Slight Corruption in Pre-training Data Makes Better Diffusion Models | Read Paper on Bytez