symisc/tiny-dream

SD1 optimized?

ClashSAN opened this issue · 1 comments

ClashSAN commented

Hi there, this project sounds very exciting!
Particularly having fast, memory efficient inference on low ram android smartphones.

SD2 is mentioned, was SD1 architecture optimized for this library as well?

symisc commented

Hello,
Actually, we reverted back to a Stable Diffusion 1.x implementation using the CLIP transformer for token embedding, 512*512 pixels output that can be upscaled to 2048x2048 high resolution output.