The researchers present a novel Noise-Free Score Distillation (NFSD) process that requires minimal changes to the original Score Distillation Sampling (SDS) framework. NFSD improves the effectiveness of pre-trained text-to-image diffusion models using a smaller Classifier-Free Guidance (CFG) scale. This method prevents the over-smoothing of results, ensuring the generated data is realistic and adheres to the desired prompt. The paper also discusses the general nature of their framework and how it supports and provides explanations for recent methods like VSD and DDS, which have shown improvements over SDS.
Publication date: 26 Oct 2023
Project Page: https://arxiv.org/abs/2310.17590v1
Paper: https://arxiv.org/pdf/2310.17590
