NAGGuider

Worth saying up front: this is the quietest node in the ComfyUI-NAG pack by a wide margin - barely any search traffic, and essentially no community discussion under its own name. That's not a red flag on the node itself; it's what you'd expect from a very specific swap. NAGGuider replaces BasicGuider, and BasicGuider is already the guider for the simplest, most stripped-down case in ComfyUI's custom-sampling graph: a single conditioning, no CFG, no negative at all. That's exactly the shape of Flux Dev and similar pure guidance-distilled models. If you're not already using BasicGuider, you probably don't need this node.

What it's for

BasicGuider exists because some models genuinely don't have a negative-prompt mechanism to plug into - Flux Dev is the canonical example, guidance-distilled so hard that there's no unconditioned pass anywhere in its native sampling, not just one turned off by a CFG-1 setting. Most of the CFG-1 conversation around NAG is really about models that do have a negative slot sitting there inert (Flux Kontext, SDXL turbo variants, Wan at CFG 1); BasicGuider's use case is a step further - no negative slot to begin with. NAGGuider adds one back, using the same attention-space extrapolation as the rest of this pack, without requiring the model to support CFG at all.

Inputs and outputs that matter

Notice the shape here is different from the other guider node in this pack, NAGCFGGuider: there's no separate positive/negative pair, because BasicGuider-style sampling never had one. Required inputs are model, conditioning (your one and only prompt - this is what BasicGuider calls conditioning, not positive), nag_negative (the thing NAG steers away from), and latent_image. Then the four shared NAG dials at their usual defaults: nag_scale (5), nag_tau (2.5), nag_alpha (0.25), and nag_sigma_end (0).

One output, GUIDER, which plugs into SamplerCustomAdvanced alongside a SAMPLER node (from KSamplerSelect) and SIGMAS (from a scheduler node) - the same modular custom-sampling graph NAGCFGGuider feeds into. If you don't already have that graph, this isn't your easiest way into NAG; KSamplerWithNAG gives you the same negative-prompting effect through the ordinary KSampler widgets you already know.

Installing it

Via ComfyUI Manager: search "ComfyUI-NAG", install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ChenDarYen/ComfyUI-NAG

then restart ComfyUI. No model downloads, no heavy dependency chain - NAG operates on the attention of a model you've already loaded.

Troubleshooting

There's genuinely thin public discussion of this specific node to draw from, so here's what's grounded rather than guessed:

If conditioning looks like it's doing nothing without NAG, that's expected - BasicGuider-style sampling was never going to give you negative control on its own. The whole reason to reach for NAGGuider instead of BasicGuider is that plain conditioning-only sampling has no negative mechanism whatsoever; NAG is what adds one.

nag_negative unwired is the most common way this appears broken. Since there's no negative field here to accidentally use instead, it's usually just left empty rather than mixed up with the wrong slot - check that it's actually connected to real conditioning, not left dangling.

Guider-internals errors after a ComfyUI update. This pack's guider classes are built against ComfyUI core's own sampler interface, and that interface has shifted before in ways that broke a sibling guider node in this same pack with a TypeError about an unexpected keyword argument. If NAGGuider starts erroring right after you update ComfyUI, look for a newer version of the pack before assuming your graph is at fault.