Würstchen brings 42x compression to image generation
The creators of Würstchen just released a lightweight text-to-image diffusion model that drastically reduces training and inference costs. By pushing latent spatial compression to an unprecedented 42x, it generates images faster than SDXL while slashing training compute down to 9,000 GPU hours compared to SD 1.4's 150,000. It makes training and running custom image models accessible without a farm of A100s.