Native NAG node (Normalized Attention Guidance) not applying negative prompt despite double inference time

[fappaz]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

When using the native Normalized Attention Guidance node with default values (scale: 5, alpha: 0.5, tau: 1.5), the negative prompt should affect the output.

Actual Behavior

The output is exactly the same as if there was no negative prompt, despite the inference taking twice as long.

Tested with both Flux.2 Klein 4B (example below) and 9B, same issue.

Negative prompt with NAG was: trees, grass

Steps to Reproduce

• Generate an image with this workflow for Flux.2 Klein 4B (or any txt2img workflow), which contains native NAG with default parameters and a negative prompt:

{"id":"92112d97-bb64-4b44-86f2-ea5691ef8f6e","revision":0,"last_node_id":367,"last_link_id":773,"nodes":[{"id":341,"type":"CLIPLoader","pos":[-680,360],"size":[340,106],"flags":{},"order":0,"mode":0,"inputs":[],"outputs":[{"name":"CLIP","type":"CLIP","links":[770,771]}],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"CLIPLoader"},"widgets_values":["qwen3_4b_fp8_scaled.safetensors","flux2","default"],"color":"#432","bgcolor":"#653"},{"id":343,"type":"VAELoader","pos":[-680,580],"size":[340,58],"flags":{},"order":1,"mode":0,"inputs":[],"outputs":[{"name":"VAE","type":"VAE","links":[773]}],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"VAELoader"},"widgets_values":["flux2-vae.safetensors"],"color":"#322","bgcolor":"#533"},{"id":8,"type":"VAEDecode","pos":[980,560],"size":[140,46],"flags":{},"order":8,"mode":0,"inputs":[{"name":"samples","type":"LATENT","link":750},{"name":"vae","type":"VAE","link":773}],"outputs":[{"name":"IMAGE","type":"IMAGE","slot_index":0,"links":[709]}],"properties":{"cnr_id":"comfy-core","ver":"0.3.75","Node name for S&R":"VAEDecode"},"widgets_values":[],"color":"#323","bgcolor":"#535"},{"id":344,"type":"SaveImage","pos":[1240,0],"size":[520,600],"flags":{},"order":9,"mode":0,"inputs":[{"name":"images","type":"IMAGE","link":709}],"outputs":[],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"SaveImage"},"widgets_values":["Simplesmente IA/flux-2-klein"],"color":"#2a363b","bgcolor":"#3f5159"},{"id":340,"type":"UNETLoader","pos":[-680,20],"size":[340,82],"flags":{},"order":2,"mode":0,"inputs":[],"outputs":[{"name":"MODEL","type":"MODEL","links":[769]}],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"UNETLoader"},"widgets_values":["flux-2-klein-4b-fp8.safetensors","default"],"color":"#223","bgcolor":"#335"},{"id":330,"type":"EmptyImage","pos":[100,-180],"size":[300,130],"flags":{},"order":3,"mode":0,"inputs":[],"outputs":[{"name":"IMAGE","type":"IMAGE","links":[759]}],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"EmptyImage"},"widgets_values":[896,1152,1,0],"color":"#2a363b","bgcolor":"#3f5159"},{"id":6,"type":"CLIPTextEncode","pos":[-140,180],"size":[540,220],"flags":{"collapsed":false},"order":4,"mode":0,"inputs":[{"name":"clip","type":"CLIP","link":770}],"outputs":[{"name":"CONDITIONING","type":"CONDITIONING","slot_index":0,"links":[756]}],"title":"CLIP Text Encode (Positive Prompt)","properties":{"cnr_id":"comfy-core","ver":"0.3.75","Node name for S&R":"CLIPTextEncode"},"widgets_values":["photo of a puppy walking