Enhanced Prompt Editor

The name undersells it. Enhanced Prompt Editor (EPE) isn't a fancy text box - it's a whole prompt-writing studio packed into a single ComfyUI node, driven entirely by a local LLM running through Ollama. Type a rough idea like "cyberpunk courier on a rooftop at night," hit Enhance, and it writes a diffusion-ready prompt shaped for modern encoders (Flux, Qwen-Image and friends): subject-first, no negations, spatially placed, quoting any text you want verbatim. No cloud API, no key, nothing leaves your machine. That's the entire pitch, and it's real.

The local-LLM-prompt-enhancer pattern is mainstream by now - an 8B model on your own card writing the instruction that your checkpoint's own text encoder will read. EPE is a polished, self-contained take on it. What makes it stand out is how much of the writing workflow it absorbs: multi-tab editor with undo, find-and-replace, word/token counts, synonyms, a "flag words" pass that catches empty qualifiers like "beautiful" or "masterpiece" - the stuff you'd normally patch together from a dozen little utilities.

How it actually works

Here's the part that trips people up: the node has zero inputs and zero outputs. Its INPUT_TYPES are empty, its return types are empty. EPE is an output node that does everything through the ComfyUI backend and its own browser UI. Instead of wiring, you use wireless targets - click + Add target, pick a text widget anywhere in your graph (CLIP Text Encode, even inside subgraphs), and on every queue the current tab's prompt is injected straight into that widget. The node reports itself changed on every run so ComfyUI always executes it, which is what makes the injection fire each queue. Prefer explicit wiring? Skip targets and just copy the prompt out.

All the AI calls go to your local Ollama server, proxied through ComfyUI's backend rather than called directly from the browser. That detail matters: it's why you don't need to fiddle with OLLAMA_ORIGINS or CORS when ComfyUI is opened from another machine.

The bits you'll actually use

• Transform rail - Enhance expands an idea; Variations writes three distinct takes; Inverter flips it into a contrasting aesthetic. Results stream in for review, keep or undo.
• Instruct edit - the sleeper hit. Type "make her hair red" or "change to golden hour" in plain language and the model rewrites around it, rippling through related details. Edits chain, and saved sequences replay against any prompt.
• Style tuning - 8 aesthetic targets plus 6 sliders (Creativity, Length, Focus, Variability, Boldness, Subject grip). Override On re-art-directs the whole look; Override Off only fills gaps your prompt leaves open. This is where model quality shows most - a bigger model follows the style system noticeably better.
• From media - Image to Prompt and Video to Prompt caption media into prompts (vision model required; the Python av package handles video frames), and Extract from Image pulls the embedded prompt out of a ComfyUI-generated PNG.
• Library - browse prompts from Civitai and Genur.art in-node, or load whole workflows from Civitai into a new canvas tab.

Installing it

Easiest path is ComfyUI Manager: search Enhanced Prompt Editor, install, restart. Or via the CLI:

comfy node install comfyui-ai-prompt-editor

Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/pixelpainter/comfyui-AI-prompt-editor
pip install -r comfyui-AI-prompt-editor/requirements.txt

The only Python dependency is av>=10.0.0, for video frames - light. The real requirement is Ollama running on the machine, plus models pulled through it:

ollama pull ministral      # text - Enhance, Variations, Instruct edit
ollama pull qwen3-vl       # vision - Image/Video to Prompt only

Then open ⚙ AI Setup in the node's title bar and pick your model. On Linux the node will even auto-start Ollama if it's on your PATH.

Where people get burned

"The AI does nothing" is almost always Ollama not being reachable - check it's installed and running, then re-check in ⚙ AI Setup. "Image to Prompt is greyed out" means you picked a text-only model; you need a vision one. "Results look generic" - bump to a bigger or better instruct model, not the smallest one; the README is upfront that larger models track the style system far better.

One thing worth knowing about this category: LLM custom nodes run arbitrary Python and have historically been a malware vector (the infamous ComfyUI_LLMVISION incident). EPE is GPL-3.0, open, and its server routes are genuinely hardened - Ollama addresses are locked to localhost/private networks, and its media fetching refuses internal URLs. It's a rare pack where reading the code makes you more comfortable, not less.

It's a small pack from a solo author (pixelpainter), still under the community's radar - install it because the tool is good, not because everyone's talking about it. If you hate hand-wrangling prompts, it's the one I'd reach for.