Applying style transfer to a full 3D environment is a challenging task that has seen many developments since the advent of neural rendering. 3D Gaussian splatting (3DGS) has recently pushed further many limits of neural rendering in terms of training speed and reconstruction quality. This work introduces SGSST: Scaling Gaussian Splatting Style Transfer, an optimization-based method to apply style transfer to pretrained 3DGS scenes. We demonstrate that a new multiscale loss based on global neural statistics, that we name SOS for Simultaneously Optimized Scales, enables style transfer to ultra-high resolution 3D scenes. Not only SGSST pioneers 3D scene style transfer at such high image resolutions, it also produces superior visual quality as assessed by thorough qualitative, quantitative and perceptual comparisons.

Understanding materials properties depends largely on the ability to determine its components, and in particular its mineral phases. Powder X-ray diffraction (XRD) is a powerful tool for such purposes. This paper presents a Transformer-based vision model (ViT) for mineral phase identification, and proportion inference to quantify the mineral phases present in a material. Our analysis shows that the tokenization strategy is a critical step for XRD pattern analysis. The results obtained for both tasks are excellent and more robust than those obtained with a CNN. The proposed approach also makes it possible to introduce visualization tools for signal analysis, to better understand how information flows through the model and how data is classified or quantified.

Diffusion or score-based models recently showed high performance in image generation. They rely on a forward and a backward stochastic differential equations (SDE). The sampling of a data distribution is achieved by numerically solving the backward SDE or its associated flow ODE. Studying the convergence of these models necessitates to control four different types of error: the initialization error, the truncation error, the discretization error and the score approximation. In this paper, we theoretically study the behavior of diffusion models and their numerical implementation when the data distribution is Gaussian. Our first contribution is to derive the analytical solutions of the backward SDE and the probability flow ODE and to prove that these solutions and their discretizations are all Gaussian processes. Our second contribution is to compute the exact Wasserstein errors between the target and the numerically sampled distributions for any numerical scheme. This allows us to monitor convergence directly in the data space, while experimental works limit their empirical analysis to Inception features. An implementation of our code is available online.

International audience

Abstract. Super-resolution (SR) is the problem that consists in reconstructing images that have been degraded by a zoom-out operator. This is an ill-posed problem that does not have a unique solution, and numerical approaches rely on a prior on high-resolution images. While optimization-based methods are generally deterministic, with the rise of image generative models more and more interest has been given to stochastic SR, that is, sampling among all possible SR images associated with a given low-resolution input. In this paper, we construct an efficient, stable, and provably exact sampler for the stochastic SR of Gaussian microtextures. Even though our approach is limited regarding the scope of images it encompasses, our algorithm is competitive with deep learning state-of-the-art methods both in terms of perceptual metric and execution time when applied to microtextures. The framework of Gaussian microtextures also allows us to rigorously discuss the limitations of various reconstruction metrics to evaluate the efficiency of SR routines. An implementation of our algorithm is available at https://github.com/emilePi/Stochastic_SR_for_Gaussian_textures .

Abstract Neural style transfer (NST) is a deep learning technique that produces an unprecedentedly rich style transfer from a style image to a content image. It is particularly impressive when it comes to transferring style from a painting to an image. NST was originally achieved by solving an optimization problem to match the global statistics of the style image while preserving the local geometric features of the content image. The two main drawbacks of this original approach is that it is computationally expensive and that the resolution of the output images is limited by high GPU memory requirements. Many solutions have been proposed to both accelerate NST and produce images with larger size. However, our investigation shows that these accelerated methods all compromise the quality of the produced images in the context of painting style transfer. Indeed, transferring the style of a painting is a complex task involving features at different scales, from the color palette and compositional style to the fine brushstrokes and texture of the canvas. This paper provides a solution to solve the original global optimization for ultra‐high resolution (UHR) images, enabling multiscale NST at unprecedented image sizes. This is achieved by spatially localizing the computation of each forward and backward passes through the VGG network. Extensive qualitative and quantitative comparisons, as well as a perceptual study, show that our method produces style transfer of unmatched quality for such high‐resolution painting styles. By a careful comparison, we show that state‐of‐the‐art fast methods are still prone to artifacts, thus suggesting that fast painting style transfer remains an open problem.

Super-resolution (SR) is an ill-posed inverse problem which consists in proposing high-resolution images consistent with a given low-resolution one. While most SR algorithms are deterministic, stochastic SR deals with designing a stochastic sampler generating any realistic SR solution. The goal of this paper is to show that stochastic SR is a well-posed and solvable problem when restricting to Gaussian stationary textures. Using Gaussian conditional sampling and exploiting the stationarity assumption, we propose an efficient algorithm based on fast Fourier transform. We also demonstrate the practical relevance of the approach for SR with a reference image. Although limited to stationary microtextures, our approach compares favorably in terms of speed and visual quality to some state of the art methods designed for a larger class of images.

