RectifID: Personalizing Rectified Flow with Anchored Classifier Guidance

TL;DR

This work tackles identity-preserving personalized image generation without subject-specific training by integrating classifier guidance with rectified flow. It introduces a fixed-point endpoint formulation to remove reliance on noise-aware classifiers and further stabilizes the process by anchoring the guided trajectory to a reference flow. The proposed anchored classifier guidance operates in a piecewise rectified flow setting and leverages off-the-shelf discriminators for flexible personalization across faces, live subjects, and objects, achieving strong identity fidelity and prompt consistency with training-free inference. The approach shows broad generalization to different base diffusion models and controllable generation tasks, offering a practical, efficient alternative to finetuning-based personalization methods.

Abstract

Customizing diffusion models to generate identity-preserving images from user-provided reference images is an intriguing new problem. The prevalent approaches typically require training on extensive domain-specific images to achieve identity preservation, which lacks flexibility across different use cases. To address this issue, we exploit classifier guidance, a training-free technique that steers diffusion models using an existing classifier, for personalized image generation. Our study shows that based on a recent rectified flow framework, the major limitation of vanilla classifier guidance in requiring a special classifier can be resolved with a simple fixed-point solution, allowing flexible personalization with off-the-shelf image discriminators. Moreover, its solving procedure proves to be stable when anchored to a reference flow trajectory, with a convergence guarantee. The derived method is implemented on rectified flow with different off-the-shelf image discriminators, delivering advantageous personalization results for human faces, live subjects, and certain objects. Code is available at https://github.com/feifeiobama/RectifID.

Figures (17)

• Figure 1: Illustration of training-free classifier guidance. Left: an off-the-shelf discriminator can be reused to steer the existing diffusion model, e.g. rectified flow, to generate identity-preserving images. Right: personalized image generation results for human faces and objects using our proposed method.
• Figure 2: Illustration of anchored classifier guidance for rectified flow. Left: we propose to guide the flow trajectory while implicitly enforcing it to flow straight and stay close to a reference trajectory. Right: comparison of the new trajectory with the reference trajectory (in the last three sampling steps).
• Figure 3: Qualitative comparison for face-centric personalization. See \ref{['fig:single_person_app', 'fig:single_person_app2', 'fig:single_person_app3', 'fig:single_person_app4']} for more samples.
• Figure 4: Qualitative comparison for subject-driven generation. $^*$ denotes finetuned with multiple images of the target subject to achieve sufficient identity consistency. See \ref{['fig:single_object_app']} for more samples.
• Figure 5: Qualitative comparison for multi-subject personalization. See \ref{['fig:multi_subject_app']} for more samples.

Theorems & Definitions (4)

• Proposition 1
• proof
• Proposition 2
• proof