outdoor"],"color":"#232","bgcolor":"#353"},{"id":359,"type":"1eb7ed83-dad4-4ff5-90a5-f7ce2800182f","pos":[620,0],"size":[280,520],"flags":{},"order":7,"mode":0,"inputs":[{"name":"image","type":"IMAGE","link":759},{"name":"model","type":"MODEL","link":763},{"name":"positive","type":"CONDITIONING","link":756},{"name":"negative","type":"CONDITIONING","link":766}],"outputs":[{"name":"output","type":"LATENT","links":[750]}],"properties":{"proxyWidgets":[["357","noise_seed"],["357","control_after_generate"],["355","steps"],["354","cfg"],["356","sampler_name"],["353","batch_size"]],"cnr_id":"comfy-core","ver":"0.13.0"},"widgets_values":[],"color":"#323","bgcolor":"#535"},{"id":365,"type":"NAGuidance","pos":[100,20],"size":[300,106],"flags":{},"order":6,"mode":0,"inputs":[{"name":"model","type":"MODEL","link":769}],"outputs":[{"name":"MODEL","type":"MODEL","links":[763]}],"properties":{"cnr_id":"comfy-core","ver":"0.15.1","Node name for S&R":"NAGuidance"},"widgets_values":[5,0.5,1.5],"color":"#223","bgcolor":"#335"},{"id":366,"type":"CLIPTextEncode","pos":[-140,460],"size":[540,100],"flags":{"collapsed":false},"order":5,"mode":0,"inputs":[{"name":"clip","type":"CLIP","link":771}],"outputs":[{"name":"CONDITIONING","type":"CONDITIONING","slot_index":0,"links":[766]}],"title":"CLIP Text Encode (Positive Prompt)","properties":{"cnr_id":"comfy-core","ver":"0.3.75","Node name for S&R":"CLIPTextEncode"},"widgets_values":["trees, grass"],"color":"#322","bgcolor":"#533"}],"links":[[709,8,0,344,0,"IMAGE"],[750,359,0,8,0,"LATENT"],[756,6,0,359,2,"CONDITIONING"],[759,330,0,359,0,"IMAGE"],[763,365,0,359,1,"MODEL"],[766,366,0,359,3,"CONDITIONING"],[769,340,0,365,0,"MODEL"],[770,341,0,6,0,"CLIP"],[771,341,0,366,0,"CLIP"],[773,343,0,8,1,"VAE"]],"groups":[],"definitions":{"subgraphs":[{"id":"1eb7ed83-dad4-4ff5-90a5-f7ce2800182f","version":1,"state":{"lastGroupId":0,"lastNodeId":367,"lastLinkId":773,"lastRerouteId":0},"revision":0,"config":{},"name":"Klein Sampler","inputNode":{"id":-10,"bounding":[500,259,120,120]},"outputNode":{"id":-20,"bounding":[1560,280,120,60]},"inputs":[{"id":"6163fb6b-dddc-46e5-83ed-cdc0112388d0","name":"image","type":"IMAGE","linkIds":[685],"localized_name":"image","pos":[600,279]},{"id":"37182bf5-0ce5-4d82-919d-dcaf34a40629","name":"model","type":"MODEL","linkIds":[742],"localized_name":"model","pos":[600,299]},{"id":"f2cf81f5-c30c-4390-9739-418cb7fb7486","name":"positive","type":"CONDITIONING","linkIds":[743],"localized_name":"positive","pos":[600,319]},{"id":"72d41ddc-be7c-4bf8-a75e-6d64af447603","name":"negative","type":"CONDITIONING","linkIds":[744],"localized_name":"negative","pos":[600,339]}],"outputs":[{"id":"ee5ae631-bb73-4f5c-aa81-8e46c5e801f2","name":"output","type":"LATENT","linkIds":[745],"localized_name":"output","pos":[1580,300]}],"widgets":[],"nodes":[{"id":354,"type":"CFGGuider","pos":[980,-20],"size":[220,100],"flags":{},"order":4,"mode":0,"inputs":[{"localized_name":"model","name":"model","type":"MODEL","link":742},{"localized_name":"positive","name":"positive","type":"CONDITIONING","link":743},{"localized_name":"negative","name":"negative","type":"CONDITIONING","link":744}],"outputs":[{"localized_name":"GUIDER","name":"GUIDER","type":"GUIDER","links":[734]}],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"CFGGuider"},"widgets_values":[1]},{"id":357,"type":"RandomNoise","pos":[980,-160],"size":[210,82],"flags":{},"order":0,"mode":0,"inputs":[],"outputs":[{"localized_name":"NOISE","name":"NOISE","type":"NOISE","links":[733]}],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"RandomNoise"},"widgets_values":[98491897113364,"fixed"]},{"id":353,"type":"EmptyFlux2LatentImage","pos":[980,320],"size":[220,106],"flags":{},"order":3,"mode":0,"inputs":[{"localized_name":"width","name":"width","type":"INT","widget":{"name":"width"},"link":738},{"localized_name":"height","name":"height","type":"INT","widget":{"name":"height"},"link":740}],"outputs":[{"localized_name":"LATENT","name":"LATENT","type":"LATENT","links":[737]}],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"EmptyFlux2LatentImage"},"widgets_values":[896,1152,1],"color":"#323","bgcolor":"#535"},{"id":355,"type":"Flux2Scheduler