Despite their remarkable performance, modern Diffusion Transformers (DiTs) are hindered by substantial resource requirements during inference, stemming from the fixed and large amount of compute needed for each denoising step. In this work, we revisit the conventional static paradigm that allocates a fixed compute budget per denoising iteration and propose a dynamic strategy instead. Our simple and sample-efficient framework enables pre-trained DiT models to be converted into flexible ones — dubbed FlexiDiT— allowing them to process inputs at varying compute budgets. We demonstrate how a single flexible model can generate images without any drop in quality, while reducing the required FLOPs by more than 40% compared to their static counterparts, for both class-conditioned and text-conditioned image generation. Our method is general and agnostic to input and conditioning modalities. We show how our approach can be readily extended for video generation, where FlexiDiT models generate samples with up to 75% less compute without compromising performance.

Diffusion models with transformer architectures have demonstrated promising capabilities in generating high-fidelity images and scalability for high resolution. However, iterative sampling process required for synthesis is very resource-intensive. A line of work has focused on distilling solutions to probability flow ODEs into few-step student models. Nevertheless, existing methods have been limited by their reliance on the most recent denoised samples as input, rendering them susceptible to exposure bias. To address this limitation, we propose AutoRegressive Distillation (ARD), a novel approach that leverages the historical trajectory of the ODE to predict future steps. ARD offers two key benefits: 1) it mitigates exposure bias by utilizing a predicted historical trajectory that is less susceptible to accumulated errors, and 2) it leverages the previous history of the ODE trajectory as a more effective source of coarse-grained information. ARD modifies the teacher transformer architecture by adding token-wise time embedding to mark each input from the trajectory history and employs a block-wise causal attention mask for training. Furthermore, incorporating historical inputs only in lower transformer layers enhances performance and efficiency. We validate the effectiveness of ARD in a class-conditioned generation on ImageNet and T2I synthesis. Our model achieves a 5× reduction in FID degradation compared to the baseline methods while requiring only 1.1% extra FLOPs on ImageNet-256. Moreover, ARD reaches FID of 1.84 on ImageNet-256 in merely 4 steps and outperforms the publicly available 1024p text-to-image distilled models in prompt adherence score with a minimal drop in FID compared to the teacher. Project page: https://github.com/alsdudrla10/ARD.

The performance of large language models (LLMs) is closely linked to their underlying size, leading to ever-growing networks and hence slower inference. Speculative decoding has been proposed as a technique to accelerate autoregressive generation, leveraging a fast draft model to propose candidate tokens, which are then verified in parallel based on their likelihood under the target model. While this approach guarantees to reproduce the target output, it incurs a substantial penalty: many high-quality draft tokens are rejected, even when they represent objectively valid continuations. Indeed, we show that even powerful draft models such as GPT-4o, as well as human text cannot achieve high acceptance rates under the standard verification scheme. This severely limits the speedup potential of current speculative decoding methods, as an early rejection becomes overwhelmingly likely when solely relying on alignment of draft and target. We thus ask the following question: Can we adapt verification to recognize correct, but non-aligned replies? To this end, we draw inspiration from the LLM-as-a-judge framework, which demonstrated that LLMs are able to rate answers in a versatile way. We carefully design a dataset coined TokenCourt to elicit the same capability in the target model by training a compact module on top of the embeddings to produce ``judgements" of the current continuation. We showcase our strategy on the Llama-3.1 family, where our 8B/405B-Judge achieves a speedup of $9\times$ over Llama-405B, while maintaining its quality on a large range of benchmarks. These benefits remain present even in optimized inference frameworks, where our method reaches up to $141$ tokens/s for 8B/70B-Judge and $129$ tokens/s for 8B/405B on $2$ and $8$ H100s respectively.

