Draw Me a Sheep: GPT to SVG

I need to close a chapter before I move on, but this time I found myself struggling to find the right words to tell the story, so I'll state it plain.

I built a tool that transforms input prompts to SVG illustrations. It runs on GPT3.5, Janet, HMTX, a home-grown scheduler that keeps objects in the file system, as opposed to a traditional database and a few other unconventional software decisions. A backstory with dozens of cats and unicorns: https://drawmeasheep.net/pages/about.html

The tool currently has a cap of 4 images a minute, and it may take up to a minute to generate an image. I have a monthly budget allocated to GPT, but besides that there are no other limitations. The app may occasionally generate "I'm a large language model" error or be rate-limited by OpenAI. But that's okay–just try again.

Sounds interesting? Check it out: https://drawmeasheep.net/

GPT3.5 doesn't always generate the most artistic illustrations, but once in a while it may draw surprisingly reasonable images. GPT4 is more creative, but is also more expensive.

On average it takes about 1k tokens to generate an image, hence the price is $0.002 per image.

I selected a few sketches and corresponding prompts to demonstrate what GPT3.5 can do. Tip: right-click and inspect source code, I told GPT to leave clarification comments – the last image will make more sense.

baba yaga

the great wave off kanagawa

polar bear



fruit fly

see no evil

hacker news