Diffusion models achieve state-of-the-art performance in generating new samples, but lack a low-dimensional latent space that encodes data into editable features. Inversion-based methods address this issue by inverting the denoising trajectory to approximate the initial noise. This study thoroughly analyzes this process, focusing on the relationship between the initial noise, the generated samples, and the corresponding latent encoding obtained through DDIM inversion. We find that the latent exhibits structural patterns that predict less diverse noise for smooth image regions (e.g., plain sky). This problem stems from the failure of the first inversion step to provide accurate and diverse noise. Consequently, the DDIM inversion space is significantly less manipulable than the original noise. While existing inversion methods do not completely resolve this issue, a simple solution—replacing the first DDIM inversion step with a forward diffusion process—successfully separates the latent encoding and enables higher-quality editing and interpolation.
