🎯 STG Guider

LTX's spatio-temporal guidance is a different animal from plain CFG, and it's what makes the distilled checkpoints hold together at 8 steps. The LTX2STGGuider is this pack's implementation - and it's a proper one, with per-step control and the exact combination formula the model was tuned around, not a stripped-down approximation.

Plain CFG runs two forwards per step: positive and negative, then combines pos + (cfg-1)*(pos-neg). STG adds a third pass that perturbs self-attention on a couple of chosen transformer blocks (a "V-shortcut" that skips the QK attention and returns V directly) and blends it in:

pred = pos + (cfg - 1)*(pos - neg) + stg_scale*(pos - perturbed)

The perturbed pass is the model second-guessing itself on spatio-temporal consistency, which is what fixes the "video knows what it should look like but forgets how it should move" failure mode. There's also an optional STD rescale so the combined prediction's magnitude stays stable against the positive pass.

The inputs that matter

The output is a single GUIDER, wired into SamplerCustom alongside your model, conditionings, sigmas, and sampler. Required inputs: model, positive, negative, and sigmas - that last one matters because per-step lists have to match its length. Then three comma-separated per-step lists:

• cfg_per_step - default 2.0, 1.5, 1.0... ramping down to 1.0. For distilled CFG=1 setups, set the whole list to 1.0.
• stg_scale_per_step - the STG strength. 0.0 disables STG for that step entirely (skips the perturbed forward, saving compute).
• stg_rescale_per_step - STD rescale, typically 1.0 throughout.

You'll rarely touch these because each one has a sigma_curve mode: set cfg_min/cfg_max (or stg_min/stg_max) once and the node interpolates linearly in sigma space. The author's note is worth respecting: it interpolates in sigma, not step index, and LTX2's schedules concentrate non-uniformly - so the curve is intentionally steeper at the start. That's a feature, not a bug.

block_indices (default 14, 19) picks which blocks get the self-attention perturbation - these two mid-depth blocks are the upstream Lightricks defaults. More blocks dramatically strengthens STG, so reduce stg_scale proportionally if you add any. Setting an out-of-range index like 9999 disables the perturbation while keeping the per-step CFG schedule alive - handy for A/B tests.

The two tricks that make it sing

First: for DMD/Echo workflows, this guider replaces CFGGuider and handles the distilled CFG=1 case cleanly. Second, the cheap compute win - the official Echo schedule's init cluster (sigmas ≈ 1.0) doesn't need the negative pass. Set cfg_per_step to 1.0 for those steps and the guider skips uncond, which is the "right architectural layer" for that saving.

One caveat: the tooltip on model is worth reading - chain this after anchor/clamp nodes. Those operate inside the model's forward; this guider orchestrates three forwards from outside, and the two coexist cleanly if the anchors are on the model first. The guider is a GUIDER, not a sampler, so pair it with SamplerCustom.

Install is the pack-wide clone (10S_Nodes into custom_nodes, restart, or ComfyUI Manager → "10S-Comfy-nodes"). No extra deps, LTX2-specific, and - like most good LTX advice - the distilled checkpoints are where it pays off most.