8 ms·
Imagine Flash: Accelerating Emu Diffusion Models with Backward Distillation
- cs702 2y agoThe authors claim to have reduced the number of iterations necessary for sampling from ~25 to 3, while improving image quality. The sample images look amazing. I've added this to my (very long) reading list.
- Dwedit 2y agoThe SDXL-Lightning model is doing something very similar. It is a "distilled" version of SDXL for very low step counts (2, 4, 8). In addition to a full model, it can be integrated into other trained SDXL models as a LORA. This seems to be an attempt to get the benefits of the Lightning model without the problems that come with it (worse prompt adherence)