This paper proposes a simple method which solves an open problem of multi-view 3D-Reconstruction for objects with unknown and generic surface materials, imaged by a freely moving camera and a freely moving point light source. The object can have arbitrary (e.g. non-Lambertian), spatially-varying (or everywhere different) surface reflectances (svBRDF). Our solution consists of two smallsized neural networks (dubbed the 'Shape-Net' and 'BRDFNet'), each having about 1,000 neurons, used to parameterize the unknown shape and unknown svBRDF, respectively. Key to our method is a special network design (namely, a ResNet with a global feedback or 'ring' connection), which has a provable guarantee for finding a valid diffeomorphic shape parameterization. Despite the underlying problem is highly non-convex hence impractical to solve by traditional optimization techniques, our method converges reliably to high quality solutions, even without initialization. Extensive experiments demonstrate the superiority of our method, and it naturally enables a wide range of special-effect applications including novel-view-synthesis, relighting, material retouching, and shape exchange without additional coding effort. We encourage the reader to view our demo video for better visualizations.