| Mar 12, 2025 | Our new paper, Denoising Score Distillation: From Noisy Diffusion Pretraining to One-Step High-Quality Generation, introduces a novel approach to training a one-step image generator using only noisy images. Remarkably, our method achieves FID scores comparable to diffusion models trained on clean images. Beyond proposing a new solution to inverse problems through distillation, our work demonstrates that distillation is not merely an acceleration technique but also enhances generation quality compared to the teacher diffusion model—both empirically and theoretically. |
