Negative Prompt Generator

Everyone agonizes over the positive prompt and then types "blurry, bad quality" into the negative box out of habit. Negative Prompt Generator is the node that fixes that: give it your positive prompt and a style, and it writes a proper comma-separated negative list for you - locally, via the same Ollama setup as the rest of this pack.

How it works

The node takes your positive prompt, the selected style, and a per-style hint list, and builds an instruction for the LLM: "Given this positive prompt and style, generate a concise negative prompt that lists what should be avoided." The style hints matter - they're what make the output style-aware rather than a generic wall of words. Cinematic gets "shaky cam, lens flare abuse, bad CGI"; photorealistic gets "painting, illustration, oversaturated"; video_wan gets temporal-coherence warnings like "static image, still frame, jump cut". It then cleans the model's response and hands you negative_prompt.

Two details that show the author thought about failure modes:

• The default temperature is 0.3, deliberately low - this is a conservative task, and you want the model listing problems, not being creative about them.
• If the LLM returns nothing usable, the node falls back to the static style hints instead of handing you an empty string. You always get something to paste in.

Inputs and outputs

• prompt (required) - your positive prompt. Feed it the generator's output for a matching pair.
• style - nine presets, same list as the rest of the pack.
• model - default qwen3:8b; smaller/faster models work fine here since the task is short.
• temperature (0.3), top_p (0.9), timeout (default 60s - this task is quick, so it's shorter than the generator's).

Single output: negative_prompt (STRING) → wire into your CLIP Text Encode (Negative) or Show Text.

Install and the one big caveat

Same pack, same install:

cd ComfyUI/custom_nodes
git clone https://github.com/Limbicnation/ComfyUI-PromptGenerator.git
cd ComfyUI-PromptGenerator
pip install -r requirements.txt

Prereqs: Ollama running and ollama pull qwen3:8b. Find it under text/generation → Negative Prompt Generator.

Now the caveat, and it's worth reading before you build a workflow around this: a negative prompt only does something if your CFG is above 1. On guidance-distilled models - Flux Dev and Schnell at their default settings, Z-Image Turbo, most of the "fast" releases - ComfyUI runs at CFG 1 and skips the unconditional pass entirely. The negative box is inert there no matter how good the list is. So this node earns its keep on SDXL-lineage models (SD 1.5, SDXL, Illustrious, Pony) and on LLM-encoded models that genuinely support negatives (Z-Image Base being the notable one that made negative support a selling point). If your whole workflow is Flux Dev at CFG 1, skip this node - you're asking an LLM to write text the sampler will ignore.

One more realistic gotcha: the node can only be as style-aware as your positive prompt is. If you feed it a vague prompt, it'll generate a vague, generic negative - the static-hints fallback is fine as a floor, but the whole point is to start from a good positive. Generate that first, then run this on the result.