KSamplerWithNAG (Advanced)

This is KSampler (Advanced) with Normalized Attention Guidance folded in - same job as KSamplerWithNAG, but built for the multi-stage sampling patterns the Advanced node exists for: base-plus-refiner chains, splitting a run across two samplers, or resuming a partially-denoised latent. If your graph passes a latent between two KSampler (Advanced) nodes, replacing both with this is the direct swap, and it's the piece of this pack that actually gets a slightly better click-through than its siblings, which tracks - people searching for the Advanced variant by name usually already know exactly why they need it.

Why it's a separate node from KSamplerWithNAG

KSampler (Advanced) exists because a single denoise pass isn't always what you want: SDXL's base+refiner split needs to hand off mid-sequence, hires-fix workflows resample a subset of steps, and some pipelines deliberately stop before full denoise so a downstream node can take over. The extra widgets - start_at_step, end_at_step, return_with_leftover_noise - are what make that possible. NAG doesn't change any of that machinery; it just needs to ride along on whichever stage is doing the actual denoising, which is why this node exists instead of you bolting NAG onto only one half of a two-stage chain.

What NAG adds on top

Same mechanism as the rest of the pack: it extrapolates the model's attention away from a nag_negative conditioning you supply, which keeps a real negative-prompting effect alive even on guidance-distilled models pinned to CFG 1 - Flux, Flux Kontext, Chroma, HiDream, Wan, Hunyuan Video - where a normal negative prompt has no CFG pass left to act through.

Inputs and outputs that matter

Everything KSampler (Advanced) already has: model, add_noise (enum, enable/disable - not a boolean here, unlike the non-Advanced node), noise_seed, steps (default 20), cfg (default 8), sampler_name, scheduler, positive, negative, latent_image, start_at_step (default 0), end_at_step (default 10000), and return_with_leftover_noise (enum, disable/enable).

On top: nag_negative (the conditioning NAG steers away from - separate from negative), and the four shared dials - nag_scale (default 5, your day-to-day strength knob), nag_tau (default 2.5) and nag_alpha (default 0.25, tune these once per model and leave them), and nag_sigma_end (default 0, raise it to switch NAG off early once composition is set, for a real speed win with barely any quality cost).

One output: LATENT, same as the standard node - feeds VAE Decode if this is the last stage, or the next KSamplerWithNAG (Advanced) if you're chaining.

The one thing worth getting right in a chain

If you're splitting a run across two of these - say, steps 0–15 on one and 15–30 on the next - decide once whether nag_negative should stay identical across both stages or change. Most workflows keep it constant: NAG is steering the same unwanted content the whole way through, and there's no established practice around varying it mid-chain. Change one setting at a time here, same as with any multi-stage sampler; splitting a run introduces enough variables on its own without also inventing new NAG behavior between stages.

Installing it

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

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

then restart. No model downloads, nothing heavy to install - it works on tensors from a model you've already got loaded.

Common issues

Mixing this up with the non-Advanced node in a chain. If one stage of your split uses plain KSamplerWithNAG and the next uses the Advanced variant, start_at_step/end_at_step bookkeeping gets easy to lose track of - keep both stages on the same node type.

Post-update TypeErrors about an unexpected keyword argument. This pack's guider internals track ComfyUI core's sampler interface closely, and a core update has broken that interface before with a real reported inner_sample() got an unexpected keyword argument 'latent_shapes' failure. If sampling errors out right after a ComfyUI update, check for a newer ComfyUI-NAG release before assuming your graph broke.

Negative prompt doing nothing. Check nag_negative specifically - it's a separate slot from negative, and leaving it unwired or empty is the most common reason NAG appears to have no effect.