ComfyUI Node
RT Qwen-Image-2.1 NAG
Normalized Attention Guidance (NAG) for Qwen-Image-2.1, Wan, and DiT models. Injects negative prompt guidance into CFG = 1.0.
RT Qwen-Image-2.1 NAG
- model
- conditioning
- model
◄nag_scale11.0►
◄nag_alpha0.25►
◄nag_tau2.5►
◄input_typedefault►
CategoryRareTutor/Sampling
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | Connect input from model or LoRA loader. | |
| conditioning | CONDITIONING | Negative prompt conditioning to guide against (NAG reference direction). | |
| nag_scale | FLOAT | 11.00–100 | Guidance strength against the negative conditioning. Default is 11.0. |
| nag_alpha | FLOAT | 0.250–1 | Blend coefficient between NAG-guided attention and positive attention. Default is 0.25. |
| nag_tau | FLOAT | 2.50–10 | L1 norm clipping threshold to prevent color burn and saturation. Default is 2.5. |
| input_type | COMBO | default | default: standard pair mode; batch: single conditioning mode for CFG=1 speed models. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |