tanelp / tiny-diffusion Goto Github PK

As the ablation study in readme.md shows, the current model configuration doesn't work well on the line dataset. The reason is the the timestep is not large enough such that Xt is not in standard gaussian distribution. If we set timestep as 250 instead of 50, the result is nice.

Simplest things are more amazing.

Hi~

Good work~ This is such a genius implementation of diffusion. It's very impressive to see that DDPM is implemented on such a small number of codes~ Really appreciate your effort and intelligence poured into this repo^^ If it would be convenient, may I ask what are the lines of code/area that are the mini tricks to help implement this DDPM model in a small number of codes, that is different from the original DDPM model

Thanks a lot~

Hi @tanelp. Wondering if there's any material I can refer to for understanding how the reconstruction from noise works. TIA