SDXL-Lightning: Progressive Adversarial Diffusion Distillation

arXiv:2402.13929v1 Announce Type: cross
Abstract: We propose a diffusion distillation method that achieves new state-of-the-art in one-step/few-step 1024px text-to-image generation based on SDXL. Our method combines progressive and adversarial distillation to achieve a balance between quality and mode coverage. In this paper, we discuss the theoretical analysis, discriminator design, model formulation, and training techniques. We open-source our distilled SDXL-Lightning models both as LoRA and full UNet weights.

abstract adversarial analysis art arxiv balance coverage cs.ai cs.cv cs.lg design diffusion discuss distillation image image generation lightning paper quality sdxl state text text-to-image training type