Recent text-to-image diffusion and flow models have enabled a range of prompt-based editing methods. These methods, however, face an inherent trade-off: the edited output must stay faithful to the original source image while also satisfying the new textual prompt. Pushing too far in one direction degrades the other.
A new arXiv preprint proposes instance-specific classifier-free guidance to address this problem. Instead of using a single global guidance strength, the method tailors the guidance to each particular source image and edit task. The authors argue that this per-instance adaptation can better balance source preservation with prompt adherence.
The paper is described as an announcement, and the abstract does not yet include experimental details or quantitative results. The contribution is framed as a general technique applicable to both diffusion and flow-based text-to-image models, potentially offering a more flexible alternative to fixed-guidance editing pipelines.