We propose to replace vector quantization (VQ) in the latent representation of VQ-VAEs with a simple scheme termed finite scalar quantization (FSQ), where we project the VAE representation down to a few dimensions (typically less than 10). Each dimension is quantized to a small set of fixed values, leading to an (implicit) codebook given by the product of these sets. By appropriately choosing the number of dimensions and values each dimension can take, we obtain the same codebook size as in VQ. On top of such discrete representations, we can train the same models that have been trained on VQ-VAE representations. For example, autoregressive and masked transformer models for image generation, multimodal generation, and dense prediction computer vision tasks. Concretely, we employ FSQ with MaskGIT for image generation, and with UViM for depth estimation, colorization, and panoptic segmentation. Despite the much simpler design of FSQ, we obtain competitive performance in all these tasks. We emphasize that FSQ does not suffer from codebook collapse and does not need the complex machinery employed in VQ (commitment losses, codebook reseeding, code splitting, entropy penalties, etc.) to learn expressive discrete representations.

Video understanding has witnessed significant progress with recent video foundation models demonstrating strong performance owing to self-supervised pre-training objectives; Masked Autoencoders (MAE) being the design of choice. Nevertheless, the majority of prior works that leverage MAE pre-training have focused on relatively short video representations (16 / 32 frames in length) largely due to hardware memory and compute limitations that scale poorly with video length due to the dense memory-intensive self-attention decoding. One natural strategy to address these challenges is to subsample tokens to reconstruct during decoding (or decoder masking). In this work, we propose an effective strategy for prioritizing tokens which allows training on longer video sequences (128 frames) and gets better performance than, more typical, random and uniform masking strategies. The core of our approach is an adaptive decoder masking strategy that prioritizes the most important tokens and uses quantized tokens as reconstruction objectives. Our adaptive strategy leverages a powerful MAGVIT-based tokenizer that jointly learns the tokens and their priority. We validate our design choices through exhaustive ablations and observe improved performance of the resulting long-video (128 frames) encoders over short-video (32 frames) counterparts. With our long-video masked autoencoder (LVMAE) strategy, we surpass state-of-the-art on Diving48 by 3.9 points and EPIC-Kitchens-100 verb classification by 2.5 points while relying on a simple core architecture and video-only pre-training (unlike some of the prior works that require millions of labeled video-text pairs or specialized encoders).

By optimizing the rate-distortion-realism trade-off, generative compression approaches produce detailed, realistic images, even at low bit rates, instead of the blurry reconstructions produced by rate-distortion optimized models. However, previous methods do not explicitly control how much detail is synthesized, which results in a common criticism of these methods: users might be worried that a misleading reconstruction far from the input image is generated. In this work, we alleviate these concerns by training a decoder that can bridge the two regimes and navigate the distortion-realism tradeoff. From a single compressed representation, the receiver can decide to either reconstruct a low mean squared error reconstruction that is close to the input, a realistic reconstruction with high perceptual quality, or anything in between. With our method, we set a new state-of-the-art in distortion-realism, pushing the frontier of achievable distortion-realism pairs, i.e., our method achieves better distortions at high realism and better realism at low distortion than ever before.

We show how bidirectional transformers trained for masked token prediction can be applied to neural image compression to achieve state-of-the-art results. Such models were previously used for image generation by progressivly sampling groups of masked tokens according to uncertainty-adaptive schedules. Unlike these works, we demonstrate that predefined, deterministic schedules perform as well or better for image compression. This insight allows us to use masked attention during training in addition to masked inputs, and activation caching during inference, to significantly speed up our models (≈4× higher inference speed) at a small increase in bitrate.1

We show how transformers can be used to vastly simplify neural video compression. Previous methods have been relying on an increasing number of architectural biases and priors, including motion prediction and warping operations, resulting in complex models. Instead, we independently map input frames to representations and use a transformer to model their dependencies, letting it predict the distribution of future representations given the past. The resulting video compression transformer outperforms previous methods on standard video compression data sets. Experiments on synthetic data show that our model learns to handle complex motion patterns such as panning, blurring and fading purely from data. Our approach is easy to implement, and we release code to facilitate future research.

