NAG Paramters Setting

NAG is not this pack's invention, and that's exactly why this node is neat. It's a settings pass-through for Normalized Attention Guidance, a sampling method by Chen Dar Yen that got attention in 2025 as a way to use your negative prompt without paying the full classifier-free-guidance tax - pitched at roughly 2x speed on Wan. The SuperUltimate pack bundles the NAG sampler directly into the repo (see the nag/ folder, credited to ComfyUI-NAG), and this node is how you configure it.

The long-video and upscale nodes all sample through plain KSampler by default. Wire a nag_params output into their optional nag_params input and they switch to the bundled NAG sampler instead. Nothing else about your workflow changes - it's an opt-in swap, not a new pipeline.

The inputs

Four floats, and they mirror the upstream NAG parameters:

• nag_scale - default 5. The strength of the attention guidance. The scale of the correction NAG applies.
• nag_tau - default 2.5. A temperature-like parameter controlling how aggressively the guidance is normalized.
• nag_alpha - default 0.25. Blend between the guided and unguided paths.
• nag_sigma_end - default 0.0. Where in the noise schedule the guidance phases out (0 = runs through the end).

The defaults are the upstream values, so a node you've never touched is already a sane starting point. Output is nag_params (NAGParamtersSetting), wired to any of the three big nodes' nag_params sockets.

Should you use it?

Honest take: it's a "try it and see" knob, not a set-and-forget win. The pitch is speed - negative-prompt guidance without the full CFG cost - which matters a lot when every VACE tile or round is a long render. But the guidance behavior is different from CFG, and people either love it or find the results subtly off. If your VACE gens are already fast enough, leave the socket empty and don't think about it. If you're grinding through a long upscale or a 9-round video, this is the single cheapest speed experiment the pack offers - one node, four defaults, and you're measuring.

Install

Nothing extra. It ships inside the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/bbaudio-2025/ComfyUI-SuperUltimateVaceTools

Restart ComfyUI (or search "SuperUltimateVaceTools" in Manager). Unlike the pack's color-matching feature, NAG needs no pip dependencies - the sampler is bundled.

Gotchas

• It's per-socket, not global. Each of the three big nodes takes its own nag_params, so you can run NAG on the upscale but keep CFG on the long-video - useful when you want to isolate which stage benefits.
• The values are unitless tuning knobs with no universal "best." Change one at a time; nag_scale is the one people usually move first.
• nag_tau clamps to the 1–10 range and nag_alpha to 0–1; the node will fight you politely if you try nonsense, so the UI limits are the real bounds.
• If a run produces odd artifacts after an update, the node's cached inputs can lag the new code - the pack-wide fix applies: right-click → Fix node (recreate).

It's a small node, and it looks like a checkbox, but it's the pack's only real "make VACE faster without touching the model" lever. Worth one render to find out if your taste agrees with NAG's.