SkipLayerGuidanceSD3_motorway_edition

The SD3 sibling of the DiT skip-layer guidance node, on the Motorway. If you're running an SD3-based model and want the detail-and-adherence boost that skip-layer guidance gives, this is the version that speaks the SD3 architecture's layer layout instead of a DiT's double/single block split.

Here's the quick version of what it does under the hood: during sampling it runs the model forward twice - once normally, once with the chosen transformer layers dropped - and steers the generation toward the difference between the two predictions. Skipping layers on purpose sounds like sabotage, but the discrepancy is a rich signal, and nudging toward it pulls out texture and structure that plain CFG can smear away. On SD3-family models it's a genuinely useful trick and one most people never try.

The structure is simpler than the DiT variant because SD3's transformer is more uniform: one layers list instead of separate double and single blocks.

What to set

• INPUT_model_key (default model) - the Motorway key holding the MODEL you're patching.
• layers (default "7, 8, 9") - comma-separated indices of transformer layers to skip. The defaults are a proven starting point; drop earlier layers if you want subtler guidance.
• scale (default 3, 0–10) - the push strength. Bump it and the effect gets more aggressive, and more prone to artifacts.
• start_percent / end_percent (defaults 0.01 / 0.15) - the denoising window where the guidance is active. Early window, shuts off after the first 15% of steps.
• OUTPUT_MODEL_key (default MODEL) - where the patched model lands in the Motorway.

Output is the MOTORWAY with MODEL written under that key, ready to feed a KSampler.

Installing it

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

Restart and it's under agilly1989 Nodes → Motorway-ed → advanced → guidance. Manager installs the pack too. There's no extra dependency - this wrapper leans entirely on ComfyUI's own node and Comfy's execution, which is the whole point of the _motorway_edition line.

Where people get burned

The two usual suspects. Key mismatch, announced as 'MotorwayClass' object has no attribute 'hash_', means the model key isn't in the Motorway when this node runs - check the key on your checkpoint/loader node and get the spelling exactly right, keys are case-sensitive.

And the SLG-specific one: the start_percent/end_percent window. The guidance only lives in that slice of denoising. Run a short-step pipeline and a 0.01–0.15 window may be over before the sampler has done anything, which is why you'd set start_percent slightly higher or just accept the effect is front-loaded. Finally, this whole pack is beta - the README literally leads with "IF THINGS BREAK ITS BECAUSE I BROKE IT." If a Motorway update eats the motorway-ed nodes, that's expected churn, not a misinstall.