[anonzzzies]

I love the idea but it is not working for me : even the example prompts are saying they cannot generate any design.

So how does it work? SD is not gpt right? So does it use sd or gpt or some clever mix, like gpt3 writing prompts for dall-e to make sure they are designs for web components?

[designhat]

Sorry, looks like a memory leak built up overnight. Should be working now.

Glad you like the idea! It's a combination of GPT-3 and traditional layout optimization similar to this paper: http://www.dgp.toronto.edu/~donovan/layout/

GPT-3 is prompted to output JSON that describes which elements should be in the design (label, icon, image, badge, etc), their copy and which elements should be grouped together. Then the layout optimization module determines how those elements should be arranged spatially.

I eventually want to experiment with fine-tuning stable diffusion to handle the layout part of the problem, by outputting colored rectangles that represent the layout, like in this paper http://interactionmining.org/rico