Karol Myszkowski

Lingyan Ruan, Mojtaba Bemana, Hans-peter Seidel et al.

Image deblurring aims to recover the latent sharp image from its blurry counterpart and has a wide range of applications in computer vision. The Convolution Neural Networks (CNNs) have performed well in this domain for many years, and until recently an alternative network architecture, namely Transformer, has demonstrated even stronger performance. One can attribute its superiority to the multi-head self-attention (MHSA) mechanism, which offers a larger receptive field and better input content adaptability than CNNs. However, as MHSA demands high computational costs that grow quadratically with respect to the input resolution, it becomes impractical for high-resolution image deblurring tasks. In this work, we propose a unified lightweight CNN network that features a large effective receptive field (ERF) and demonstrates comparable or even better performance than Transformers while bearing less computational costs. Our key design is an efficient CNN block dubbed LaKD, equipped with a large kernel depth-wise convolution and spatial-channel mixing structure, attaining comparable or larger ERF than Transformers but with a smaller parameter scale. Specifically, we achieve +0.17dB / +0.43dB PSNR over the state-of-the-art Restormer on defocus / motion deblurring benchmark datasets with 32% fewer parameters and 39% fewer MACs. Extensive experiments demonstrate the superior performance of our network and the effectiveness of each module. Furthermore, we propose a compact and intuitive ERFMeter metric that quantitatively characterizes ERF, and shows a high correlation to the network performance. We hope this work can inspire the research community to further explore the pros and cons of CNN and Transformer architectures beyond image deblurring tasks.

Chao Wang, Ana Serrano, Xingang Pan et al.

Most in-the-wild images are stored in Low Dynamic Range (LDR) form, serving as a partial observation of the High Dynamic Range (HDR) visual world. Despite limited dynamic range, these LDR images are often captured with different exposures, implicitly containing information about the underlying HDR image distribution. Inspired by this intuition, in this work we present, to the best of our knowledge, the first method for learning a generative model of HDR images from in-the-wild LDR image collections in a fully unsupervised manner. The key idea is to train a generative adversarial network (GAN) to generate HDR images which, when projected to LDR under various exposures, are indistinguishable from real LDR images. The projection from HDR to LDR is achieved via a camera model that captures the stochasticity in exposure and camera response function. Experiments show that our method GlowGAN can synthesize photorealistic HDR images in many challenging cases such as landscapes, lightning, or windows, where previous supervised generative models produce overexposed images. We further demonstrate the new application of unsupervised inverse tone mapping (ITM) enabled by GlowGAN. Our ITM method does not need HDR images or paired multi-exposure images for training, yet it reconstructs more plausible information for overexposed regions than state-of-the-art supervised learning models trained on such data.

Uğur Çoğalan, Mojtaba Bemana, Hans-Peter Seidel et al.

Video frame interpolation (VFI) enables many important applications that might involve the temporal domain, such as slow motion playback, or the spatial domain, such as stop motion sequences. We are focusing on the former task, where one of the key challenges is handling high dynamic range (HDR) scenes in the presence of complex motion. To this end, we explore possible advantages of dual-exposure sensors that readily provide sharp short and blurry long exposures that are spatially registered and whose ends are temporally aligned. This way, motion blur registers temporally continuous information on the scene motion that, combined with the sharp reference, enables more precise motion sampling within a single camera shot. We demonstrate that this facilitates a more complex motion reconstruction in the VFI task, as well as HDR frame reconstruction that so far has been considered only for the originally captured frames, not in-between interpolated frames. We design a neural network trained in these tasks that clearly outperforms existing solutions. We also propose a metric for scene motion complexity that provides important insights into the performance of VFI methods at the test time.

Martin Balint, Karol Myszkowski, Hans-Peter Seidel et al.

When dealing with difficult inverse problems such as inverse rendering, using Monte Carlo estimated gradients to optimise parameters can slow down convergence due to variance. Averaging many gradient samples in each iteration reduces this variance trivially. However, for problems that require thousands of optimisation iterations, the computational cost of this approach rises quickly. We derive a theoretical framework for interleaving sampling and optimisation. We update and reuse past samples with low-variance finite-difference estimators that describe the change in the estimated gradients between each iteration. By combining proportional and finite-difference samples, we continuously reduce the variance of our novel gradient meta-estimators throughout the optimisation process. We investigate how our estimator interlinks with Adam and derive a stable combination. We implement our method for inverse path tracing and demonstrate how our estimator speeds up convergence on difficult optimisation tasks.

