Embedded Files
Posture-Requested Personalized Image Generation (RotateID)
Introduction
Introduction
Identity-preserving personalization has become remarkably good at one thing: give it a single photo of a person and it will reproduce that face. Ask it to turn the head, though, to look over the shoulder, glance back, or tilt the chin up, and most methods quietly refuse. They collapse back to a near-frontal view and hand back a face that is recognizably the right person but in the wrong pose.
The evaluation practice hides this. Face recognition encoders read identity reliably near frontal but under-report it as the same person turns away, so a method that simply stays frontal scores higher on identity than one that actually performs the requested turn. CLIP alignment, meanwhile, cannot tell head direction at all. The result is a benchmark that rewards exactly the failure it should be catching.
Framework
Framework
Rotate-ID
the generation system, is built on a few-step rectified-flow diffusion transformer (FLUX.2-klein) and works from a single reference image. An in-context structural pose control drives the requested head turn. Identity is preserved by re-allocating the frozen backbone's attention toward the reference image, a mechanism verified causally and then distilled into a learned gate of roughly 200 parameters (the gamma-gate), which refines identity at no cost to the achieved pose.
Pose-rectified performance metrics
The evaluation half, close the two measurement gaps: an explainable 3D pose check verifies that each requested turn actually happened, a pose-adapted read-out scores identity with yaw-matched recognizers, and a pose-aware score credits identity only at the pose that was actually produced. This is instantiated as a benchmark of 40 posture requests spanning eight pose categories.
Results
Results
Under these metrics, Rotate-ID reaches 82.6% yaw accuracy against 36.1% for PhotoMakerV2, the strongest identity-personalization baseline, more than double. It realizes over-the-shoulder rotations at measured yaw angles no baseline attains, and once the frontal-collapsing baselines lose their apparent identity advantage, it remains the strongest tuning-free method under pose-aware identity.
Page updated
Google Sites
Report abuse