LTXV Sigma Tail / Trim

LTXV Sigma Tail / Trim is the odd one out in this pack: it has nothing to do with LoRAs, prompts, or segments. It's a small utility that trims a ComfyUI SIGMAS tensor so you can run a second or third pass starting partway through the denoising schedule - think "I like the first pass but I want to refine only the low-noise tail." Normally that means converting sigmas to floats, slicing them in some clunky helper node, and converting back. This node slices the native tensor directly, so it works even when your SIGMAS-to-float helper nodes are being difficult.

It's a genuinely useful little thing to keep on the canvas, and it's refreshingly model-agnostic - you can use it with an LTX 2.3 workflow or anything else that produces a SIGMAS tensor.

How it works

You feed it the output of a BasicScheduler (or KSamplerSelect/SamplerCustom sigmas) and pick one of four trim modes:

• drop_first_count - chop the first N sigma values (default 2). The standard "skip the high-noise steps" move.
• keep_last_count - keep the final N values, terminal zero included.
• keep_last_fraction - keep the last 75% (say) of the schedule, handy when you want "roughly the last third."
• start_at_or_below_sigma - start at the first step whose sigma is at or below a threshold like 0.85. This is the mode that maps to "I want to denoise everything below this noise level."

ensure_zero_end (on by default) re-appends the terminal zero if your slice chopped it off - samplers expect that last step, so leave it on until you have a reason not to. csv_precision just controls the number of digits in the CSV output.

Outputs

• sigmas - the trimmed tensor, wired back into your SamplerCustom/KSampler.
• sigmas_csv - a comma-separated text dump, for eyeballing or for feeding a text display node.
• info - a formatted summary with input/output counts, start index, first/last values, and the reason string for the mode. This is the "did I actually trim what I meant to trim?" check, and it's worth glancing at.

Installing and gotchas

Standard pack install, under "LTXV/Timing":

cd ComfyUI/custom_nodes
git clone https://github.com/Jinx138/ComfyUI-LTXV-TimeGated-LoRA.git

Restart ComfyUI, hard-refresh. No pip packages, no model downloads, no PromptRelay dependency - this node works fully standalone, which the README explicitly calls out.

Two small traps. First, the trimmed schedule must keep at least two values - the node raises a clear error if your settings would slice below that, so ease off keep_last_count or keep_last_fraction if you hit it. Second, think about what "second pass" means here: dropping the first steps of a schedule is not the same as lowering denoise strength on a latents input. This is a refinement tool for people doing iterative pass experiments - if that's not your workflow yet, the mode you'll reach for is start_at_or_below_sigma, which is the most intuition-friendly of the four.