Mojtaba Bemana, Thomas Leimkühler, Karol Myszkowski et al.

We demonstrate generating HDR images using the concerted action of multiple black-box, pre-trained LDR image diffusion models. Relying on a pre-trained LDR generative diffusion models is vital 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 "exposure brackets'', to produce a single HDR image. We operate multiple denoising processes to generate multiple LDR brackets that together form a valid HDR result. The key to making this work is to introduce a consistency term into the diffusion process to couple the brackets such that they agree across the exposure range they share while accounting for possible differences due to the quantization error. We demonstrate state-of-the-art unconditional and conditional or restoration-type (LDR2HDR) generative modeling results, yet in HDR.

Chao Wang, Zhihao Xia, Thomas Leimkuehler et al.

While consumer displays increasingly support more than 10 stops of dynamic range, most image assets such as internet photographs and generative AI content remain limited to 8-bit low dynamic range (LDR), constraining their utility across high dynamic range (HDR) applications. Currently, no generative model can produce high-bit, high-dynamic range content in a generalizable way. Existing LDR-to-HDR conversion methods often struggle to produce photorealistic details and physically-plausible dynamic range in the clipped areas. We introduce LEDiff, a method that enables a generative model with HDR content generation through latent space fusion inspired by image-space exposure fusion techniques. It also functions as an LDR-to-HDR converter, expanding the dynamic range of existing low-dynamic range images. Our approach uses a small HDR dataset to enable a pretrained diffusion model to recover detail and dynamic range in clipped highlights and shadows. LEDiff brings HDR capabilities to existing generative models and converts any LDR image to HDR, creating photorealistic HDR outputs for image generation, image-based lighting (HDR environment map generation), and photographic effects such as depth of field simulation, where linear HDR data is essential for realistic quality.

Uğur Çoğalan, Mojtaba Bemana, Karol Myszkowski et al.

We present MILO (Metric for Image- and Latent-space Optimization), a lightweight, multiscale, perceptual metric for full-reference image quality assessment (FR-IQA). MILO is trained using pseudo-MOS (Mean Opinion Score) supervision, in which reproducible distortions are applied to diverse images and scored via an ensemble of recent quality metrics that account for visual masking effects. This approach enables accurate learning without requiring large-scale human-labeled datasets. Despite its compact architecture, MILO outperforms existing metrics across standard FR-IQA benchmarks and offers fast inference suitable for real-time applications. Beyond quality prediction, we demonstrate the utility of MILO as a perceptual loss in both image and latent domains. In particular, we show that spatial masking modeled by MILO, when applied to latent representations from a VAE encoder within Stable Diffusion, enables efficient and perceptually aligned optimization. By combining spatial masking with a curriculum learning strategy, we first process perceptually less relevant regions before progressively shifting the optimization to more visually distorted areas. This strategy leads to significantly improved performance in tasks like denoising, super-resolution, and face restoration, while also reducing computational overhead. MILO thus functions as both a state-of-the-art image quality metric and as a practical tool for perceptual optimization in generative pipelines.

Krzysztof Wolski, Adarsh Djeacoumar, Alireza Javanmardi et al.

The real world exhibits rich structure and detail across many scales of observation. It is difficult, however, to capture and represent a broad spectrum of scales using ordinary images. We devise a novel paradigm for learning a representation that captures an orders-of-magnitude variety of scales from an unstructured collection of ordinary images. We treat this collection as a distribution of scale-space slices to be learned using adversarial training, and additionally enforce coherency across slices. Our approach relies on a multiscale generator with carefully injected procedural frequency content, which allows to interactively explore the emerging continuous scale space. Training across vastly different scales poses challenges regarding stability, which we tackle using a supervision scheme that involves careful sampling of scales. We show that our generator can be used as a multiscale generative model, and for reconstructions of scale spaces from unstructured patches. Significantly outperforming the state of the art, we demonstrate zoom-in factors of up to 256x at high quality and scale consistency.

Chao Wang, Krzysztof Wolski, Bernhard Kerbl et al.