Despite considerable progress on end-to-end optimized deep networks for image compression, video coding remains a challenging task. Recently proposed methods for learned video compression use optical flow and bilinear warping for motion compensation and show competitive rate-distortion performance relative to hand-engineered codecs like H.264 and HEVC. However, these learning-based methods rely on complex architectures and training schemes including the use of pre-trained optical flow networks, sequential training of sub-networks, adaptive rate control, and buffering intermediate reconstructions to disk during training. In this paper, we show that a generalized warping operator that better handles common failure cases, e.g. disocclusions and fast motion, can provide competitive compression results with a greatly simplified model and training procedure. Specifically, we propose scale-space flow, an intuitive generalization of optical flow that adds a scale parameter to allow the network to better model uncertainty. Our experiments show that a low-latency video compression model (no B-frames) using scale-space flow for motion compensation can outperform analogous state-of-the art learned video compression models while being trained using a much simpler procedure and without any pre-trained optical flow networks.

A popular approach to learning encoders for lossy compression is to use additive uniform noise during training as a differentiable approximation to test-time quantization. We demonstrate that a uniform noise channel can also be implemented at test time using universal quantization (Ziv, 1985). This allows us to eliminate the mismatch between training and test phases while maintaining a completely differentiable loss function. Implementing the uniform noise channel is a special case of the more general problem of communicating a sample, which we prove is computationally hard if we do not make assumptions about its distribution. However, the uniform special case is efficient as well as easy to implement and thus of great interest from a practical point of view. Finally, we show that quantization can be obtained as a limiting case of a soft quantizer applied to the uniform noise channel, bridging compression with and without quantization.

We extensively study how to combine Generative Adversarial Networks and learned compression to obtain a state-of-the-art generative lossy compression system. In particular, we investigate normalization layers, generator and discriminator architectures, training strategies, as well as perceptual losses. In contrast to previous work, i) we obtain visually pleasing reconstructions that are perceptually similar to the input, ii) we operate in a broad range of bitrates, and iii) our approach can be applied to high-resolution images. We bridge the gap between rate-distortion-perception theory and practice by evaluating our approach both quantitatively with various perceptual metrics, and with a user study. The study shows that our method is preferred to previous approaches even if they use more than 2x the bitrate.

We address interactive full image annotation, where the goal is to accurately segment all object and stuff regions in an image. We propose an interactive, scribble-based annotation framework which operates on the whole image to produce segmentations for all regions. This enables sharing scribble corrections across regions, and allows the annotator to focus on the largest errors made by the machine across the whole image. To realize this, we adapt Mask-RCNN [22] into a fast interactive segmentation framework and introduce an instance-aware loss measured at the pixel-level in the full image canvas, which lets predictions for nearby regions properly compete for space. Finally, we compare to interactive single object segmentation on the COCO panoptic dataset [11, 27, 34]. We demonstrate that our interactive full image segmentation approach leads to a 5% IoU gain, reaching 90% IoU at a budget of four extreme clicks and four corrective scribbles per region.

We propose the first practical learned lossless image compression system, L3C, and show that it outperforms the popular engineered codecs, PNG, WebP and JPEG 2000. At the core of our method is a fully parallelizable hierarchical probabilistic model for adaptive entropy coding which is optimized end-to-end for the compression task. In contrast to recent autoregressive discrete probabilistic models such as PixelCNN, our method i) models the image distribution jointly with learned auxiliary representations instead of exclusively modeling the image distribution in RGB space, and ii) only requires three forward-passes to predict all pixel probabilities instead of one for each pixel. As a result, L3C obtains over two orders of magnitude speedups when sampling compared to the fastest PixelCNN variant (Multiscale-PixelCNN). Furthermore, we find that learning the auxiliary representation is crucial and outperforms predefined auxiliary representations such as an RGB pyramid significantly.

