Maintenance fork of ChenDarYen/ComfyUI-NAG by Dar-Yen Chen. All credit for the original implementation goes to the upstream author; this fork only adds compatibility patches and packaging.

See NOTICE.md for the change list and LICENSE for the unmodified MIT license terms.

Install

Comfy Registry: comfyui-nag-delcado (publisher @delcado)
From source: git clone https://github.com/Delcado19/ComfyUI-NAG.git into your ComfyUI/custom_nodes/ directory.

What this fork fixes vs. upstream

chroma/layers.py None-stub crash on ComfyUI v0.21+ Upstream chroma/layers.py:10 does class NAGDoubleStreamBlock(DoubleStreamBlock): after importing DoubleStreamBlock / SingleStreamBlock from comfy.ldm.chroma.layers. Recent ComfyUI versions deprecate those symbols and set them to None (the real classes moved to comfy.ldm.flux.layers), so the import fails with TypeError: NoneType takes no arguments. This fork detects both the None stub and a missing legacy import path, then falls back to the new location. Closes the same root cause as upstream issues #79, #60, #54, #53, #55.
Flux / Flux Kontext missing flipped_img_txt attribute Recent ComfyUI Flux blocks no longer expose flipped_img_txt. This fork treats a missing attribute as ComfyUI's current txt, img attention ordering, avoiding an AttributeError during Flux and Flux Kontext sampling.
Flux2 / Flux.2 klein position-ID compatibility Flux2 models can use a different positional-ID axis count than older Flux models. This fork derives the NAG Flux text-ID shape from the loaded model instead of assuming three axes, matching ComfyUI's current Flux behavior.
Flux2 / Flux.2 klein gated MLP compatibility Current ComfyUI Flux2 / Flux.2 klein variants can use a gated single-stream MLP layout. This fork keeps those core single-stream blocks untouched for the gated or four-axis-ID variants, passes the combined positive/negative RoPE data through to the core blocks, and applies NAG only to the double-stream path to avoid the shape regressions that were crashing sampling. The combined RoPE is also left-padded along the txt axis to the shared padded txt length when the positive and negative contexts differ in token count, fixing RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 6052 but got size 4388 on Flux.2 klein 9B NAG workflows where the positive and uncond contexts have different lengths.
SamplerCustomAdvanced empty-token / batch compatibility NAG's direct nag_negative conditioning path now repeats the NAG negative tensor to the latent batch size using ComfyUI's batching helper and replaces zero-token text conditioning with a minimal zero-token fallback. This avoids shape errors such as shape '[2, 0, 4096]' is invalid for input of size 4096 when custom samplers run batched Flux-style workflows with empty or very short NAG negative prompts. Flux2 paths that omit pooled CLIP output for nag_negative now either use a neutral vector fallback or skip vector conditioning when the model has no vector branch, avoiding 'NoneType' object has no attribute 'to' crashes in SamplerCustomAdvanced. Flux2 global-modulation blocks are handled through ComfyUI's tuple-based modulation path, so NAG no longer expects removed img_mod / txt_mod block attributes.
Flux2 / current ComfyUI forward compatibility The Flux NAG wrapper now preserves the loaded model's patch_size when reshaping outputs, so Flux2 128-channel latents are not collapsed into the older Flux 32-channel layout. It also passes optional ComfyUI Flux forward parameters by keyword, keeping the NAG path aligned with current ComfyUI signatures that added positional arguments after attn_mask.
README Usage section direction — closes upstream #39.
Packaging / Registry — pyproject.toml plus GitHub Actions for smoke-import test and Comfy-Registry publish.

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

comfyui-nag

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

We currently support Flux, Flux Kontext, Flux2 / Flux.2 klein, Wan, Vace Wan, Hunyuan Video, HiDream, Chroma, 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.

ComfyUI validates these ranges before running the prompt. Values outside the allowed range are rejected rather than clamped automatically.

| Input | Minimum | Maximum | Default | Notes | |-------|---------|---------|---------|-------| | nag_scale | 0.0 | 100.0 | 5.0 | NAG guidance is applied only when nag_scale > 1.0. | | nag_tau | 1.0 | 10.0 | 2.5 | Normalization threshold. | | nag_alpha | 0.0 | 1.0 | 0.25 | Attention blend factor. | | nag_sigma_end | 0.0 | 20.0 | 0.0 | For Flux-style models, 0.75 is a useful speed/quality default. |

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.