In this work, we study the problem of single-image super-resolution (SISR) of Sentinel-2 imagery. We show that thanks to its unique sensor specification, namely the inter-band shift and alias, that deep-learning methods are able to recover fine details. By training a model using a simple L1loss, results are free of hallucinated details. For this study, we build a dataset of pairs of images Sentinel-2/PlanetScope to train and evaluate our super-resolution (SR) model.

Determinantal point processes (DPPs) are probabilistic models of configurations that favor diversity or repulsion. They have recently gained influence in the machine learning community, mainly because of their ability to elegantly and efficiently subsample large sets of data. In this paper, we consider DPPs from an image processing perspective, meaning that the data we want to subsample are pixels or patches of a given image. To this end, our framework is discrete and finite. First, we adapt their basic definition and properties to DPPs defined on the pixels of an image, that we call determinantal pixel processes (DPixPs). We are mainly interested in the repulsion properties of such a process and we apply DPixPs to texture synthesis using shot noise models. Finally, we study DPPs on the set of patches of an image. Because of their repulsive property, DPPs provide a strong tool to subsample discrete distributions such as that of image patches.

Abstract This paper describes a novel approach for on demand volumetric texture synthesis based on a deep learning framework that allows for the generation of high‐quality three‐dimensional (3D) data at interactive rates. Based on a few example images of textures, a generative network is trained to synthesize coherent portions of solid textures of arbitrary sizes that reproduce the visual characteristics of the examples along some directions. To cope with memory limitations and computation complexity that are inherent to both high resolution and 3D processing on the GPU, only 2D textures referred to as ‘slices’ are generated during the training stage. These synthetic textures are compared to exemplar images via a perceptual loss function based on a pre‐trained deep network. The proposed network is very light (less than 100k parameters), therefore it only requires sustainable training (i.e. few hours) and is capable of very fast generation (around a second for 2563 voxels) on a single GPU. Integrated with a spatially seeded pseudo‐random number generator (PRNG) the proposed generator network directly returns a color value given a set of 3D coordinates. The synthesized volumes have good visual results that are at least equivalent to the state‐of‐the‐art patch‐based approaches. They are naturally seamlessly tileable and can be fully generated in parallel.

In this work we introduce a statistical framework in order to analyze the spatial redundancy in natural images. This notion of spatial redundancy must be defined locally. To do so, we define an auto-similarity function which, given one image, computes a dissimilarity measurement between patches. To derive a criterion for taking a decision on the similarity between two patches, we present an a contrario model. Namely, two patches are said to be similar if the associated dissimilarity measurement is unlikely to happen in a background model. Choosing Gaussian random fields as background models, we derive nonasymptotic expressions for the probability distribution function of similarity measurements. We present an algorithm in order to assess redundancy in natural images and discuss applications in denoising, periodicity analysis, and texture ranking.

Exemplar-based texture synthesis consists in producing new synthetic images which have the same perceptual characteristics as a given texture sample while exhibiting sufficient innovation (to avoid verbatim copy). In this paper, we propose to address this problem with a model obtained as local transformations of Gaussian random fields. The local transformations operate on $3 \times 3$ patches and are designed to solve a semi-discrete optimal transport problem in order to reimpose the patch distribution of the exemplar texture. The semi-discrete optimal transport problem is solved with a stochastic gradient algorithm, whose convergence speed is evaluated on several practical transport cases. After studying the properties of such transformed Gaussian random fields, we propose a multiscale extension of the model which aims at preserving the patch distribution of the exemplar texture at multiple scales. Experiments demonstrate that this multiscale model is able to synthesize structured textures while keeping several mathematical guarantees, and with low requirements in synthesis time and memory storage. In particular, a single patch optimal transport map is shown to be better than iterated nearest neighbor assignments in terms of statistical guarantees. Besides, once the model is estimated, the resulting synthesis algorithm is fast and highly parallel since it amounts to performing weighted nearest neighbor patch assignments at each scale.

Abstract Designing realistic noise patterns from scratch is hard. To solve this problem, recent contributions have proposed involved spectral analysis algorithms that enable procedural noise models to faithfully reproduce some class of textures. The aim of this paper is to propose the simplest and most efficient noise model that allows for the reproduction of any Gaussian texture. Texton noise is a simple sparse convolution noise that sums randomly scattered copies of a small bilinear texture called texton . We introduce an automatic algorithm to compute the texton associated with an input texture image that concentrates the input frequency content into the desired texton support. One of the main features of texton noise is that its evaluation only consists to sum 30 texture fetches on average. Consequently, texton noise generates Gaussian textures with an unprecedented evaluation speed for noise by example. A second main feature of texton noise is that it allows for high‐quality on‐the‐fly anisotropic filtering by simply invoking existing GPU hardware solutions for texture fetches. In addition, we demonstrate that texton noise can be applied on any surface using parameterization‐free surface noise and that it allows for noise mixing.

