Recent advances in feed-forward 3D Gaussian Splatting have led to rapid improvements in efficient scene reconstruction from sparse views. However, most existing approaches construct Gaussian primitives directly aligned with the pixels in one or more of the input images. This leads to redundancies in the representation when input views overlap and constrains the position of the primitives to lie along the input rays without full flexibility in 3D space. Moreover, these pixel-aligned approaches do not naturally generalize to dynamic scenes, where effectively leveraging temporal information requires resolving both redundant and newly appearing content across frames. To address these limitations, we introduce a novel Fuse-and-Refine module that enhances existing feed-forward models by merging and refining the primitives in a canonical 3D space. At the core of our method is an efficient hybrid Splat-Voxel representation – from an initial set of pixel-aligned Gaussian primitives, we aggregate local features into a coarse-to-fine voxel hierarchy, and then use a sparse voxel transformer to process these voxel features and generate refined Gaussian primitives. By fusing and refining an arbitrary number of inputs into a consistent set of primitives, our representation effectively reduces redundancy and naturally adapts to temporal frames, enabling history-aware online reconstruction of dynamic scenes. Trained on large-scale static scene datasets, our model learns an effective global strategy to process around 200k primitives within 15ms and significantly enhances reconstruction quality compared to pixel-aligned reconstruction approaches. Without additional training, our model generalizes to video by fusing primitives across time, yielding a more temporally coherent result compared to baseline methods with graceful handling of occluded content. Our approach achieves state-of-the-art performance in both static and streaming scene reconstructions while running at interactive rates (15 fps with 350ms delay) on a single H100 GPU.

Abstract Intuitively editing the appearance of materials from a single image is a challenging task given the complexity of the interactions between light and matter, and the ambivalence of human perception. This problem has been traditionally addressed by estimating additional factors of the scene like geometry or illumination, thus solving an inverse rendering problem and subduing the final quality of the results to the quality of these estimations. We present a single‐image appearance editing framework that allows us to intuitively modify the material appearance of an object by increasing or decreasing high‐level perceptual attributes describing such appearance (e.g., glossy or metallic). Our framework takes as input an in‐the‐wild image of a single object, where geometry, material, and illumination are not controlled, and inverse rendering is not required. We rely on generative models and devise a novel architecture with Selective Transfer Unit (STU) cells that allow to preserve the high‐frequency details from the input image in the edited one. To train our framework we leverage a dataset with pairs of synthetic images rendered with physically‐based algorithms, and the corresponding crowd‐sourced ratings of high‐level perceptual attributes. We show that our material editing framework outperforms the state of the art, and showcase its applicability on synthetic images, in‐the‐wild real‐world photographs, and video sequences.

Abstract Single‐image appearance editing is a challenging task, traditionally requiring the estimation of additional scene properties such as geometry or illumination. Moreover, the exact interaction of light, shape and material reflectance that elicits a given perceptual impression is still not well understood. We present an image‐based editing method that allows to modify the material appearance of an object by increasing or decreasing high‐level perceptual attributes, using a single image as input. Our framework relies on a two‐step generative network, where the first step drives the change in appearance and the second produces an image with high‐frequency details. For training, we augment an existing material appearance dataset with perceptual judgements of high‐level attributes, collected through crowd‐sourced experiments, and build upon training strategies that circumvent the cumbersome need for original‐edited image pairs. We demonstrate the editing capabilities of our framework on a variety of inputs, both synthetic and real, using two common perceptual attributes (Glossy and Metallic), and validate the perception of appearance in our edited images through a user study.

We analyze the effect of motion in the perception of material appearance. First, we create a set of stimuli containing 72 realistic materials, rendered with varying degrees of linear motion blur. Then we launch a large-scale study on Mechanical Turk to rate a given set of perceptual attributes, such as brightness, roughness, or the perceived strength of reflections. Our statistical analysis shows that certain attributes undergo a significant change, varying appearance perception under motion. In addition, we further investigate the perception of brightness, for the particular cases of rubber and plastic materials. We create new stimuli, with ten different luminance levels and seven motion degrees. We launch a new user study to retrieve their perceived brightness. From the users’ judgements, we build two-dimensional maps showing how perceived brightness varies as a function of the luminance and motion of the material.

We present a model to measure the similarity in appearance between different materials, which correlates with human similarity judgments. We first create a database of 9,000 rendered images depicting objects with varying materials, shape and illumination. We then gather data on perceived similarity from crowdsourced experiments; our analysis of over 114,840 answers suggests that indeed a shared perception of appearance similarity exists. We feed this data to a deep learning architecture with a novel loss function, which learns a feature space for materials that correlates with such perceived appearance similarity. Our evaluation shows that our model outperforms existing metrics. Last, we demonstrate several applications enabled by our metric, including appearance-based search for material suggestions, database visualization, clustering and summarization, and gamut mapping.