","pos":[980,480],"size":[220,106],"flags":{},"order":5,"mode":0,"inputs":[{"localized_name":"width","name":"width","type":"INT","widget":{"name":"width"},"link":739},{"localized_name":"height","name":"height","type":"INT","widget":{"name":"height"},"link":741}],"outputs":[{"localized_name":"SIGMAS","name":"SIGMAS","type":"SIGMAS","links":[736]}],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"Flux2Scheduler"},"widgets_values":[4,896,1152]},{"id":326,"type":"GetImageSize","pos":[680,460],"size":[160,124],"flags":{},"order":2,"mode":0,"inputs":[{"localized_name":"image","name":"image","type":"IMAGE","link":685}],"outputs":[{"localized_name":"width","name":"width","type":"INT","links":[738,739]},{"localized_name":"height","name":"height","type":"INT","links":[740,741]},{"localized_name":"batch_size","name":"batch_size","type":"INT","links":null}],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"GetImageSize"},"widgets_values":[],"color":"#2a363b","bgcolor":"#3f5159"},{"id":358,"type":"SamplerCustomAdvanced","pos":[1280,40],"size":[212.3638671875,326],"flags":{},"order":6,"mode":0,"inputs":[{"localized_name":"noise","name":"noise","type":"NOISE","link":733},{"localized_name":"guider","name":"guider","type":"GUIDER","link":734},{"localized_name":"sampler","name":"sampler","type":"SAMPLER","link":735},{"localized_name":"sigmas","name":"sigmas","type":"SIGMAS","link":736},{"localized_name":"latent_image","name":"latent_image","type":"LATENT","link":737}],"outputs":[{"localized_name":"output","name":"output","type":"LATENT","links":[745]},{"localized_name":"denoised_output","name":"denoised_output","type":"LATENT","links":[]}],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"SamplerCustomAdvanced"},"widgets_values":[],"color":"#323","bgcolor":"#535"},{"id":356,"type":"KSamplerSelect","pos":[980,140],"size":[220,58],"flags":{},"order":1,"mode":0,"inputs":[],"outputs":[{"localized_name":"SAMPLER","name":"SAMPLER","type":"SAMPLER","links":[735]}],"properties":{"cnr_id":"comfy-core","ver":"0.9.2","Node name for S&R":"KSamplerSelect"},"widgets_values":["euler"]}],"groups":[],"links":[{"id":738,"origin_id":326,"origin_slot":0,"target_id":353,"target_slot":0,"type":"INT"},{"id":740,"origin_id":326,"origin_slot":1,"target_id":353,"target_slot":1,"type":"INT"},{"id":739,"origin_id":326,"origin_slot":0,"target_id":355,"target_slot":0,"type":"INT"},{"id":741,"origin_id":326,"origin_slot":1,"target_id":355,"target_slot":1,"type":"INT"},{"id":733,"origin_id":357,"origin_slot":0,"target_id":358,"target_slot":0,"type":"NOISE"},{"id":734,"origin_id":354,"origin_slot":0,"target_id":358,"target_slot":1,"type":"GUIDER"},{"id":735,"origin_id":356,"origin_slot":0,"target_id":358,"target_slot":2,"type":"SAMPLER"},{"id":736,"origin_id":355,"origin_slot":0,"target_id":358,"target_slot":3,"type":"SIGMAS"},{"id":737,"origin_id":353,"origin_slot":0,"target_id":358,"target_slot":4,"type":"LATENT"},{"id":685,"origin_id":-10,"origin_slot":0,"target_id":326,"target_slot":0,"type":"IMAGE"},{"id":742,"origin_id":-10,"origin_slot":1,"target_id":354,"target_slot":0,"type":"MODEL"},{"id":743,"origin_id":-10,"origin_slot":2,"target_id":354,"target_slot":1,"type":"CONDITIONING"},{"id":744,"origin_id":-10,"origin_slot":3,"target_id":354,"target_slot":2,"type":"CONDITIONING"},{"id":745,"origin_id":358,"origin_slot":0,"target_id":-20,"target_slot":0,"type":"LATENT"}],"extra":{"workflowRendererVersion":"LG"}}]},"config":{},"extra":{"frontendVersion":"1.41.7","VHS_latentpreview":false,"VHS_latentpreviewrate":0,"VHS_MetadataImage":true,"VHS_KeepIntermediate":true,"workflowRendererVersion":"LG","ds":{"scale":0.7682123190779837,"offset":[665.55660119324,361.0221052620578]}},"version":0.4}