Abstract The realistic synthesis and rendering of film grain is a crucial goal for many amateur and professional photographers and film‐makers whose artistic works require the authentic feel of analogue photography. The objective of this work is to propose an algorithm that reproduces the visual aspect of film grain texture on any digital image. Previous approaches to this problem either propose unrealistic models or simply blend scanned images of film grain with the digital image, in which case the result is inevitably limited by the quality and resolution of the initial scan. In this work, we introduce a stochastic model to approximate the physical reality of film grain, and propose a resolution‐free rendering algorithm to simulate realistic film grain for any digital input image. By varying the parameters of this model, we can achieve a wide range of grain types. We demonstrate this by comparing our results with film grain examples from dedicated software, and show that our rendering results closely resemble these real film emulsions. In addition to realistic grain rendering, our resolution‐free algorithm allows for any desired zoom factor, even down to the scale of the microscopic grains themselves.

Inpainting consists in computing a plausible completion of missing parts of an image given the available content. In the restricted framework of texture images, the image can be seen as a realization of a random field model, which gives a stochastic formulation of image inpainting: on the masked exemplar one estimates a random texture model which can then be conditionally sampled in order to fill the hole. In this paper is proposed an instance of such stochastic inpainting methods, dealing in particular with the case of Gaussian textures. First, a simple procedure is proposed for estimating a Gaussian texture model based on a masked exemplar, which, although quite naive, gives sufficient results for our inpainting purpose. Next, the conditional sampling step is solved with the traditional algorithm for Gaussian conditional simulation. The main difficulty of this step is to solve a very large linear system, which, in the case of stationary Gaussian textures, can be done efficiently with a conjugate gradient descent (using a Fourier representation of the covariance operator). Several experiments show that the corresponding inpainting algorithm is able to inpaint large holes (of any shape) in a texture, with a reasonable computational time. Moreover, several comparisons illustrate that the proposed approach performs better on texture images than state-of-the-art inpainting methods.

Image inpainting consists in filling missing regions of an image by inferring from the surrounding content. In the case of texture images, inpainting can be formulated in terms of conditional simulation of a stochastic texture model. Many texture synthesis methods thus have been adapted to texture inpainting, but these methods do not offer theoretical guarantees since the conditional sampling is in general only approximate. Here we show that in the case of Gaussian textures, inpainting can be addressed with perfect conditional simulation relying on kriging estimation. We thus obtain a microtexture inpainting algorithm that is able to fill holes of any shape and size in an efficient manner while respecting exactly a stochastic model.

Abstract We present a novel shape‐aware method for synthesizing 2D and 3D discrete element textures consisting of collections of distinct vector graphics objects. Extending the long‐proven point process framework, we propose a shape process, a novel stochastic model based on spatial measurements that fully take into account the geometry of the elements. We demonstrate that our approach is well‐suited for discrete texture synthesis by example. Our model enables for both robust statistical parameter estimation and reliable output generation by Monte Carlo sampling. Our numerous experiments show that contrary to current state‐of‐the‐art techniques, our algorithm manages to capture anisotropic element distributions and systematically prevents undesirable collisions between objects.

Procedural noise is a fundamental tool in Computer Graphics. However, designing noise patterns is hard. In this paper, we present Gabor noise by example , a method to estimate the parameters of bandwidth-quantized Gabor noise, a procedural noise function that can generate noise with an arbitrary power spectrum, from exemplar Gaussian textures, a class of textures that is completely characterized by their power spectrum. More specifically, we introduce (i) bandwidth-quantized Gabor noise, a generalization of Gabor noise to arbitrary power spectra that enables robust parameter estimation and efficient procedural evaluation; (ii) a robust parameter estimation technique for quantized-bandwidth Gabor noise, that automatically decomposes the noisy power spectrum estimate of an exemplar into a sparse sum of Gaussians using non-negative basis pursuit denoising; and (iii) an efficient procedural evaluation scheme for bandwidth-quantized Gabor noise, that uses multi-grid evaluation and importance sampling of the kernel parameters. Gabor noise by example preserves the traditional advantages of procedural noise, including a compact representation and a fast on-the-fly evaluation, and is mathematically well-founded.

This paper explores the mathematical and algorithmic properties of two sample-based texture models: random phase noise (RPN) and asymptotic discrete spot noise (ADSN). These models permit to synthesize random phase textures. They arguably derive from linearized versions of two early Julesz texture discrimination theories. The ensuing mathematical analysis shows that, contrarily to some statements in the literature, RPN and ADSN are different stochastic processes. Nevertheless, numerous experiments also suggest that the textures obtained by these algorithms from identical samples are perceptually similar. The relevance of this study is enhanced by three technical contributions providing solutions to obstacles that prevented the use of RPN or ADSN to emulate textures. First, RPN and ADSN algorithms are extended to color images. Second, a preprocessing is proposed to avoid artifacts due to the nonperiodicity of real-world texture samples. Finally, the method is extended to synthesize textures with arbitrary size from a given sample.