This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHubThe NAGuidance node applies Normalized Attention Guidance to a model. This technique enables the use of negative prompts with distilled or schnell models by modifying the model’s attention mechanism during the sampling process to steer the generation away from undesired concepts.
Inputs
| Parameter | Data Type | Required | Range | Description |
|---|---|---|---|---|
model | MODEL | Yes | - | The model to apply Normalized Attention Guidance to. |
nag_scale | FLOAT | Yes | 0.0 - 50.0 | The guidance scale factor. Higher values push the generation further from the negative prompt. (default: 5.0) |
nag_alpha | FLOAT | Yes | 0.0 - 1.0 | The blending factor for the normalized attention. A value of 1.0 fully replaces the original attention, while 0.0 has no effect. (default: 0.5) |
nag_tau | FLOAT | Yes | 1.0 - 10.0 | A scaling factor used to limit the normalization ratio. (default: 1.5) |
Outputs
| Output Name | Data Type | Description |
|---|---|---|
model | MODEL | The patched model with Normalized Attention Guidance enabled. |