Yukang Cao Chenyang Si Jinghao Wang Ziwei Liu

Abstract

We present FreeMorph, the first tuning-free method for image morphing thataccommodates inputs with different semantics or layouts. Unlike existingmethods that rely on finetuning pre-trained diffusion models and are limited bytime constraints and semantic/layout discrepancies, FreeMorph delivershigh-fidelity image morphing without requiring per-instance training. Despitetheir efficiency and potential, tuning-free methods face challenges inmaintaining high-quality results due to the non-linear nature of the multi-stepdenoising process and biases inherited from the pre-trained diffusion model. Inthis paper, we introduce FreeMorph to address these challenges by integratingtwo key innovations. 1) We first propose a guidance-aware sphericalinterpolation design that incorporates explicit guidance from the input imagesby modifying the self-attention modules, thereby addressing identity loss andensuring directional transitions throughout the generated sequence. 2) Wefurther introduce a step-oriented variation trend that blends self-attentionmodules derived from each input image to achieve controlled and consistenttransitions that respect both inputs. Our extensive evaluations demonstratethat FreeMorph outperforms existing methods, being 10x ~ 50x faster andestablishing a new state-of-the-art for image morphing.