SkipLayerGuidanceSD3

SkipLayerGuidanceSD3 is where skip-layer guidance started in ComfyUI, and the honest thing to say is that it's now mostly historical. It was the first SLG implementation, built for SD 3.5 back in October 2024 by Dango233@StabilityAI - you can feel the family's shared DNA in the name. Then, in July 2025, ComfyUI shipped the generic SkipLayerGuidanceDiT node, and this one quietly became a thin wrapper around it. The source of SkipLayerGuidanceSD3 is literally a one-line call into the DiT node passing your layers as the double_layers argument.

So why would you use it? Two reasons: you downloaded an SD3 workflow that references it and want the same behavior with zero thought, or you're on an SD3-lineage model and the simpler, SD3-shaped interface (one layers string instead of separate double/single fields) is exactly what you want. Otherwise, prefer SkipLayerGuidanceDiT - it's the same mechanism plus single-layer control.

How it works

Identical mechanism to the DiT node, minus the single-layer option. During the sigma window (default 1%–15%), the model runs an extra forward pass with your chosen transformer blocks patched out to no-ops. The difference between that degraded prediction and the normal conditional is added back on top of the CFG result, scaled by scale:

cfg_result = cfg_result + (cond_pred - slg) * scale

The skipped layers produce a prediction with coarse structure but no fine detail; subtracting it isolates exactly the detail those layers contribute, and re-adding it with force sharpens the render. Inspired by Perturbed Attention Guidance (arXiv 2403.17377) and, per the source, intended to "enhance guidance towards detailed structure."

Inputs and what to set

• model - the MODEL to patch.
• layers - comma-separated block indices, default 7, 8, 9. On SD3.5, layers 7–9 are late double blocks where detail concentrates. SD3.5 has fewer total blocks than Flux, so check the architecture if a higher index seems to do nothing.
• scale - default 3.0, range 0–10. Detail boost strength; 2–4 is the sensible band.
• start_percent / end_percent - default 0.01 / 0.15. When the extra pass runs; keep it short, since each in-window step costs double.

Output is the patched MODEL. Ships with ComfyUI core.

Common issues & troubleshooting

It's a wrapper, so behavior tracks the DiT node. Improvements and quirks in SkipLayerGuidanceDiT show up here automatically. If you're debugging, read the DiT node's docs - you're debugging that.

Only double blocks, no single-layer option. That's the main functional gap versus the generic node. If you want to skip single blocks too, switch to SkipLayerGuidanceDiT.

Same cost profile, same experimental flag. One extra forward pass during the window; is_experimental applies. On SD3's distilled variants - which run near CFG 1 - the uncond-based benefits shrink, so keep the window short and expect a modest effect. If a workflow's SLG settings were tuned on an older ComfyUI build, an update may shift results; re-check scale before blaming the model.