Nodes/ComfyUI_agilly1989_motorway/SDTurboScheduler_motorway_edition
ComfyUI Node

SDTurboScheduler_motorway_edition

The SD-Turbo scheduler, 1–4 steps, riding the Motorway

By agilly1989·Created 2 years ago·Updated about a year ago· 7
SDTurboScheduler_motorway_edition
  • MOTORWAY 🚌💨
  • MOTORWAY 🚌💨
INPUT_model_keymodel
steps1
denoise1.00
OUTPUT_SIGMAS_keySIGMAS

Here's the mental model: SD-Turbo models are distilled to be fast, and they only behave well in a very narrow lane. SDTurboScheduler_motorway_edition is the clone of ComfyUI's built-in SDTurboScheduler, and its whole job is to lay down the sigmas (the noise schedule) that turbo models actually like. Distilled models - SDXL Turbo, SD 1.5 Turbo, and friends - were trained to produce an image in 1 to 4 steps, and they fall apart if you treat them like a normal checkpoint. The scheduler is half of making that work; the sampler is the other half (the community standard is Euler). This node replaces the generic schedulers in the custom-sampling chain so you don't hand a turbo model a 20-step Karras schedule it was never trained for.

The KB's distillation notes give you the real-world numbers: SDXL Turbo wants CFG around 2 and 1–4 steps with Euler; SD 1.5 Turbo wants around 1 step. This scheduler encodes the "tiny number of steps" assumption directly, which is why its step range is capped the way it is.

The widgets:

  • steps (default 1, 1–10) - the step count for the schedule. 1 is the aggressive SD 1.5 Turbo default; SDXL Turbo usually wants 2–4. This is the knob that defines the whole character of the output.
  • denoise (default 1.0, 0–1) - how much of the schedule to actually run. 1.0 is full denoise. Lower values give the img2img "partial redraw" behavior.

Plumbing: it reads the model off the bus via INPUT_model_key (default model) - it needs the model to size the schedule correctly - and writes the resulting sigmas back under OUTPUT_SIGMAS_key (default SIGMAS). That SIGMAS value is what a later sampling node (on the bus or wired) consumes for its schedule.

Like the other sampler/scheduler nodes in this pack, this one is a factory - no images flow through it. It produces a SIGMAS object for the custom-sampling chain.

Installing it

Part of ComfyUI_agilly1989_motorway. ComfyUI Manager: search ComfyUI_agilly1989_motorway, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/agilly1989/ComfyUI_agilly1989_motorway

No extra dependencies, no model files - though you'll want an actual turbo checkpoint to use it for anything.

The catch

The pack's usual gotchas: release 1.1.7 ships with the clones disabled, so uncomment the two ClonedNodeMapping lines in custom_nodes/ComfyUI_agilly1989_motorway/__init__.py and restart to see any *_motorway_edition node. And the README - first line "BIG BROKEN WITH ASYNC WILL FIX WHEN I GET THE TIME/MOTIVATION" - is an honest beta from a solo dev. One more real-world note: a turbo scheduler is only useful with a turbo model. Slap it on a regular SDXL checkpoint and you'll get the "1 step" look, which is to say a smeared mess. Match the scheduler to the model family and this works great; mismatch it and it's the fastest way to confuse yourself.

Categoryagilly1989 Nodes/Motorway-ed/sampling/custom_sampling/schedulers

Inputs (5)

NameTypeDefaultDescription
MOTORWAY 🚌💨MOTORWAY 🚌💨
INPUT_model_keySTRINGmodel
stepsINT11–10
denoiseFLOAT1.000–1
OUTPUT_SIGMAS_keySTRINGSIGMAS

Outputs (1)

NameTypeDescription
MOTORWAY 🚌💨MOTORWAY 🚌💨