We present a learned image compression system based on GANs, operating at extremely low bitrates. Our proposed framework combines an encoder, decoder/generator and a multi-scale discriminator, which we train jointly for a generative learned compression objective. The model synthesizes details it cannot afford to store, obtaining visually pleasing results at bitrates where previous methods fail and show strong artifacts. Furthermore, if a semantic label map of the original image is available, our method can fully synthesize unimportant regions in the decoded image such as streets and trees from the label map, proportionally reducing the storage cost. A user study confirms that for low bitrates, our approach is preferred to state-of-the-art methods, even when they use more than double the bits.

Deep Neural Networks trained as image auto-encoders have recently emerged as a promising direction for advancing the state-of-the-art in image compression. The key challenge in learning such networks is twofold: To deal with quantization, and to control the trade-off between reconstruction error (distortion) and entropy (rate) of the latent image representation. In this paper, we focus on the latter challenge and propose a new technique to navigate the rate-distortion trade-off for an image compression auto-encoder. The main idea is to directly model the entropy of the latent representation by using a context model: A 3D-CNN which learns a conditional probability model of the latent distribution of the auto-encoder. During training, the auto-encoder makes use of the context model to estimate the entropy of its representation, and the context model is concurrently updated to learn the dependencies between the symbols in the latent representation. Our experiments show that this approach, when measured in MS-SSIM, yields a state-of-the-art image compression system based on a simple convolutional auto-encoder.

Designing a logo for a new brand is a lengthy and tedious back-and-forth process between a designer and a client. In this paper we explore to what extent machine learning can solve the creative task of the designer. For this, we build a dataset - LLD - of 600k+ logos crawled from the world wide web. Training Generative Adversarial Networks (GANs) for logo synthesis on such multi-modal data is not straightforward and results in mode collapse for some state-of-the-art methods. We propose the use of synthetic labels obtained through clustering to disentangle and stabilize GAN training, and validate this approach on CIFAR-10 and ImageNet-small to demonstrate its generality. We are able to generate a high diversity of plausible logos and demonstrate latent space exploration techniques to ease the logo design task in an interactive manner. GANs can cope with multi-modal data by means of synthetic labels achieved through clustering, and our results show the creative potential of such techniques for logo synthesis and manipulation. Our dataset and models are publicly available at https://data.vision.ee.ethz.ch/sagea/lld/.

We propose and study the problem of distribution-preserving lossy compression. Motivated by recent advances in extreme image compression which allow to maintain artifact-free reconstructions even at very low bitrates, we propose to optimize the rate-distortion tradeoff under the constraint that the reconstructed samples follow the distribution of the training data. The resulting compression system recovers both ends of the spectrum: On one hand, at zero bitrate it learns a generative model of the data, and at high enough bitrates it achieves perfect reconstruction. Furthermore, for intermediate bitrates it smoothly interpolates between learning a generative model of the training data and perfectly reconstructing the training samples. We study several methods to approximately solve the proposed optimization problem, including a novel combination of Wasserstein GAN and Wasserstein Autoencoder, and present an extensive theoretical and empirical characterization of the proposed compression systems.

After decades of research, the real (biological) age estimation from a single face image reached maturity thanks to the availability of large public face databases and impressive accuracies achieved by recently proposed methods. The estimation of “apparent age” is a related task concerning the age perceived by human observers. Significant advances have been also made in this new research direction with the recent Looking At People challenges. In this paper we make several contributions to age estimation research. (i) We introduce APPA-REAL, a large face image database with both real and apparent age annotations. (ii)We study the relationship between real and apparent age. (iii) We develop a residual age regression method to further improve the performance. (iv) We show that real age estimation can be successfully tackled as an apparent age estimation followed by an apparent to real age residual regression. (v) We graphically reveal the facial regions on which the CNN focuses in order to perform apparent and real age estimation tasks.

