NAI 负面提示词

The negative prompt has a worse reputation than it deserves in 2026. On the guidance-distilled flow models everyone's excited about, the negative field is genuinely inert - but NovelAI's models are the older tag-based SDXL lineage where a real negative prompt still steers the output. This node is how you supply one in the NovelAI pack, and it exists mainly to stop you from plugging it in backwards.

What it is

NAINegativePromptV2 is a multiline text field, default empty, that outputs a typed NAI_NEGATIVE_PROMPT port. Type the things you want to suppress and wire the output into the negative input on the sampler.

The typed port is the whole point. In this pack the positive and negative prompt ports are distinct types, so ComfyUI will refuse to connect your negative prompt to the positive sampler input. That's a small mercy you only appreciate after watching someone burn a hundred generations because a generic STRING port accepted anything. You can also route it through the negative join and processor variants the pack ships if you're composing it from parts.

What to actually put in it

NovelAI's default negative - the one the tiling nodes in this pack ship with - is a good starting point:

blurry, lowres, upscaled, worst quality, bad quality, jpeg artifacts, very displeasing, chromatic aberration, color noise, colored speckles, multiple views, logo

That covers the classic SDXL-era failure modes: compression artifacts, chroma noise, and "please don't look upscaled." Since NovelAI is a booru-tag model, tag vocabulary works here just as well as in the positive box - bad hands, extra fingers, watermark, signature are all real tokens it understands. Keep the list short and specific. A wall of forty tags mostly dilutes the ones that matter.

One thing worth knowing: an empty negative prompt is a legitimate choice. If your positive prompt is doing the work and the negative is fighting it, leaving it blank can give you cleaner results. This node defaults to empty for a reason.

Install and setup

It ships in the ComfyUI-NovelAI-GENYTOOLS pack. Manager: search "ComfyUI-NovelAI-GENYTOOLS". Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/XTOGENY/ComfyUI-NovelAI-GENYTOOLS.git
pip install -r ComfyUI-NovelAI-GENYTOOLS/requirements.txt

Restart, then add your NovelAI Persistent API Token in ComfyUI settings under NovelAI 2.0: Persistent API Token. No models to download - inference runs on NovelAI's servers. The node itself spends nothing; the sampler downstream is where Anlas goes.

Gotchas

• You can't plug this into the positive port, and that's a feature, not a bug. If you need a negative prompt as text for something else, right-click and convert the widget, or just read the summary string off the processor.
• The UI labels are Chinese ("NAI 负面提示词"). The port types don't lie even when the language does.
• If your negative prompt seems to do nothing, check you actually wired it into the sampler's negative input and not the positive one - the typed ports catch the common mistake, but leaving it unwired on a fresh sampler is still the usual cause.