Petty Paint SD Turbo Scheduler

SD Turbo is the model that made "preview in a second" a normal workflow: a distilled checkpoint that was trained to look good in one to four steps instead of the usual twenty-plus. The catch is that ComfyUI's stock schedulers will happily sample it for 30 steps and give you a washed-out mess - Turbo needs a sigma schedule built for very few steps. That's exactly what PettyPaintSDTurboScheduler produces: a SIGMAS output you feed straight into a KSampler's sigmas input, with nothing else to patch.

It's a small node and it does one thing, but it's the difference between "Turbo is useless in ComfyUI" and a sub-second preview loop.

How it works

Instead of a normal scheduler curve, it hand-builds the timestep ladder that SD Turbo's one-step training implies: the model sampling sigmas for timesteps 999, 899, 799, … 99 (ten points, evenly spaced), then slices that ladder by your steps and denoise, and tacks on a trailing zero sigma so the sampler actually reaches clean output. With the defaults - 1 step, denoise 1.0 - you get the classic single-step Turbo jump: one sigma at 999, one zero. Bump steps to 2–4 and you're in the territory where Turbo quality still climbs a little; past 4 you're mostly wasting time.

The inputs that matter

• model - the checkpoint you're sampling, so the node can ask it for its real sigma values (it's model-aware, not hardcoded to SD 1.5).
• steps - integer 1–10, default 1. 1–4 is the sweet spot for Turbo.
• denoise - float 0–1, default 1.0. Lower it for img2img-style partial renoise.

Both steps and denoise are forceInput, so wire them from somewhere - a constant node, or the pack's PettyPaintArguments bundle if you're standardizing on one settings source. The single output, SIGMAS, plugs into the sigmas input of any KSampler, including ComfyUI's core one.

Installing it

Part of the petty-paint pack:

cd ComfyUI/custom_nodes
git clone https://github.com/mephisto83/petty-paint-comfyui-node

Then restart ComfyUI, or use ComfyUI Manager and search "petty-paint-comfyui-node". Declared dependency is just Flask==2.1.2, which this node ignores.

Common issues

The number-one mistake is using it with a normal (non-distilled) checkpoint. One step on an SD 1.5 or SDXL model gives you a flat gray blob - Turbo sigmas are only valid for Turbo/LCM-style distilled models, so keep it on the checkpoints it's meant for. And remember it's a scheduler, not a sampler: it only produces the sigmas wire, so you still need a KSampler downstream of it. If you're new to distillation models, the general rule the community lands on is: Turbo for speed, LCM for a middle ground, and full multi-step samplers when quality beats latency.