Their main claim of “faster” unfortunately is false.
> Running on an Nvidia A100 GPU, Paella took 0.5 seconds to produce a 256x256-pixel image in eight steps, while Stable Diffusion took 3.2 seconds
Using the latest methods (torch 2.0 compile, improved schedulers) stable diffusion only takes about 1 second to generate a 512x512 image on an a100 gpu. A 256x256 image 1/4 the size presumably takes less than half that time.
So the corrected title is “Like diffusion but slightly slower and lower quality.”
> Running on an Nvidia A100 GPU, Paella took 0.5 seconds to produce a 256x256-pixel image in eight steps, while Stable Diffusion took 3.2 seconds
Using the latest methods (torch 2.0 compile, improved schedulers) stable diffusion only takes about 1 second to generate a 512x512 image on an a100 gpu. A 256x256 image 1/4 the size presumably takes less than half that time.
So the corrected title is “Like diffusion but slightly slower and lower quality.”