alt_ltx23_scheduler

alt_ltx23_scheduler

A variation on the ComfyUI LTX 2.3 scheduler (LTXVScheduler) that exposes the power exponent, so you can decouple how high the schedule holds from how steeply it leaves the midpoint.

The node is registered as LTXVScheduler (Power) under model/sampling/schedulers. The stock node is left untouched.

⚠️ This mirrors the v0.24.0 stock LTXV scheduler math. If Lightricks changes the upstream formula in a future ComfyUI release, this fork won't track it automatically.

The formula

Each linspace value s (running 1 → 0 across the steps) is mapped to a sigma:

                    e^k
σ(s) =  ───────────────────────────         k = sigma_shift,  derived from
        e^k + (1/s − 1) ^ power                  max_shift / base_shift / tokens

• k (the shift) is a linear function of the latent token count, with base_shift at 1024 tokens and max_shift at 4096 tokens. Setting max_shift == base_shift pins k to that value regardless of token count.
• power is the new knob. It only touches the (1/s − 1) term, which equals 1 at the midpoint (s = 0.5) — so power changes the slope at the midpoint without changing the value there. power = 1.0 reproduces the stock node exactly.

power — the headline parameter

Held at max_shift = base_shift = 3.5 (so k = 3.5, ~0.97 at step 10), varying only power. All three curves share the same midpoint value; higher power drops out of it faster and spreads the descent more evenly instead of cliffing at the final step.

xychart-beta
    title "Effect of power (k fixed at 3.5, 20 steps)"
    x-axis "Step" [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20]
    y-axis "Sigma" 0 --> 1
    line [1.000,0.998,0.997,0.995,0.993,0.990,0.987,0.984,0.980,0.976,0.971,0.964,0.957,0.947,0.934,0.917,0.892,0.854,0.786,0.635,0.000]
    line [1.000,1.000,0.999,0.998,0.996,0.994,0.992,0.988,0.984,0.978,0.971,0.961,0.947,0.929,0.903,0.864,0.805,0.711,0.551,0.286,0.000]
    line [1.000,1.000,1.000,0.999,0.998,0.997,0.994,0.991,0.987,0.980,0.971,0.957,0.936,0.906,0.859,0.786,0.674,0.508,0.290,0.084,0.000]

Lines top-to-bottom in the second half: power = 1.0 (stock, flat-then-cliff), power = 1.5, power = 2.0 (steepest exit, gentlest final step). Recommended starting range: 1.5 – 2.0.

| step | power 1.0 | power 1.5 | power 2.0 | |-----:|:---------:|:---------:|:---------:| | 10 | 0.971 | 0.971 | 0.971 | | 12 | 0.957 | 0.947 | 0.936 | | 14 | 0.934 | 0.903 | 0.859 | | 16 | 0.892 | 0.805 | 0.674 | | 18 | 0.786 | 0.551 | 0.290 | | 19 | 0.635 | 0.286 | 0.084 |

max_shift / base_shift — the shift k

With power = 1 (stock behavior), raising the shift holds the curve nearer to 1 for longer — but because it's a plain logistic, a higher hold necessarily flattens the midpoint and pushes the descent into the last few steps. This is the coupling that power was added to break.

xychart-beta
    title "Effect of shift k (power = 1, 20 steps)"
    x-axis "Step" [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20]
    y-axis "Sigma" 0 --> 1
    line [1.000,0.993,0.986,0.978,0.969,0.959,0.948,0.935,0.921,0.905,0.886,0.864,0.838,0.807,0.769,0.721,0.660,0.578,0.463,0.290,0.000]
    line [1.000,0.997,0.994,0.991,0.988,0.984,0.979,0.974,0.968,0.961,0.953,0.943,0.931,0.915,0.896,0.870,0.834,0.780,0.691,0.514,0.000]
    line [1.000,0.999,0.998,0.997,0.995,0.994,0.992,0.990,0.988,0.985,0.982,0.978,0.973,0.967,0.959,0.948,0.932,0.906,0.858,0.742,0.000]

Lines bottom-to-top: k = 2.05 (stock default), k = 3.0, k = 4.0. Note how k = 4.0 clings to 1.0 then falls off a cliff — exactly the shape power is meant to fix.

Tip: set max_shift == base_shift to make k independent of the latent's token count, so the schedule doesn't drift when you change resolution or frame count.

stretch / terminal

stretch linearly remaps the non-zero sigmas so the last one lands on terminal instead of near 0. It's a rescale, not a reshape — it lifts the tail without changing the overall curvature. Shown here on top of k = 3.5, power = 2.0.

xychart-beta
    title "Effect of stretch / terminal (k = 3.5, power = 2.0)"
    x-axis "Step" [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20]
    y-axis "Sigma" 0 --> 1
    line [1.000,1.000,1.000,0.999,0.998,0.997,0.994,0.991,0.987,0.980,0.971,0.957,0.936,0.906,0.859,0.786,0.674,0.508,0.290,0.084,0.000]
    line [1.000,1.000,1.000,0.999,0.998,0.997,0.995,0.991,0.987,0.981,0.971,0.958,0.937,0.907,0.861,0.790,0.680,0.516,0.303,0.100,0.000]
    line [1.000,1.000,1.000,0.999,0.999,0.997,0.996,0.993,0.990,0.985,0.978,0.967,0.951,0.928,0.892,0.837,0.751,0.624,0.458,0.300,0.000]

Lines bottom-to-top in the tail: stretch off, terminal = 0.1, terminal = 0.3. The final step still snaps to 0 either way.

sigma_max — scale the whole curve

A final scalar multiply applied after everything else. The schedule's natural peak is 1.0 at the first step, so sigma_max simply becomes the new starting sigma and every other value scales by the same factor. 1.0 is a no-op; the trailing 0 is preserved (0 × x = 0). Shown on k = 3.5, power = 2.0.

xychart-beta
    title "Effect of sigma_max (k = 3.5, power = 2.0)"
    x-axis "Step" [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20]
    y-axis "Sigma" 0 --> 1
    line [1.000,1.000,1.000,0.999,0.998,0.997,0.994,0.991,0.987,0.980,0.971,0.957,0.936,0.906,0.859,0.786,0.674,0.508,0.290,0.084,0.000]
    line [0.750,0.750,0.750,0.749,0.749,0.747,0.746,0.743,0.740,0.735,0.728,0.718,0.702,0.679,0.644,0.590,0.506,0.381,0.218,0.063,0.000]
    line [0.500,0.500,0.500,0.500,0.499,0.498,0.497,0.496,0.493,0.490,0.485,0.478,0.468,0.453,0.429,0.393,0.337,0.254,0.145,0.042,0.000]

Lines top-to-bottom: sigma_max = 1.0, 0.75, 0.5 — the same shape, uniformly scaled down. Unlike stretch/terminal (which lift only the tail), this rescales the entire curve including its peak.

Inputs

| input | default | notes | |--------------|:-------:|-------| | steps | 20 | number of sampler steps | | max_shift | 2.05 | shift k at 4096 tokens | | base_shift | 0.95 | shift k at 1024 tokens | | power | 1.0 | midpoint-slope exponent; 1.0 = stock, try 1.5–2.0 | | sigma_max | 1.0 | scales the whole curve; becomes the new starting sigma (1.0 = no-op) | | stretch | true | (advanced) rescale tail to terminal | | terminal | 0.1 | (advanced) value the last non-zero sigma maps to | | latent | — | optional; its token count drives k between the two shifts |

Install

Clone/copy this folder into ComfyUI/custom_nodes/ and restart ComfyUI. The node loads via the standard comfy_entrypoint extension mechanism.