We propose the Anchored Regression Network (ARN), a nonlinear regression network which can be seamlessly integrated into various networks or can be used stand-alone when the features have already been fixed. Our ARN is a smoothed relaxation of a piecewise linear regressor through the combination of multiple linear regressors over soft assignments to anchor points. When the anchor points are fixed the optimal ARN regressors can be obtained with a closed form global solution, otherwise ARN admits end-to-end learning with standard gradient based methods. We demonstrate the power of the ARN by applying it to two very diverse and challenging tasks: age prediction from face images and image super-resolution. In both cases, ARNs yield strong results.

We present a new approach to learn compressible representations in deep architectures with an end-to-end training strategy. Our method is based on a soft (continuous) relaxation of quantization and entropy, which we anneal to their discrete counterparts throughout training. We showcase this method for two challenging applications: Image compression and neural network compression. While these tasks have typically been approached with different methods, our soft-to-hard quantization approach gives results competitive with the state-of-the-art for both.

We propose an information-theoretic framework for analog signal separation. Specifically, we consider the problem of recovering two analog signals, modeled as general random vectors, from the noiseless sum of linear measurements of the signals. Our framework is inspired by the groundbreaking work of Wu and Verdú (2010) on analog compression and encompasses, inter alia, inpainting, declipping, super-resolution, the recovery of signals corrupted by impulse noise, and the separation of (e.g., audio or video) signals into two distinct components. The main results we report are general achievability bounds for the compression rate, i.e., the number of measurements relative to the dimension of the ambient space the signals live in, under either measurability or Hölder continuity imposed on the separator. Furthermore, we find a matching converse for sources of mixed discrete-continuous distribution. For measurable separators our proofs are based on a new probabilistic uncertainty relation, which shows that the intersection of generic subspaces with general sets of sufficiently small Minkowski dimension is empty. Hölder continuous separators are dealt with by introducing the concept of regularized probabilistic uncertainty relations. The probabilistic uncertainty relations we develop are inspired by embedding results in dynamical systems theory due to Sauer et al. (1991) and-conceptually-parallel classical Donoho-Stark and Elad-Bruckstein uncertainty principles at the heart of compressed sensing theory. Operationally, the new uncertainty relations take the theory of sparse signal separation beyond traditional sparsity-as measured in terms of the number of non-zero entries-to the more general notion of low description complexity as quantified by Minkowski dimension. Finally, our approach also allows to significantly strengthen key results in Wu and Verdú (2010).

A+ aka Adjusted Anchored Neighborhood Regression - is a state-of-the-art method for exemplar-based single image super-resolution with low time complexity at both train and test time. By robustly training a clustered regression model over a low-resolution dictionary, its performance keeps improving with the dictionary size - even when using tens of thousands of regressors. However, this can pose a memory issue where the model size can grow to more than a gigabyte, limiting applicability in memory constrained scenarios. To address this, we propose Regressor Basis Learning (RB), a novel variant of A+ where we restrict the regressor set to a learned low-dimensional subspace, such that each regressor is coded as a linear combination of few basis regressors. We learn the regressor basis by alternating between closed form solutions of the optimal coding of the regressor set (given the basis) and the optimal regressor basis (given the coding). We validate RB on several standard benchmarks and achieve comparable performance to A+ but by using orders of magnitude fewer basis regressors, ie. 32 basis regressors instead of 1024 regressors. This makes our RB method ideal for memory constrained applications.

Subspace clustering refers to the problem of clustering high-dimensional data points into a union of low-dimensional linear subspaces, where the number of subspaces, their dimensions and orientations are all unknown. In this paper, we propose a variation of the recently introduced thresholding-based subspace clustering (TSC) algorithm, which applies spectral clustering to an adjacency matrix constructed from the nearest neighbors of each data point with respect to the spherical distance measure. The new element resides in an individual and data-driven choice of the number of nearest neighbors. Previous performance results for TSC, as well as for other subspace clustering algorithms based on spectral clustering, come in terms of an intermediate performance measure, which does not address the clustering error directly. Our main analytical contribution is a performance analysis of the modified TSC algorithm (as well as the original TSC algorithm) in terms of the clustering error directly.