Radiance field methods represent the state of the art in reconstructing complex scenes from multi-view photos. However, these reconstructions often suffer from one or both of the following limitations: First, they typically represent scenes in low dynamic range (LDR), which restricts their use to evenly lit environments and hinders immersive viewing experiences. Secondly, their reliance on a pinhole camera model, assuming all scene elements are in focus in the input images, presents practical challenges and complicates refocusing during novel-view synthesis. Addressing these limitations, we present a lightweight method based on 3D Gaussian Splatting that utilizes multi-view LDR images of a scene with varying exposure times, apertures, and focus distances as input to reconstruct a high-dynamic-range (HDR) radiance field. By incorporating analytical convolutions of Gaussians based on a thin-lens camera model as well as a tonemapping module, our reconstructions enable the rendering of HDR content with flexible refocusing capabilities. We demonstrate that our combined treatment of HDR and depth of field facilitates real-time cinematic rendering, outperforming the state of the art.

Uğur Çoğalan, Mojtaba Bemana, Hans-Peter Seidel et al.

Full-reference image quality metrics (FR-IQMs) aim to measure the visual differences between a pair of reference and distorted images, with the goal of accurately predicting human judgments. However, existing FR-IQMs, including traditional ones like PSNR and SSIM and even perceptual ones such as HDR-VDP, LPIPS, and DISTS, still fall short in capturing the complexities and nuances of human perception. In this work, rather than devising a novel IQM model, we seek to improve upon the perceptual quality of existing FR-IQM methods. We achieve this by considering visual masking, an important characteristic of the human visual system that changes its sensitivity to distortions as a function of local image content. Specifically, for a given FR-IQM metric, we propose to predict a visual masking model that modulates reference and distorted images in a way that penalizes the visual errors based on their visibility. Since the ground truth visual masks are difficult to obtain, we demonstrate how they can be derived in a self-supervised manner solely based on mean opinion scores (MOS) collected from an FR-IQM dataset. Our approach results in enhanced FR-IQM metrics that are more in line with human prediction both visually and quantitatively.

Mojtaba Bemana, Karol Myszkowski, Jeppe Revall Frisvad et al.

We tackle the problem of generating novel-view images from collections of 2D images showing refractive and reflective objects. Current solutions assume opaque or transparent light transport along straight paths following the emission-absorption model. Instead, we optimize for a field of 3D-varying Index of Refraction (IoR) and trace light through it that bends toward the spatial gradients of said IoR according to the laws of eikonal light transport.

Chao Wang, Bin Chen, Hans-Peter Seidel et al.

High Dynamic Range (HDR) content is becoming ubiquitous due to the rapid development of capture technologies. Nevertheless, the dynamic range of common display devices is still limited, therefore tone mapping (TM) remains a key challenge for image visualization. Recent work has demonstrated that neural networks can achieve remarkable performance in this task when compared to traditional methods, however, the quality of the results of these learning-based methods is limited by the training data. Most existing works use as training set a curated selection of best-performing results from existing traditional tone mapping operators (often guided by a quality metric), therefore, the quality of newly generated results is fundamentally limited by the performance of such operators. This quality might be even further limited by the pool of HDR content that is used for training. In this work we propose a learning-based self-supervised tone mapping operator that is trained at test time specifically for each HDR image and does not need any data labeling. The key novelty of our approach is a carefully designed loss function built upon fundamental knowledge on contrast perception that allows for directly comparing the content in the HDR and tone mapped images. We achieve this goal by reformulating classic VGG feature maps into feature contrast maps that normalize local feature differences by their average magnitude in a local neighborhood, allowing our loss to account for contrast masking effects. We perform extensive ablation studies and exploration of parameters and demonstrate that our solution outperforms existing approaches with a single set of fixed parameters, as confirmed by both objective and subjective metrics.

Luca Surace, Marek Wernikowski, Cara Tursun et al.

A foveated image can be entirely reconstructed from a sparse set of samples distributed according to the retinal sensitivity of the human visual system, which rapidly decreases with increasing eccentricity. The use of Generative Adversarial Networks has recently been shown to be a promising solution for such a task, as they can successfully hallucinate missing image information. As in the case of other supervised learning approaches, the definition of the loss function and the training strategy heavily influence the quality of the output. In this work,we consider the problem of efficiently guiding the training of foveated reconstruction techniques such that they are more aware of the capabilities and limitations of the human visual system, and thus can reconstruct visually important image features. Our primary goal is to make the training procedure less sensitive to distortions that humans cannot detect and focus on penalizing perceptually important artifacts. Given the nature of GAN-based solutions, we focus on the sensitivity of human vision to hallucination in case of input samples with different densities. We propose psychophysical experiments, a dataset, and a procedure for training foveated image reconstruction. The proposed strategy renders the generator network flexible by penalizing only perceptually important deviations in the output. As a result, the method emphasized the recovery of perceptually important image features. We evaluated our strategy and compared it with alternative solutions by using a newly trained objective metric, a recent foveated video quality metric, and user experiments. Our evaluations revealed significant improvements in the perceived image reconstruction quality compared with the standard GAN-based training approach.

