MD: GITS Scheduler (Boomerang)

Schedulers decide where your sampling steps actually go. A dumb schedule spreads them evenly; a smart one concentrates them where the image is changing most. GITS ("gradient-informed timestep scaling") is the pack's smart one - it warps the schedule with a tanh curve so steps cluster in the high-curvature middle of denoising, where most of the detail actually gets decided, and thins out over the stable early and late regions.

The "boomerang" in the category name is the shape: the tanh curve spends a burst of steps mid-flight and comes back. The core file describes it exactly that way - "GITS 'Boomerang' (tanh) S-Curve Schedule Generation."

How it works

Feed it a model and it auto-detects that model's native sigma range (max/min noise limits), which the tooltip insists on - it's how the schedule ends up mathematically correct instead of guessed. You pick a steps count and a curvature_preset, and it returns three outputs: sigmas (wire into your KSampler's scheduler slot), plot (a rendered image of the schedule so you can actually see where steps landed), and schedule_info (an analytics string).

The curvature presets are the dial:

• Linear (0.0) - no clustering.
• Standard (1.0) - balanced for the usual 20–30 steps.
• Heavy (2.0) - made for 8–12 step Lightning/Turbo workflows. The tooltip is explicit: use Heavy for distilled models.
• Extreme (3.5) - ultra-low step counts only.
• Custom - activates the curvature_scale slider (0–5) for manual clustering intensity.

If that language sounds familiar, it's the same intuition behind AuraFlow's shift and the flow-matching schedulers the ecosystem is converging on: effort distribution is model-specific, and a schedule tuned to a distilled 8-step model is a different beast from one for a 30-step run. GITS is this pack's take on that knob.

The inputs that matter

• steps - 8–12 with Heavy, 20–30 with Standard. The tooltip gives you the pairing directly.
• denoise - truncates the schedule for img2img. 1.0 for txt2img, 0.4–0.7 for refinement. Lowering it slices from the start, which preserves the low-noise steps - the opposite of how some people expect it to work.
• sigma_override_max / sigma_override_min - manual noise-range overrides. Leave at 0 (auto) unless you know why; the tooltip warns overrides can cause burn-in.

Why you'd bother

If your images are a mess at low steps, the fix is often not "more steps" but "put the steps in the right place." That's the whole argument for this node. It's also unusually debug-friendly for a scheduler - the plot output shows you the schedule shape, and debug_mode (with an "Analytics Report" option) prints step-density metrics, so you can actually see where the steps went instead of trusting a formula.

The honest caveat: it's a niche tool from a niche pack. If you're happy with the built-in schedulers plus a standard shift node, GITS is solving a problem you may not feel. But for distilled checkpoints and for anyone who wants the schedule visible, it's a thoughtful alternative.

Installing

It's part of MD Nodes:

cd path/to/ComfyUI/custom_nodes
git clone https://github.com/MDMAchine/ComfyUI_MD_Nodes.git
cd ComfyUI_MD_Nodes && pip install -r requirements.txt

Or via ComfyUI Manager (search MD Nodes), then restart. As with the whole pack, the install is heavier than the node warrants - audio libraries, matplotlib, and client SDKs all come along.