Towards High Fidelity Face Relighting with Realistic Shadows

De-rendering 3D Objects in the Wild

“…6 shows two examples. As a result of the underlying explicit image formation model, our method produces visually correct relighting results, which are more color accurate than [58] and contain much fewer artifacts than [12]. This demonstrates that our method can not only perform very well in intrinsic image decomposition, but also compete with methods, that were specifically designed for certain sub-tasks.…”

“…Single Image Relighting. To demonstrate the usefulness of de-rendering, we perform relighting on the CelebA-HQ dataset and compare with state-of-the-art methods [58] and [12], which were specifically designed for face relighting. Fig.…”

“…The image de-rendering task can also be learned with direct supervision, often using synthetic data, such as ShapeNet [6] objects [28,47], synthetic faces [12,43], nearplanar surfaces [24], indoor scenes [22] or other synthetic objects [13,25,41]. However, generating large-scale realistic synthetic data that capture the level of complexity of the real world is often challenging, and hence it remains questionable how well these methods generalize to real images.…”

“…The light network follows a classical encoder architecture. All 12 . We adapt the models to fit our requirements, for example, define camera views and modify the material so that we can extract the diffuse albedo.…”

Deep Portrait Delighting

“…None of them have a dedicated delighting solution independent from their application. For example, the relighting application area has seen rapid progress with deep neural networks able to render convincing nonlambertian shading effects [20,32,33,42,45,49]. However, the delighting phase is often abstracted.…”