Prompt All in One · Negative

Negatives get a bad editor everywhere, and this fixes that

"Prompt All in One · Negative" looks like a joke of a node: one multiline prompt input, one STRING output, and the execute function just returns your text. But that's the whole architecture of this pack - a faithful port of the beloved A1111 extension sd-webui-prompt-all-in-one. The canvas node is a dumb passthrough so the workflow stays simple; the actual work happens in a floating React editor you open by clicking Open Prompt Editor on the node.

That editor is where the negative side finally gets treated as a first-class citizen. Your negative is split into individual tags you can drag-reorder, toggle on and off individually or in batch, blacklist, weight-bump, and wrap in brackets. There's a grouped, searchable tag library (English and Chinese ship with the pack) and its own history of your last 100 negative edits, separate from the positive node's. If you ever watched a negative prompt scroll past the edge of an A1111 text box and lose the one keyword that fixed your hands, this is the pitch.

Two things to know before you fall in love

First, the wiring is exactly what you expect: feed STRING into the text input of the negative text-encode node (the CLIPTextEncode whose conditioning goes to the sampler's negative slot). Since output is a plain string, nothing else changes in your workflow.

Second, and this matters more: your negative only does something while CFG is above 1. Classifier-free guidance pushes the sampler away from whatever you type, but at CFG 1 the negative pass isn't even computed - ComfyUI skips it as an optimization. That's true of Flux Dev at its default, and it's true of basically every Turbo/Lightning/Z-Image Turbo build. If you've been meticulously curating a negative for a distilled model and seeing no effect, the node isn't broken and neither are you - the mechanism is inert at your CFG. The KB's own digging confirms the negative prompt held its ground in community use precisely because so much of what people run is still SDXL-lineage, where it works and stays worth having. On those models, keep your negative short and name the defect you can actually see rather than the forty-keyword boilerplate - and remember the weight-bump button produces (x:1.3) syntax, which only CLIP-family encoders respect. On an LLM-encoded model it's just literal punctuation.

Installing it

The easy path is ComfyUI Manager: search comfyui-prompt-all-in-one. The registry release includes the pre-built frontend. Manually:

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

Restart ComfyUI after. Two traps that send people to the issue tracker. The pack registers nodes through ComfyUI's new V3 extension API only - no legacy fallback - so on an older ComfyUI the nodes simply never appear; update ComfyUI. And if you cloned from source, dist/ is gitignored, meaning the floating editor UI is missing until you build it: cd ui && npm ci && npm run build. If "Open Prompt Editor" silently does nothing, that's your culprit, and the fastest fix is reinstalling via Manager instead.

Honest caveats

This pack is days old - created September 2026, zero stars, nobody's stress-tested it yet - so treat the "why is this still rough" moments accordingly. Editor state (disabled tags, ordering) serializes into the workflow's node properties, and per-user data lives under ComfyUI/user/<user-id>/prompt_all_in_one/. One thing the A1111 original quietly let you get away with does not transfer: the editor will validate <lora:...> and embedding: references against what you have installed, but ComfyUI won't auto-load a LoRA from text. You still wire in a LoRA Loader. Validation yes; loading, never.