Denoising Diffusion Probabilistic Models (DDPMs) achieve state-of-the-art performance in synthesizing new images from random noise, but they lack meaningful latent space that encodes data into features. Recent DDPM-based editing techniques try to mitigate this issue by inverting images back to their approximated staring noise. In this work, we study the relation between the initial Gaussian noise, the samples generated from it, and their corresponding latent encodings obtained through the inversion procedure. First, we interpret their spatial distance relations to show the inaccuracy of the DDIM inversion technique by localizing latent representations manifold between the initial noise and generated samples. Then, we demonstrate the peculiar relation between initial Gaussian noise and its corresponding generations during diffusion training, showing that the high-level features of generated images stabilize rapidly, keeping the spatial distance relationship between noises and generations consistent throughout the training.

本研究解决了去噪扩散概率模型（DDPMs）在生成新图像时缺乏有意义的潜在空间问题。通过研究初始高斯噪声、由其生成的样本和通过反演过程获得的潜在编码之间的关系，提出了一种新颖的看法，指出反演技术的局限性并证明生成图像的高层特征在训练中迅速稳定，空间距离关系保持一致。这项工作为进一步优化图像生成模型提供了重要见解。

归来与再出发：噪声、图像及其在扩散模型中的反演关系研究