirements.txt. No model downloads.