🅛🅣🅧 STG Guider

STG Guider is the beginner-friendly half of LTX's guidance story. It bundles Spatiotemporal Skip Guidance into a single node with three knobs, and it plugs into an LTXV sampler the same way a KSampler's CFG widget would - except instead of just steering toward your prompt, it also steers away from a deliberately degraded version of the model to tighten up motion and structure. If STG Guider Advanced looked like a spreadsheet and scared you off, this is the one to start with.

It comes from ComfyUI-LTXVideo, Lightricks' own node pack for their LTX video models. STG showed up in the 0.9.6 workflows and stuck around because it genuinely helps the thing LTX is weakest at - coherent motion - without costing you a training run or extra models.

How it works

Spatiotemporal Skip Guidance is a training-free method: during sampling it runs the transformer a second time with certain attention layers skipped, producing a worse prediction, then pushes the result away from that worse prediction. The effect is a bit like classifier-free guidance, but the "negative" here is the model's own crippled self rather than your negative prompt. Because video transformers carry motion in specific layers, skipping them and steering away sharpens temporal consistency and detail - the node's description puts it as improving quality "without sacrificing diversity or motion fidelity."

The inputs that matter

You feed in model, positive, and negative conditioning, then set three floats:

• cfg - classic classifier-free guidance strength. Default is 1, which is correct for distilled LTX checkpoints. Dev checkpoints can take more.
• stg - how hard the skip-guidance pushes. Default 1. This is the dial that adds or removes the STG "punch"; turn it up for more structure, but not too far (see below).
• rescale - default 0.7, a normalization that keeps the guided prediction from drifting in overall brightness/contrast. Leave it near default unless colors start shifting.

The single output is a GUIDER, which wires into the guider input of LTXV Base Sampler, Tiled Sampler, or Looping Sampler. That's the whole node - three numbers and a socket.

Installing it

Through ComfyUI Manager: search LTXVideo under Install Custom Nodes, install, restart. Or manually: cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo and restart ComfyUI. It lands under the lightricks/LTXV menu. The node has no weights of its own; the LTX checkpoint and (for LTX-2) the Gemma text encoder are the heavy downloads, pulled on first use.

Where people get burned

• Cranking STG too high. The most common complaint is that heavy STG gives hair and skin an exaggerated, unnatural look. If your subject starts looking rubbery, drop the stg value before touching anything else.
• Raising cfg on a distilled model. Distilled LTX checkpoints are trained for CFG 1. Push cfg up on them and you often get worse output, not better - save higher CFG for the dev checkpoints.
• Expecting it to fix bad prompts. STG sharpens what the model is already doing; it won't rescue a short, vague prompt. LTX is famously prompt-hungry - long, specific, chronological descriptions do more than any guider setting.
• Reaching for it when you need per-step control. STG Guider applies one setting for the whole denoise. If you want heavy guidance early and none late (usually the better recipe on dev checkpoints), that's exactly what STG Guider Advanced is for.