Cross-view Fusion — A variable number of masked input views is encoded with DINOv2 and camera-aware modulation. One hero view anchors the query stream while additional views provide cross-attention memory, producing triplane features that are consistent across viewpoints.

Two-path Prediction — The geometry and appearance path predicts density, mesh structure, and spatially varying albedo, roughness, metallic, and bump normals from a shared triplane embedding. In parallel, the illumination path fuses mask-aware tokens with object tokens to infer RENI++ latents and recover HDR environment lighting.

Disentangled Training — A differentiable Monte Carlo renderer with Multiple Importance Sampling ties both paths together. This enforces physically meaningful material-lighting separation and allows mixed-domain training across synthetic PBR-supervised data, synthetic RGB-only data, and real-world captures.

PBR Prediction — On Polyhaven + Blender Shiny evaluation, ReLi3D reports 25.00 dB basecolor PSNR, 22.69 dB roughness PSNR, and 32.73 dB metallic PSNR in the single-view setting, with further gains in multi-view mode.

Illumination Estimation — The model recovers plausible sky color and dominant light direction even from a single view, and improves source localization when more views or background evidence are available. Compared to prior RENI++ usage, the multi-view illumination path produces sharper and more coherent environment predictions.

Overall Reconstruction — In the combined GSO + Stanford ORB evaluation, ReLi3D reaches 19.57 PSNR and 0.902 SSIM in the single-view setting, and improves further with more views. In UCO3D, it reports 15.28 PSNR single-view and up to 15.73 PSNR in multi-view settings.

Cross-domain Real-world Performance — The model is trained on 174k objects across synthetic PBR, synthetic RGB-only, and real-world UCO3D captures. On Stanford ORB, ReLi3D outperforms baselines across geometry, image quality, and basecolor metrics, and continues to improve from 1-view to 16-view input.