ComfyUI-NAG

Implementation of Normalized Attention Guidance: Universal Negative Guidance for Diffusion Models for ComfyUI.

NAG restores effective negative prompting in few-step diffusion models, and complements CFG in multi-step sampling for improved quality and control.

Paper: https://arxiv.org/abs/2505.21179

Code: https://github.com/ChenDarYen/Normalized-Attention-Guidance

Wan2.1 Demo: https://huggingface.co/spaces/ChenDY/NAG_wan2-1-fast

LTX Video Demo: https://huggingface.co/spaces/ChenDY/NAG_ltx-video-distilled

Flux-Dev Demo: https://huggingface.co/spaces/ChenDY/NAG_FLUX.1-dev

News

2025-07-06: Add three new nodes:

• KSamplerWithNAG (Advanced) as a drop-in replacement for KSampler (Advanced).
• SamplerCustomWithNAG for SamplerCustom.
• NAGGuider for BasicGuider.

2025-07-02: HiDream is now supported!

2025-07-02: Add support for TeaCache and WaveSpeed to accelerate NAG sampling!

2025-06-30: Fix a major bug affecting Flux, Flux Kontext and Chroma, resulting in degraded guidance. Please update your NAG node!

2025-06-29: Add compile model support. You can now use compile model nodes like TorchCompileModel to speed up NAG sampling!

2025-06-28: Flux Kontext is now supported. Check out the workflow!

2025-06-26: Hunyuan video is now supported!

2025-06-25: Wan video generation is now supported (GGUF compatible)! Try it out with the new workflow!

Nodes

• KSamplerWithNAG, KSamplerWithNAG (Advanced), SamplerCustomWithNAG
• BasicGuider, NAGCFGGuider

Usage

To use NAG, simply replace

• KSampler with KSamplerWithNAG.
• KSamplerWithNAG (Advanced) with KSampler (Advanced).
• SamplerCustomWithNAG with SamplerCustom.
• NAGGuider with BasicGuider.
• CFGGuider with NAGCFGGuider.

We currently support Flux, Flux Kontext, Wan, Vace Wan, Hunyuan Video, Choroma, SD3.5, SDXL and SD.

Example workflows are available in the ./workflows directory!

Key Inputs

When working with a new model, it's recommended to first find a good combination of nag_tau and nag_alpha, which ensures that the negative guidance is effective without introducing artifacts.

Once you're satisfied, keep nag_tau and nag_alpha fixed and tune only nag_scale in most cases to control the strength of guidance.

Using nag_sigma_end to reduce computation without much quality drop.

For flow-based models like Flux, nag_sigma_end = 0.75 achieves near-identical results with significantly improved speed. For diffusion-based SDXL, a good default is nag_sigma_end = 4.

• nag_scale: The scale for attention feature extrapolation. Higher values result in stronger negative guidance.
• nag_tau: The normalisation threshold. Higher values result in stronger negative guidance.
• nag_alpha: Blending factor between original and extrapolated attention. Higher values result in stronger negative guidance.
• nag_sigma_end: NAG will be active only until nag_sigma_end.

Rule of Thumb

• For image-reference tasks (e.g., Image2Video), use lower nag_tau and nag_alpha to preserve the reference content more faithfully.
• For models that require more sampling steps and higher CFG, also prefer lower nag_tau and nag_alpha.
• For few-step models, you can use higher nag_tau and nag_alpha to have stronger negative guidance.