Title: Training-free Image Inversion for One-Step Diffusion Models

Abstract:

This study presents TFinv, a new training-free inversion framework designed for one-step diffusion models, specifically aimed at overcoming significant hurdles in real-image inversion and editing tasks. Our analysis highlights two primary obstacles that impede effective inversion and manipulation of real images: (1) Initial Latent Editability, which pertains to the disparity between the starting noise and the target Gaussian distribution, and (2) the Caption Gap, defined as the misalignment between textual descriptions and image embeddings. Both elements critically impact the efficiency of inversion and the subsequent editability of one-step diffusion models.

To address these issues, we introduce two innovative methods. First, iterative noise alignment (iterNA) is employed to reduce the distributional divergence, thereby aligning the noise with a standard Gaussian distribution. Second, suffix learning (suffL) improves the correlation between text captions and image representations by incorporating learnable suffix prompt tokens. Together, these techniques allow for the accurate transformation of input images into their corresponding initial noise states, thereby enabling effective image editing. Additionally, we develop a mask-based approach for localized modifications that maintain the integrity of the background. Extensive evaluations on the PIE-Bench dataset demonstrate that TFinv not only sets a new state-of-the-art standard for one-step diffusion editing but also surpasses current multi-step methods in terms of efficiency. The source code is accessible at https://github.com/tttao-uwu/TFinv.git.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC