Exposure Diffusion: HDR Image Generation by Consistent LDR denoising

We demonstrate generating high-dynamic range (HDR) images using the concerted action of multiple black-box, pre-trained low-dynamic range (LDR) image diffusion models. Common diffusion models are not HDR as, first, there is no sufficiently large HDR image dataset available to re-train them, and second, even if it was, re-training such models is impossible for most compute budgets. Instead, we seek inspiration from the HDR image capture literature that traditionally fuses sets of LDR images, called "brackets", to produce a single HDR image. We operate multiple denoising processes to generate multiple LDR brackets that together form a valid HDR result. To this end, we introduce an exposure consistency term into the diffusion process to couple the brackets such that they agree across the exposure range they share. We demonstrate HDR versions of state-of-the-art unconditional and conditional as well as restoration-type (LDR2HDR) generative modeling.

NeRF Analogies: Example-Based Visual Attribute Transfer for NeRFs

A Neural Radiance Field (NeRF) encodes the specific relation of 3D geometry and appearance of a scene. We here ask the question whether we can transfer the appearance from a source NeRF onto a target 3D geometry in a semantically meaningful way, such that the resulting new NeRF retains the target geometry but has an appearance that is an analogy to the source NeRF. To this end, we generalize classic image analogies from 2D images to NeRFs. We leverage correspondence transfer along semantic affinity that is driven by semantic features from large, pre-trained 2D image models to achieve multi-view consistent appearance transfer. Our method allows exploring the mix-and-match product space of 3D geometry and appearance. We show that our method outperforms traditional stylization-based methods and that a large majority of users prefer our method over several typical baselines.

4D-ONIX: A deep learning approach for reconstructing 3D movies from sparse X-ray projections

The X-ray flux provided by X-ray free-electron lasers and storage rings offers new spatiotemporal possibilities to study in-situ and operando dynamics, even using single pulses of such facilities. X-ray Multi-Projection Imaging (XMPI) is a novel technique that enables volumetric information using single pulses of such facilities and avoids centrifugal forces induced by state-of-the-art time-resolved 3D methods such as time-resolved tomography. As a result, XMPI can acquire 3D movies (4D) at least three orders of magnitude faster than current methods. However, it is exceptionally challenging to reconstruct 4D from highly sparse projections as acquired by XMPI with current algorithms. Here, we present 4D-ONIX, a Deep Learning (DL)-based approach that learns to reconstruct 3D movies (4D) from an extremely limited number of projections. It combines the computational physical model of X-ray interaction with matter and state-of-the-art DL methods. We demonstrate the potential of 4D-ONIX to generate high-quality 4D by generalizing over multiple experiments with only two projections per timestamp for binary droplet collisions. We envision that 4D-ONIX will become an enabling tool for 4D analysis, offering new spatiotemporal resolutions to study processes not possible before.

Neural Bounding

Bounding volumes are an established concept in computer graphics and vision tasks but have seen little change since their early inception. In this work, we study the use of neural networks as bounding volumes. Our key observation is that bounding, which so far has primarily been considered a problem of computational geometry, can be redefined as a problem of learning to classify space into free and empty. This learning-based approach is particularly advantageous in high-dimensional spaces, such as animated scenes with complex queries, where neural networks are known to excel. However, unlocking neural bounding requires a twist: allowing -- but also limiting -- false positives, while ensuring that the number of false negatives is strictly zero. We enable such tight and conservative results using a dynamically-weighted asymmetric loss function. Our results show that our neural bounding produces up to an order of magnitude fewer false positives than traditional methods.

Zero Grads Ever Given: Learning Local Surrogate Losses for Non-Differentiable Graphics

Gradient-based optimization is now ubiquitous across graphics, but unfortunately can not be applied to problems with undefined or zero gradients. To circumvent this issue, the loss function can be manually replaced by a "surrogate" that has similar minima but is differentiable. Our proposed framework, ZeroGrads, automates this process by learning a neural approximation of the objective function, the surrogate, which in turn can be used to differentiate through arbitrary black-box graphics pipelines. We train the surrogate on an actively smoothed version of the objective and encourage locality, focusing the surrogate's capacity on what matters at the current training episode. The fitting is performed online, alongside the parameter optimization, and self-supervised, without pre-computed data or pre-trained models. As sampling the objective is expensive (it requires a full rendering or simulator run), we devise an efficient sampling scheme that allows for tractable run-times and competitive performance at little overhead. We demonstrate optimizing diverse non-convex, non-differentiable black-box problems in graphics, such as visibility in rendering, discrete parameter spaces in procedural modelling or optimal control in physics-driven animation. In contrast to more traditional algorithms, our approach scales well to higher dimensions, which we demonstrate on problems with up to 35k interlinked variables.

