AC_Negative_prompt提示词

Here's the thing nobody tells you about negative prompts in 2026: whether they do anything at all depends entirely on the model you're running. The ComfyUI-AC_FUNV2.0 pack's AC_Negative_prompt提示词 ships 56 curated negative prompts (反向提示词) - photography, landscape, interior, anime, and character-realistic variants - and they're textbook SD1.5/SDXL-era negatives: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, blurry. That's the classic list, and it's exactly the right tool for the models that were trained alongside it.

Mechanically, same as the rest of the pack: ModeList is a dropdown of the .txt files in the Negative folder, pick one, the node reads it, and the contents arrive on a single STRING output. Wire it into the negative CLIP Text Encode socket of your workflow, not the positive one.

The "if and only if" part, stated plainly. A negative prompt is not a filter and removes nothing - classifier-free guidance runs a second pass and steers away from what you wrote. That mechanism only kicks in when CFG is above 1. The KB's research on this is unambiguous: negative prompts are alive and working for the SDXL lineage (Illustrious, NoobAI, Pony, and friends at CFG 4–9), and they are dead weight on distilled models - Flux, Schnell, Turbo and friends run at CFG 1, where the negative pass is literally skipped for speed, and the text in your negative box does nothing at all. So this node is genuinely useful on an SDXL-finetune workflow and close to decorative on a modern distilled one.

That's not the author's fault - this pack's negative shelf is squarely aimed at the SDXL era it was written for, and on that stack these files are a nice time-saver: 56 ready-made negatives sorted by subject, so you don't retype the standard "bad hands, bad anatomy" list for the tenth time.

Two tips that make it better:

• The negatives are keyed to subject. A landscape negative is not the same as a character-realistic one - grab the one that matches your positive, don't just pick the first entry.
• If you're on a distilled model and still want quality insurance, the modern move (per the prompt-engineering research) is to restate constraints positively - "clean studio background, sharp focus, plain seamless backdrop" in the positive prompt beats any negative on CFG-1 models.

Install is the pack standard:

cd ComfyUI/custom_nodes
git clone https://github.com/A719689614/ComfyUI-AC_FUNV2.0

restart, then under ♑AC_FUNV2.0. No models, no dependencies beyond stock ComfyUI (skip the torch==2.11.0 line in requirements.txt if pip complains).

The honest verdict: if you're on SDXL or one of its finetunes, this is a perfectly good negative-prompt library and the folder-based extension point (drop your own .txt in Negative/, restart) lets you build on it. If your primary stack is Flux or any CFG-1 distilled model, keep the node around for the occasional SDXL workflow but don't expect it to improve your Flux outputs - nothing will, because on that stack the negative prompt isn't being ignored, it's not even being computed.