SamplerDPMPP_3M_SDE_motorway_edition

If SamplerDPMPP_2M_SDE is the sampler most people default to, SamplerDPMPP_3M_SDE is the one people switch to when they're about to render the final version. This motorway-edition clone of ComfyUI's built-in SamplerDPMPP_3M_SDE does everything its 2M sibling does - the convergent multistep solver plus controlled stochastic SDE noise - but with a third-order integrator instead of a second-order one. In practice that usually means a bit more convergence accuracy and a bit more resolved detail at the same step budget, in exchange for a little more compute per step.

It's a sampler factory, not a full KSampler. It takes a model, does its math, and hands back a SAMPLER object that you feed into SamplerCustom's sampler port (or, in Motorway land, park on the bus for a later SamplerCustom_motorway_edition to pick up). For SD 1.5 and SDXL - the DDPM-style model families where DPM++ shines - 3M SDE is a common "final render" pick over 2M SDE: same general character, slightly cleaner convergence, and the SDE noise keeps the image from going plastic.

What it does on the Motorway

This node is one of the generated *_motorway_edition clones from the ComfyUI_agilly1989_motorway pack. The pack's whole idea is a named-key data bus - the Motorway - where you store values under typed key names instead of wiring them port to port. MotorwayStart kicks off a bus, Motorway AxB ramps store and fetch keys, and every motorway-edition node is a clone of a built-in ComfyUI node re-plumbed to pull its inputs from the bus by key and push its outputs back by key. Each one returns the bus so the chain keeps going.

The widgets you actually set are the same trio as the 2M SDE edition:

• eta (default 1.0) - scales the extra SDE noise. Standard value is 1.0.
• s_noise (default 1.0) - per-step stochastic noise strength. Drop toward 0 and you lose the SDE flavor.
• noise_device - gpu (faster, the default) or cpu (offloads the noise generation if VRAM is tight).

The model comes off the bus by key - the family convention, INPUT_…_key, which the SDTurboScheduler edition in this same pack shows as INPUT_model_key defaulting to model - and the finished sampler is written back under your OUTPUT_SAMPLER_key (default SAMPLER). Tell any downstream motorway node to read SAMPLER and it's got the sampler.

Installing it

It ships inside ComfyUI_agilly1989_motorway - no standalone download. ComfyUI Manager: search ComfyUI_agilly1989_motorway, install, restart. Or from the terminal:

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

No extra dependencies, no model files. Pure Python glue over ComfyUI's own nodes.

The catch

The recurring gotcha: release 1.1.7 ships with the clone nodes disabled. The two ClonedNodeMapping lines in custom_nodes/ComfyUI_agilly1989_motorway/__init__.py are commented out (a "HOTFIX - REMOVED MONKEYPATCH" commit), so unless you uncomment them and restart, SamplerDPMPP_3M_SDE_motorway_edition won't exist in your menu. And the README itself sets expectations low - "BIG BROKEN WITH ASYNC WILL FIX WHEN I GET THE TIME/MOTIVATION", beta warnings, and a note that only "most" of the cloned nodes work. It's an honest little experiment from a solo dev, and fun to play with - just don't bet a production workflow on it.