GITSScheduler

GITSScheduler is a fossil with a fancier name than it deserves. It's the scheduler that picks your sigma schedule from a set of precomputed, "optimal" noise levels instead of computing one from a formula. Back in mid-2024, when it landed in ComfyUI core, that idea got a genuine burst of attention - then it fell off a cliff. Today it's near-zero in community discussion, and understanding why is half the point of this node.

What GITS actually is

The name comes from the GITS method in the zju-pi/diff-sampler research repo ("Guided Interval Time Steps" in spirit: sweep the noise schedule, find good ones, ship the table). Someone did that homework for SD 1.5-era models and encoded the results as a giant lookup table: for a coeff value between 0.80 and 1.50, the node has precomputed schedules for 2 up to 20 steps, each an EDM-style curve running from sigma_max ≈ 14.61 down to sigma_min ≈ 0.029 - the classic SD 1.5 noise range.

If you ask for more than 20 steps, it takes the 20-step table and log-linearly interpolates it out to your count. The denoise input just truncates the schedule to the first round(steps * denoise) steps, same as every other scheduler in this family.

The inputs that matter

• coeff (FLOAT, 0.80–1.50, default 1.20) - the shape knob. It picks which table. Sensitive, but 1.2 is the sane starting point.
• steps (INT, 2–1000, default 10) - how many sigmas you get.
• denoise (FLOAT, 0–1, default 1.0) - partial runs.

One SIGMAS output, straight into a sampler's sigmas port or KSampler's.

Why it died (and when it's still worth a try)

It shipped in June 2024, spiked in July–August 2024, and then the mention count basically flatlines through 2025 and 2026. The reason is the same one that killed the whole "clever schedule" genre: GITS's tables were tuned for DDPM-style models - SD 1.5, SDXL, and their finetunes. Flow-matching models (Flux, Wan, Z-Image) train on a near-straight trajectory and actively dislike aggressive schedule reshaping. Karras and exponential fail outright there, and GITS is cut from the same cloth. On a flow-matching checkpoint it's not a mild regression, it's the wrong tool.

That leaves the honest use case: SD 1.5/SDXL work at low step counts where you want a hand-tuned curve and are tired of rolling your own. On those it can beat the plain normal schedule at 8–15 steps - the tables were genuinely measured, not guessed.

Common issues

• Using it on the wrong model family. If output degrades badly, check you're on SD 1.5/SDXL territory, not Flux or Wan.
• Ignoring coeff. It's not a decoration. 0.8 and 1.5 are meaningfully different curves and, on some checkpoints, the difference between sharp and mushy.
• Expecting a modern result. It's a fixed-curve trick from a single research lab. Treat it as a niche option, not an upgrade. If a workflow you grabbed online uses it, you can usually swap in a plain normal or beta scheduler with comparable results - which is, in a nutshell, why the community moved on.