Negative PQ / 负面队列

Despite the name, Negative PQ is not really a queue. The other nodes in the ComfyUI_PromptQueue pack step through a list of prompts over time; this one is a negative-prompt library with templates - you type your negatives (or pick a preset), it encodes them into a negative CONDITIONING, and you wire that into the KSampler's negative input. There's no indexing, no counter, no file loading. It's the simplest node in the pack and honestly the most useful if you juggle several checkpoints, because each model family wants a different blocklist.

How it works

Feed it a clip, give it text, get one negative CONDITIONING out. The template system is the interesting part: the 模板 field defaults to {p}, which is replaced by each line of 多行文本. So with {p}, bad hands, extra fingers and lines of blurry / lowres, you get two negatives to pick from - but since there's a single CONDITIONING output, effectively you keep one assembled negative at a time. Empty input encodes a blank string, so leaving it blank is a safe no-op rather than a crash.

The presets are the point

The 预设模板 dropdown ships with four real-world negative stacks baked in: PonyNeg (with score_6, score_5... quality tags and simplified, abstract, anime, drawing...), NovaNeg, NOVAFurryNeg, and NoobAINeg. These match the positive presets the other nodes ship - the author clearly built this for the Pony/Illustrious/NoobAI anime-SDXL crowd, and they're solid starting points for that family. Each one includes {p} so your own additions still slot in.

You can also save and delete your own via 保存预设 and 删除预设 (the selected preset's name is the key, template content is the value, stored in _prompt_queue_state/presets.json). Standard pack gotcha applies: after saving or deleting, run the workflow once and refresh the browser before the dropdown updates.

The caveat that actually burns people

A negative prompt only steers generation when the sampler is running real classifier-free guidance - i.e. CFG above 1. On guidance-distilled checkpoints like Flux, Turbo models, or Z-Image Turbo at their designed CFG of 1, the negative input is silently ignored (ComfyUI skips the unconditional pass entirely), and this node's work does nothing even though it looks wired correctly. If your output is full of "bad anatomy" despite this node, first check you're not on a distilled model at CFG 1. On SDXL-lineage models - Pony, Illustrious, NoobAI, which is what the bundled presets target - it works exactly as intended.

Install

No dependencies, no model files. The usual:

cd ComfyUI/custom_nodes
git clone https://github.com/cyan9977/ComfyUI_PromptQueue.git

Restart and find Negative PQ / 负面队列 under the PromptQueue category. Parameter labels are Chinese-only (多行文本 = multiline text, 预设模板 = preset template, 模板 = template, 保存预设 = save preset, 删除预设 = delete preset). It encodes only the one output, so drop it next to your KSampler's negative socket and let the presets do the arguing about which blocklist your checkpoint wants.