Skip to main content
Enhances guidance towards detailed structure by using another set of CFG negative with skipped layers. This generic version of SkipLayerGuidance can be used on every DiT model and is inspired by Perturbed Attention Guidance. The original experimental implementation was created for SD3.

Inputs

ParameterDescriptionData TypeRequiredRange
modelThe model to apply skip layer guidance toMODELYes-
double_layersComma-separated layer numbers for double blocks to skip (default: “7, 8, 9”)STRINGYes-
single_layersComma-separated layer numbers for single blocks to skip (default: “7, 8, 9”)STRINGYes-
scaleGuidance scale factor (default: 3.0)FLOATYes0.0 - 10.0
start_percentStarting percentage for guidance application (default: 0.01)FLOATYes0.0 - 1.0
end_percentEnding percentage for guidance application (default: 0.15)FLOATYes0.0 - 1.0
rescaling_scaleRescaling scale factor to adjust the output magnitude (default: 0.0, meaning no rescaling)FLOATYes0.0 - 10.0
Note: If both double_layers and single_layers are empty (contain no layer numbers), the node returns the original model without applying any guidance.

Outputs

Output NameDescriptionData Type
modelThe modified model with skip layer guidance appliedMODEL
This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): 35ca97f87913d0efcbb502ca040a8f357da112b202184d8786f63acd76809808