Turns out you can hook up a local LLM to ComfyUI and have it write your Danbooru tags for you. It's pretty neat.
The Annoying Part of AI Art
Look, we all know the drill. You've got this cool idea in your head - maybe a cyberpunk cat or whatever - but then you sit down to actually write the prompt and suddenly you're googling "what's the difference between cyberpunk and sci-fi tags again?"
Then you spend 20 minutes trying to remember if it's circuit_board or circuitry, whether you need both glowing and neon, and which quality tags actually do anything with your model. By the time you're done, you've forgotten what you originally wanted to make.
It's just... tedious.
A Simple Solution
So I built a workflow that takes care of the annoying part. You write something normal like "a cat made of circuit boards" and a local LLM figures out all the technical tag stuff for you.
Here's what it actually does:
Takes your basic description
Runs it through a local LLM (using Ollama)
The LLM adds appropriate Danbooru tags and enhances the description
Feeds the improved prompt into SDXL
That's it. Nothing revolutionary, just convenient.
What It Actually Looks Like
I fed it this: "A cat silhouette made of circuit boards"
The LLM gave me back:
"An abstract stylized circuit board morphed into a cat's silhouette, glowing neon blue against a pitch black background, conveying intelligence and technology."
Plus these tags: catgirl, made_in_japan, neon, sci-fi, circuit_board, mecha, technology, futuristic, artificial_intelligence, AI
Which is honestly better than what I would have written myself. The LLM picked up on stuff I wouldn't have thought of, like suggesting the neon blue glow and the black background. It also threw in some anime-style tags that actually improved the final image.
Why It's Useful
Less thinking about tags: The LLM knows which tags work well together and which ones are redundant.
Better prompts: It often suggests improvements to your original idea that you wouldn't have thought of.
Consistency: Every prompt gets the same level of detail, so your results are more predictable.
Learning tool: You can see how it interprets your ideas and pick up better prompting habits.
The Setup
The workflow uses Ollama running locally with the a local model of yoru choice. I don't have too much Memory to spare, so i chose a 8B model. With a Keep_alive of 0 it gets unloaded after use. You can of course subsitute whatever you want here, but l3-8b-stheno-v3.2 Is particularily good at NSFW as well... Everything stays on your machine, so it's fast and private. It's hooked into a standard SDXL pipeline with some upscaling and post-processing.
The LLM has a system prompt that tells it to focus on art-related enhancements and Danbooru tagging conventions. You can tweak this if you want different behavior.
Is It Worth It?
If you generate a lot of images and get tired of writing the same types of tags over and over, yeah, it's pretty handy. It's not going to suddenly make you a better artist or anything, but it does remove some of the friction from the process.
The biggest benefit is that it lets you focus on the creative part (what do I want to make?) instead of the technical part (how do I describe this to the AI?).
Plus, it's kind of fun to see how the LLM interprets your ideas. Sometimes it suggests things that make the final image way better than what you originally had in mind.
Sometimes you might just want to turn it off.
The Reality Check
This isn't some magical solution that's going to transform AI art forever. It's just a practical tool that automates the boring part of prompt writing. The LLM sometimes gets things wrong, you still need to know your models and settings, and you'll probably want to edit the output sometimes.
But for day-to-day generation, especially if you're making a lot of images and you like to iterate and see what shows up, it's genuinely useful. Instead of spending time on tag management, you can spend time on actually making stuff.
If you try it out, let me know how it works for you. Always curious to see how other people use this kind of setup.