• Disable the NAG node, then run the workflow again.

Notice how the output is the same, despite the it/s taking longer when NAG is enabled.

Debug Logs

When NAG is enabled:

got prompt
Requested to load Flux2TEModel_
Unloaded partially: 891.00 MB freed, 729.04 MB remains loaded, 405.00 MB buffer reserved, lowvram patches: 0
loaded completely; 4717.33 MB usable, 4207.26 MB loaded, full load: True
Requested to load Flux2
loaded completely; 5246.23 MB usable, 3882.02 MB loaded, full load: True
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:10<00:00,  2.60s/it]
Requested to load AutoencoderKL
Unloaded partially: 2262.02 MB freed, 1620.04 MB remains loaded, 324.00 MB buffer reserved, lowvram patches: 0
loaded completely; 517.90 MB usable, 160.31 MB loaded, full load: True
Prompt executed in 23.54 seconds

When NAG is disabled, keeping same workflow/seed/prompt:

got prompt
Prompt executed in 0.01 seconds
got prompt
loaded completely; 5246.23 MB usable, 3882.02 MB loaded, full load: True
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:04<00:00,  1.12s/it]
Requested to load AutoencoderKL
Unloaded partially: 2262.02 MB freed, 1620.04 MB remains loaded, 324.00 MB buffer reserved, lowvram patches: 0
loaded completely; 517.90 MB usable, 160.31 MB loaded, full load: True
Prompt executed in 10.04 seconds

Other

Specs

ComfyUI: 0.15.1
ComfyUI_frontend: v1.41.7
OS: win32
Python Version: 3.11.6 (tags/v3.11.6:8b6ee5b, Oct 2 2023, 14:57:12) [MSC v.1935 64 bit (AMD64)]
Embedded Python: true
Pytorch Version: 2.9.0+cu128
Arguments: ComfyUI\main.py --windows-standalone-build --disable-all-custom-nodes
RAM Total: 31.36 GB
Device Name: cuda:0 NVIDIA GeForce RTX 3080 Laptop GPU : cudaMallocAsync
VRAM Total: 8 GB

[jovan2009]

I also noticed a while ago the doubling of the inference time when I tried using this node. Since it was a Wan 2.2 workflow and I wasn't sure it is supposed to work with Wan models I didn't bother investigating further, I continued to use Kijai NAG node and I left it at that. I didn't realize that it is not applying negative prompt, I didn't do a on / off comparison.

[GlamoramaAttack]

When using the native Normalized Attention Guidance node with default values (scale: 5, alpha: 0.5, tau: 1.5), the negative prompt should affect the output.

A NAG node that supports Flux 2 klein is this one: https://github.com/BigStationW/ComfyUI-NAG - only that one (none of the other two that have the same repo name, ComfyUI-NAG). But it's slow and the effect is debatable ( https://www.reddit.com/r/StableDiffusion/comments/1qmhfkz/i_implemented_nag_normalized_attention_guidance/ ). It works with objects but less so with humans. And for objects prompts can be similar effective, with faster generation time.

Since it was a Wan 2.2 workflow and I wasn't sure it is supposed to work with Wan models I didn't bother investigating further, I continued to use Kijai NAG node and I left it at that.

KJ's NAG node is only for LTX-2.

[fappaz]

When using the native Normalized Attention Guidance node with default values (scale: 5, alpha: 0.5, tau: 1.5), the negative prompt should affect the output.

A NAG node that supports Flux 2 klein is this one: https://github.com/BigStationW/ComfyUI-NAG - only that one (none of the other two that have the same repo name, ComfyUI-NAG). But it's slow and the effect is debatable ( https://www.reddit.com/r/StableDiffusion/comments/1qmhfkz/i_implemented_nag_normalized_attention_guidance/ ). It works with objects but less so with humans. And for objects prompts can be similar effective, with faster generation time.

Thanks for sharing. I'm aware of the 3 custom nodes (totally agree with your comments), but wanted to check the native node, since Comfy's PR claims that it should work with Flux.2 as well as many other models:

Use the Normalized Attention Guidance node.
Flux, Flux2, Klein, Chroma, Chroma radiance, Hunyuan Video, etc..

Perhaps we're just using it wrong, so might be just a case of providing documentation.