⛔ Generate Negative Prompt

This is the one node in the pack whose own description talks itself down: it's flagged EXPERIMENTAL, and the author says outright that "the model is quite weak and random though, so it doesn't work well. It mostly just generates random negative prompts trained on CivitAI negative prompts." That's an unusually honest label to put on your own node, and worth taking at face value before you plug it into anything you care about.

What it's trying to do

You give it a positive prompt, and it runs a small text-generation model to spit out a negative prompt that's supposed to pair with it - the idea being you skip typing "blurry, low quality, extra limbs" by hand every time. The controls are the standard knobs you'd see on any Hugging Face generate() call: num_beams for beam search width, temperature for randomness, top_k and top_p for how wide a pool of next-words the model samples from, and max_length capping how long the output can run.

The inputs and outputs that matter

• input_prompt - your positive prompt. This is what the model conditions its guess on.
• temperature and top_p - the two worth touching if the output feels stale or too random. Lower temperature = more predictable (and often more repetitive); higher = more varied (and more likely to go off the rails).
• blocked_words - a multiline list of words to exclude from the output, one per line. Useful damage control given the model's source material.
• max_length, num_beams, top_k - leave at defaults unless you're specifically chasing longer output or more deterministic phrasing.

Output is a single negative_prompt string, meant to wire straight into your negative CLIPTextEncode.

How to install it

Via ComfyUI Manager: search ComfyUI-mnemic-nodes, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/MNeMoNiCuZ/ComfyUI-mnemic-nodes

then restart. This node doesn't need the Groq API key setup the pack's LLM/VLM/audio nodes require - it runs its own small local model, not a hosted API.

Common issues & troubleshooting

The output doesn't make much sense, or seems disconnected from your prompt. That's the documented behavior, not a bug you're hitting - the author already flagged this as weak and random. Treat it as a starting point to edit by hand, not a finished negative prompt.

NSFW words show up unprompted. Also called out directly in the node's own description: it was trained on real Civitai negative-prompt text, which includes plenty of that vocabulary, and it doesn't filter it out on its own. Use blocked_words if that's a problem for your workflow.

It doesn't seem to change your output at all. Check what checkpoint and sampler settings you're actually running. Negative prompts only do anything through classifier-free guidance's unconditioned pass, which means they're inert at CFG 1 - and a lot of the fast, guidance-distilled models in circulation now (Flux Dev/Schnell-style, Turbo LoRA setups) run right at CFG 1 by design. If you're on one of those, no negative prompt generator, this one included, is going to have any effect - the field just isn't being used by the sampler at all.