So I tried fine-tuning GPT-3.5 to help me write SD prompts, and the results were... kinda unexpected lol.

I’ve never been good at writing those long, detailed prompts. Tweaking details is fine, but starting from scratch and cranking out a huge block of text? That gives me a headache. So I figured, why not let an LLM be my prompt assistant? I fine-tuned a GPT-3.5 to specifically turn crappy prompts into good ones.

The idea’s pretty simple: first, I collected about 3,000 prompts written by Midjourney pros. What’s interesting is that the average length is only around 25 words—so longer isn’t always better. Then I used GPT to compress those good prompts into short, rough versions, paired them up, and fed about 2,000 pairs into the fine-tuning. The whole training took about an hour and cost me just over seven bucks. Just keep in mind, the longer your system prompt, the more expensive it gets.

The output is all over the place—same input can give you totally different styles. Quality varies, but the real win is that it helps me break out of creative ruts. For example, I threw in “vacuum cleaner Transformer,” and it spat out something like a 20-foot-tall robot that turns into a vacuum cleaner. When I used that to generate images, the results were surprisingly awesome. If you’re not chasing consistency, it’s a blast to use as an idea generator.

That 25-token dataset is pretty counterintuitive, I always thought you’d want to pile on as many as possible.

Seven bucks per training run, that’s way cheaper than I thought.

Same here, man. Writing long prompts always makes me freeze up.

That trick of reverse-engineering bad data is genius, way less hassle than writing it all by hand.

Yeah, I was surprised by that 25-word prompt too. If you pile on too much, the model just loses the plot.