In this paper, we propose RI3D, a novel 3DGS-based approach that harnesses the power of diffusion models to reconstruct high-quality novel views given a sparse set of input images. Our key contribution is separating the view synthesis process into two tasks of reconstructing visible regions and hallucinating missing regions, and introducing two personalized diffusion models, each tailored to one of these tasks. Specifically, one model ('repair') takes a rendered image as input and predicts the corresponding high-quality image, which in turn is used as a pseudo ground truth image to constrain the optimization. The other model ('inpainting') primarily focuses on hallucinating details in unobserved areas. To integrate these models effectively, we introduce a two-stage optimization strategy: the first stage reconstructs visible areas using the repair model, and the second stage reconstructs missing regions with the inpainting model while ensuring coherence through further optimization. Moreover, we augment the optimization with a novel Gaussian initialization method that obtains per-image depth by combining 3D-consistent and smooth depth with highly detailed relative depth. We demonstrate that by separating the process into two tasks and addressing them with the repair and inpainting models, we produce results with detailed textures in both visible and missing regions that outperform state-of-the-art approaches on a diverse set of scenes with extremely sparse inputs.

In this paper, we present PanoDreamer, a novel method for producing a coherent 360° 3D scene from a single input image. Unlike existing methods that generate the scene sequentially, we frame the problem as single-image panorama and depth estimation. Once the coherent panoramic image and its corresponding depth are obtained, the scene can be reconstructed by inpainting the small occluded regions and projecting them into 3D space. Our key contribution is formulating single-image panorama and depth estimation as two optimization tasks and introducing alternating minimization strategies to effectively solve their objectives. We demonstrate that our approach outperforms existing techniques in single-image 360° 3D scene reconstruction in terms of consistency and overall quality.

Rendering novel, relit views of a human head, given a monocular portrait image as input, is an inherently underconstrained problem. The traditional graphics solution is to explicitly decompose the input image into geometry, material and lighting via differentiable rendering; but this is constrained by the multiple assumptions and approximations of the underlying models and parameterizations of these scene components. We propose 3DPR, an image-based relighting model that leverages generative priors learnt from multi-view One-Light-at-A-Time (OLAT) images captured in a light stage. We introduce a new diverse and large-scale multi-view 4K OLAT dataset of 139 subjects to learn a high-quality prior over the distribution of high-frequency face reflectance. We leverage the latent space of a pre-trained generative head model that provides a rich prior over face geometry learnt from in-the-wild image datasets. The input portrait is first embedded in the latent manifold of such a model through an encoder-based inversion process. Then a novel triplane-based reflectance network trained on our lightstage data is used to synthesize high-fidelity OLAT images to enable image-based relighting. Our reflectance network operates in the latent space of the generative head model, crucially enabling a relatively small number of lightstage images to train the reflectance model. Combining the generated OLATs according to a given HDRI environment maps yields physically accurate environmental relighting results. Through quantitative and qualitative evaluations, we demonstrate that 3DPR outperforms previous methods, particularly in preserving identity and in capturing lighting effects such as specularities, self-shadows, and subsurface scattering.

Abstract Lightweight material capture methods require a material prior, defining the subspace of plausible textures within the large space of unconstrained texel grids. Previous work has either used deep neural networks (trained on large synthetic material datasets) or procedural node graphs (constructed by expert artists) as such priors. In this paper, we propose a semi‐procedural differentiable material prior that represents materials as a set of (typically procedural) grayscale noises and patterns that are processed by a sequence of lightweight learnable convolutional filter operations. We demonstrate that the restricted structure of this architecture acts as an inductive bias on the space of material appearances, allowing us to optimize the weights of the convolutions per‐material, with no need for pre‐training on a large dataset. Combined with a differentiable rendering step and a perceptual loss, we enable single‐image tileable material capture comparable with state of the art. Our approach does not target the pixel‐perfect recovery of the material, but rather uses noises and patterns as input to match the target appearance. To achieve this, it does not require complex procedural graphs, and has a much lower complexity, computational cost and storage cost. We also enable control over the results, through changing the provided patterns and using guide maps to push the material properties towards a user‐driven objective.

Recent methods (e.g. MaterialGAN) have used unconditional GANs to generate per-pixel material maps, or as a prior to reconstruct materials from input photographs. These models can generate varied random material appearance, but do not have any mechanism to constrain the generated material to a specific category or to control the coarse structure of the generated material, such as the exact brick layout on a brick wall. Furthermore, materials reconstructed from a single input photo commonly have artifacts and are generally not tileable, which limits their use in practical content creation pipelines. We propose TileGen, a generative model for SVBRDFs that is specific to a material category, always tileable, and optionally conditional on a provided input structure pattern. TileGen is a variant of StyleGAN whose architecture is modified to always produce tileable (periodic) material maps. In addition to the standard “style” latent code, TileGen can optionally take a condition image, giving a user direct control over the dominant spatial (and optionally color) features of the material. For example, in brick materials, the user can specify a brick layout and the brick color, or in leather materials, the locations of wrinkles and folds. Our inverse rendering approach can find a material perceptually matching a single target photograph by optimization. This reconstruction can also be conditional on a user-provided pattern. The resulting materials are tileable, can be larger than the target image, and are editable by varying the condition.

In this paper, we propose a novel optimization-based method to estimate the reflectance properties of a near planar surface from a single input image. Specifically, we perform test-time optimization by directly updating the parameters of a neural network to minimize the test error. Since single image SVBRDF estimation is a highly ill-posed problem, such an optimization is prone to overfitting. Our main contribution is to address this problem by introducing a training mechanism that takes the test-time optimization into account. Specifically, we train our network by minimizing the training loss after one or more gradient updates with the test loss. By training the network in this manner, we ensure that the network does not overfit to the input image during the test-time optimization process. Additionally, we propose a learned reflectance loss to augment the typically used rendering loss during the test-time optimization. We do so by using an auxiliary network that estimates pseudo ground truth reflectance parameters and train it in combination with the main network. Our approach is able to converge with a small number of iterations of the test-time optimization and produces better results compared to the state-of-the-art methods.

Abstract In this paper, we propose a deep learning approach for estimating the spatially‐varying BRDFs (SVBRDF) from a single image. Most existing deep learning techniques use pixel‐wise loss functions which limits the flexibility of the networks in handling this highly unconstrained problem. Moreover, since obtaining ground truth SVBRDF parameters is difficult, most methods typically train their networks on synthetic images and, therefore, do not effectively generalize to real examples. To avoid these limitations, we propose an adversarial framework to handle this application. Specifically, we estimate the material properties using an encoder‐decoder convolutional neural network (CNN) and train it through a series of discriminators that distinguish the output of the network from ground truth. To address the gap in data distribution of synthetic and real images, we train our network on both synthetic and real examples. Specifically, we propose a strategy to train our network on pairs of real images of the same object with different lighting. We demonstrate that our approach is able to handle a variety of cases better than the state‐of‐the‐art methods.