Megahertz X-ray Multi-projection imaging

X-ray time-resolved tomography is one of the most popular X-ray techniques to probe dynamics in three dimensions (3D). Recent developments in time-resolved tomography opened the possibility of recording kilohertz-rate 3D movies. However, tomography requires rotating the sample with respect to the X-ray beam, which prevents characterization of faster structural dynamics. Here, we present megahertz (MHz) X-ray multi-projection imaging (MHz-XMPI), a technique capable of recording volumetric information at MHz rates and micrometer resolution without scanning the sample. We achieved this by harnessing the unique megahertz pulse structure and intensity of the European X-ray Free-electron Laser with a combination of novel detection and reconstruction approaches that do not require sample rotations. Our approach enables generating multiple X-ray probes that simultaneously record several angular projections for each pulse in the megahertz pulse burst. We provide a proof-of-concept demonstration of the MHz-XMPI technique's capability to probe 4D (3D+time) information on stochastic phenomena and non-reproducible processes three orders of magnitude faster than state-of-the-art time-resolved X-ray tomography, by generating 3D movies of binary droplet collisions. We anticipate that MHz-XMPI will enable in-situ and operando studies that were impossible before, either due to the lack of temporal resolution or because the systems were opaque (such as for MHz imaging based on optical microscopy).

Neural Field Convolutions by Repeated Differentiation

Neural fields are evolving towards a general-purpose continuous representation for visual computing. Yet, despite their numerous appealing properties, they are hardly amenable to signal processing. As a remedy, we present a method to perform general continuous convolutions with general continuous signals such as neural fields. Observing that piecewise polynomial kernels reduce to a sparse set of Dirac deltas after repeated differentiation, we leverage convolution identities and train a repeated integral field to efficiently execute large-scale convolutions. We demonstrate our approach on a variety of data modalities and spatially-varying kernels.

Plateau-free Differentiable Path Tracing

Current differentiable renderers provide light transport gradients with respect to arbitrary scene parameters. However, the mere existence of these gradients does not guarantee useful update steps in an optimization. Instead, inverse rendering might not converge due to inherent plateaus, i.e., regions of zero gradient, in the objective function. We propose to alleviate this by convolving the high-dimensional rendering function that maps scene parameters to images with an additional kernel that blurs the parameter space. We describe two Monte Carlo estimators to compute plateau-free gradients efficiently, i.e., with low variance, and show that these translate into net-gains in optimization error and runtime performance. Our approach is a straightforward extension to both black-box and differentiable renderers and enables optimization of problems with intricate light transport, such as caustics or global illumination, that existing differentiable renderers do not converge on.

3inGAN: Learning a 3D Generative Model from Images of a Self-similar Scene

We introduce 3inGAN, an unconditional 3D generative model trained from 2D images of a single self-similar 3D scene. Such a model can be used to produce 3D "remixes" of a given scene, by mapping spatial latent codes into a 3D volumetric representation, which can subsequently be rendered from arbitrary views using physically based volume rendering. By construction, the generated scenes remain view-consistent across arbitrary camera configurations, without any flickering or spatio-temporal artifacts. During training, we employ a combination of 2D, obtained through differentiable volume tracing, and 3D Generative Adversarial Network (GAN) losses, across multiple scales, enforcing realism on both its 3D structure and the 2D renderings. We show results on semi-stochastic scenes of varying scale and complexity, obtained from real and synthetic sources. We demonstrate, for the first time, the feasibility of learning plausible view-consistent 3D scene variations from a single exemplar scene and provide qualitative and quantitative comparisons against recent related methods.

Learning to Rasterize Differentiable

Differentiable rasterization changes the common formulation of primitive rasterization -- which has zero gradients almost everywhere, due to discontinuous edges and occlusion -- to an alternative one, which is not subject to this limitation and has similar optima. These alternative versions in general are ''soft'' versions of the original one. Unfortunately, it is not clear, what exact way of softening will provide the best performance in terms of converging the most reliability to a desired goal. Previous work has analyzed and compared several combinations of softening. In this work, we take it a step further and, instead of making a combinatorical choice of softening operations, parametrize the continuous space of all softening operations. We study meta-learning a parametric S-shape curve as well as an MLP over a set of inverse rendering tasks, so that it generalizes to new and unseen differentiable rendering tasks with optimal softness.