Pose-ICL: 3D-Aware In-Context Learning for Pose-Controllable Subject Customization

Subject Customization is a foundational task in modern image generation. By providing a few reference images and a text prompt, users can generate images of a specific object in any desired scene. However, existing methods still struggle to achieve effective pose control for customized subjects. In practice, they often exhibit inaccurate poses or inconsistent cross-pose appearances. These limitations suggest that understanding objects in a volumetric manner remains a significant challenge for 2D-native backbones. To address this challenge, we propose Pose-ICL, a tuning-free framework that leverages 3D-aware In-Context Learning (ICL) to directly adapt to new subjects through multiple paired image-pose references. Its core mechanism, Surface-Anchored Position Embedding (SAPE), equips the model with explicit 3D awareness by anchoring image tokens to the surface coordinates of a volumetric bounding box. Dedicated refinements ensure its seamless compatibility with existing DiT models. Extensive evaluations on both 3D assets and real-world subjects demonstrate that Pose-ICL significantly outperforms current methods in both pose accuracy and identity consistency.

Modern image generation systems can create new images of a specific object from only a few reference pictures and a text prompt. However, they often have difficulty controlling the object’s pose. When asked to show the same object from a new viewpoint, the result may face the wrong direction or lose important visual details that make the object recognizable. We propose Pose-ICL, a method that helps the model treat the object more like a 3D shape rather than a collection of separate 2D pictures. It learns from several reference images together with their poses, and uses a simple 3D box around the object to connect image features to surface locations. This gives the model a better understanding of how the object should look from different views. Experiments show that Pose-ICL generates images with more accurate poses and more consistent object identity, making pose-controllable customization easier and more dependable.