Polarization-conditioned diffusion model for shape from polarization

Shape from polarization (SfP) provides physically informative cues for surface normal recovery by exploiting the polarization state of reflected light. Existing SfP methods remain limited by the imperfect nature of the physical model and the scarcity of labeled datasets. Recent diffusion models are attractive for SfP because of their strong generative capability in geometry estimation. Their direct application remains challenging due to the need for polarization-guided diffusion and reduced stochastic ambiguity during inference. To address this issue, we propose PcdSfP, a polarization-conditioned diffusion model for shape from polarization. The unpolarized intensity, degree of polarization (DoP), and angle of polarization (AoP) are encoded as semantic anchors and polarimetric conditioning cues to guide a pre-trained latent diffusion backbone toward geometrically consistent normal recovery. A spatially adaptive noise injection strategy is further introduced to reduce ambiguity during inference, and joint supervision in latent and pixel spaces is adopted to preserve high-frequency boundaries and local geometric details. Experiments demonstrate that the proposed method effectively integrates the physical observables of polarimetric imaging with the structural priors of diffusion models, achieving improved accuracy and detail recovery in surface normal estimation across both object-level and scene-level SfP tasks.