LTXV Spatio-Temporal Guidance (STG)

LTX is the speed tier of local video - seconds per clip where Wan takes minutes - and the classic complaint that comes with it is that output can look soft and smeary, with motion that drifts into mush. LTXV Spatio-Temporal Guidance (STG) is the built-in dial for exactly that: it trades a little speed for sharper spatial detail and more coherent motion. It's the same trick Lightricks shipped as STGGuiderAdvanced back in the 0.9.6 era, now baked into core.

How it works

STG is a "guide away from a broken pass" node. It clones your model and registers a post-CFG function. On every sampling step inside its active window it runs one extra forward pass where the self-attention of the transformer blocks you pick is degraded to a value passthrough - the attention weights, positional info, and temporal mixing are skipped, and each block just pushes its value projection through. Long-range mixing is gone, so that pass predicts mush. Then the node steers the real result away from it: result + (clean_pred - mush_pred) * scale.

The logic is that the degraded pass has no idea how things relate across space or time, so "not that" is a gradient toward more spatial detail and more motion coherence. It's the same family as CFG and the other guidance nodes - find a degenerate version, then run from it.

The inputs that matter

• model - your LTX checkpoint from the loader. Cloned internally; the original stays untouched.
• scale - guidance strength. Default 1.0, range up to 100. 0 is a hard no-op (the node returns early, no extra pass), which makes it your cheapest off switch.
• blocks - comma-separated transformer block indices to degrade. Default is 29. Only the digits are parsed - "12, 29, 31" works, and stray text is ignored. The LTX-AV transformers are 48 blocks deep (0–47), so 29 is a mid-to-late block, and Lightricks' old guidance singled it out for a reason.
• start_percent / end_percent - the slice of the denoise where guidance runs; defaults cover the whole run.

Single MODEL output that feeds your sampler like normal. It stacks with LTXV Modality Guidance and the dual-CFG guider - that trio is ComfyUI's "multimodal guider," one dial for detail, one for A/V sync, one for prompt adherence.

Getting it

Core node, nothing to install. It landed in ComfyUI's nodes_lt.py alongside native LTX 2.5 support in mid-August 2026, so it's days old - update if you don't see it. No model files of its own; it patches whatever checkpoint you give it.

What actually happens, and the honest gotchas

• It costs a forward pass per step. In the 0.9.x custom-node days people measured roughly 20–25% slower iterations with STG on; expect the same ballpark here. On LTX that still leaves you fast by video-model standards, but don't pretend it's free.
• Over-apply and it turns ugly. The community verdict from the older STG era is consistent: crank it and hair and skin get "an exaggerated and unnatural appearance." Start at 1.0, nudge up in small steps, stop at the first sign of plastic.
• The results were honestly disputed even back then. One careful i2v A/B found no clear quality gain from STG at any layer; the counter was "it's the little details done right - one popped eye is enough to ruin a video." So treat it as a tool, not a checkbox, and A/B it on your own clips.
• Distilled models don't need it. The 0.9.6 release notes said the distilled checkpoint doesn't require CFG or STG. This is a dev-model quality dial - on distilled weights you're mostly burning your extra pass for nothing.

Pick a couple of blocks to play with once you've got the scale dialed. Block choice is genuinely a taste thing - 29 is the shipped default, and it's the right first stop, not the law.