Uğur Çoğalan, Mojtaba Bemana, Karol Myszkowski et al.

We seek to reconstruct sharp and noise-free high-dynamic range (HDR) video from a dual-exposure sensor that records different low-dynamic range (LDR) information in different pixel columns: Odd columns provide low-exposure, sharp, but noisy information; even columns complement this with less noisy, high-exposure, but motion-blurred data. Previous LDR work learns to deblur and denoise (DISTORTED->CLEAN) supervised by pairs of CLEAN and DISTORTED images. Regrettably, capturing DISTORTED sensor readings is time-consuming; as well, there is a lack of CLEAN HDR videos. We suggest a method to overcome those two limitations. First, we learn a different function instead: CLEAN->DISTORTED, which generates samples containing correlated pixel noise, and row and column noise, as well as motion blur from a low number of CLEAN sensor readings. Second, as there is not enough CLEAN HDR video available, we devise a method to learn from LDR video in-stead. Our approach compares favorably to several strong baselines, and can boost existing methods when they are re-trained on our data. Combined with spatial and temporal super-resolution, it enables applications such as re-lighting with low noise or blur.

Mojtaba Bemana, Karol Myszkowski, Hans-Peter Seidel et al.

We suggest to represent an X-Field -a set of 2D images taken across different view, time or illumination conditions, i.e., video, light field, reflectance fields or combinations thereof-by learning a neural network (NN) to map their view, time or light coordinates to 2D images. Executing this NN at new coordinates results in joint view, time or light interpolation. The key idea to make this workable is a NN that already knows the "basic tricks" of graphics (lighting, 3D projection, occlusion) in a hard-coded and differentiable form. The NN represents the input to that rendering as an implicit map, that for any view, time, or light coordinate and for any pixel can quantify how it will move if view, time or light coordinates change (Jacobian of pixel position with respect to view, time, illumination, etc.). Our X-Field representation is trained for one scene within minutes, leading to a compact set of trainable parameters and hence real-time navigation in view, time and illumination.

Mojtaba Bemana, Karol Myszkowski, Hans-Peter Seidel et al.

We suggest representing light field (LF) videos as "one-off" neural networks (NN), i.e., a learned mapping from view-plus-time coordinates to high-resolution color values, trained on sparse views. Initially, this sounds like a bad idea for three main reasons: First, a NN LF will likely have less quality than a same-sized pixel basis representation. Second, only few training data, e.g., 9 exemplars per frame are available for sparse LF videos. Third, there is no generalization across LFs, but across view and time instead. Consequently, a network needs to be trained for each LF video. Surprisingly, these problems can turn into substantial advantages: Other than the linear pixel basis, a NN has to come up with a compact, non-linear i.e., more intelligent, explanation of color, conditioned on the sparse view and time coordinates. As observed for many NN however, this representation now is interpolatable: if the image output for sparse view coordinates is plausible, it is for all intermediate, continuous coordinates as well. Our specific network architecture involves a differentiable occlusion-aware warping step, which leads to a compact set of trainable parameters and consequently fast learning and fast execution.

Vamsi Kiran Adhikarla, Marek Vinkler, Denis Sumin et al.

Light fields become a popular representation of three dimensional scenes, and there is interest in their processing, resampling, and compression. As those operations often result in loss of quality, there is a need to quantify it. In this work, we collect a new dataset of dense reference and distorted light fields as well as the corresponding quality scores which are scaled in perceptual units. The scores were acquired in a subjective experiment using an interactive light-field viewing setup. The dataset contains typical artifacts that occur in light-field processing chain due to light-field reconstruction, multi-view compression, and limitations of automultiscopic displays. We test a number of existing objective quality metrics to determine how well they can predict the quality of light fields. We find that the existing image quality metrics provide good measures of light-field quality, but require dense reference light- fields for optimal performance. For more complex tasks of comparing two distorted light fields, their performance drops significantly, which reveals the need for new, light-field-specific metrics.