Jasper Research just released a cookbook on how to build a text-to-image model from scratch.
It shares the full reasoning and intermediate results, making it ideal if you want to deep-dive into text-to-image models, or if you are curious about how frontier labs build them.
The cookbook also includes a 100M-image dataset and a codebase with a tiny model, so you can train a text-to-image model from scratch.
Here are the links:
Cookbook: https://huggingface.co/spaces/jasperai/t2i-technical-interactive-report
nano t2i: https://github.com/gojasper/nano-t2i
Monet Dataset: https://huggingface.co/datasets/jasperai/monet
[link] [留言]