Tracking the full body motions of users in XR (AR/VR) devices is a fundamental challenge to bring a sense of authentic social presence. Due to the absence of dedicated leg sensors, currently available body tracking methods adopt a synthesis approach to generate plausible motions given a 3-point signal from the head and controller tracking. In order to enable mixed reality features, modern XR devices are capable of estimating depth information of the headset surroundings using available sensors combined with dedicated machine learning models. Such egocentric depth sensing cannot drive the body directly, as it is not registered and is incomplete due to limited field-of-view and body self-occlusions. For the first time, we propose to lever-age the available depth sensing signal combined with self-supervision to learn a multi-modal pose estimation model capable of tracking full body motions in real time on XR devices. We demonstrate how current 3-point motion syn-thesis models can be extended to point cloud modalities using a semantic point cloud encoder network combined with a residual network for multi-modal pose estimation. These modules are trained jointly in a self-supervised way, leveraging a combination of real unregistered point clouds and simulated data obtained from motion capture. We compare our approach against several state-of-the-art systems for XR body tracking and show that our method accurately tracks a diverse range of body motions. XR-MBT tracks legs in XR for the first time, whereas traditional synthesis approaches based on partial body tracking are blind.

Diffusion models have recently revolutionized the field of image synthesis due to their ability to generate photorealistic images. However, one of the major drawbacks of diffusion models is that the image generation process is costly. A large image-to-image network has to be applied many times to iteratively refine an image from random noise. While many recent works propose techniques to reduce the number of required steps, they generally treat the underlying denoising network as a black box. In this work, we investigate the behavior of the layers within the network and find that 1) the layers' output changes smoothly over time, 2) the layers show distinct patterns of change, and 3) the change from step to step is often very small. We hypothesize that many layer computations in the denoising network are redundant. Leveraging this, we introduce block caching, in which we reuse outputs from layer blocks of previous steps to speed up inference. Furthermore, we propose a technique to automatically determine caching schedules based on each block's changes over timesteps. In our experiments, we show through FID, human evaluation and qualitative analysis that Block Caching allows to generate images with higher visual quality at the same computational cost. We demonstrate this for different state-of-the-art models (LDM and EMU) and solvers (DDIM and DPM). Project page: fwmb.github.io/blockcaching

AR/VR applications promise to provide people with a genuine feeling of mutual presence when communicating via their personalized avatars. While realistic avatars are essential in various social settings, the vast possibilities of a virtual world can also generate interest in using stylized avatars for other purposes. We introduce StyleAvatar, the first method for semantic stylization of animatable head avatars. StyleAvatar directly stylizes the avatar representation, rather than stylizing its renders. Specifically, given a model generating the avatar, StyleAvatar first disentangles geometry and texture manipulations, and then stylizes the avatar by fine-tuning a subset of the model’s weights. Our method has multiple virtues, including the ability to describe styles using images or text, preserving the avatar’s animatable capacity, providing control over identity preservation, and disentangling texture and geometry modifications. Experiments have shown that our approach consistently works across skin tones, challenging hair styles, extreme views, and diverse facial expressions.1

With the recent surge in popularity of AR/VR applications, realistic and accurate control of 3D full-body avatars has become a highly demanded feature. A particular challenge is that only a sparse tracking signal is available from standalone HMDs (Head Mounted Devices), often limited to tracking the user's head and wrists. While this signal is resourceful for reconstructing the upper body motion, the lower body is not tracked and must be synthesized from the limited information provided by the upper body joints. In this paper, we present AGRoL, a novel conditional diffusion model specifically designed to track full bodies given sparse upper-body tracking signals. Our model is based on a simple multi-layer perceptron (MLP) architecture and a novel conditioning scheme for motion data. It can predict accurate and smooth full-body motion, particularly the challenging lower body movement. Unlike common diffusion architectures, our compact architecture can run in real-time, making it suitable for online body-tracking applications. We train and evaluate our model on AMASS motion capture dataset, and demonstrate that our approach outperforms state-of-the-art methods in generated motion accuracy and smoothness. We further justify our design choices through extensive experiments and ablation studies.

This paper proposes a novel approach for rendering a pre-trained Neural Radiance Field (NeRF) in real-time on resource-constrained devices. We introduce Re-ReND, a method enabling Real-time Rendering of NeRFs across Devices. Re-ReND is designed to achieve real-time performance by converting the NeRF into a representation that can be efficiently processed by standard graphics pipelines. The proposed method distills the NeRF by extracting the learned density into a mesh, while the learned color information is factorized into a set of matrices that represent the scene’s light field. Factorization implies the field is queried via inexpensive MLP-free matrix multiplications, while using a light field allows rendering a pixel by querying the field a single time—as opposed to hundreds of queries when employing a radiance field. Since the proposed representation can be implemented using a fragment shader, it can be directly integrated with standard rasterization frameworks. Our flexible implementation can render a NeRF in real-time with low memory requirements and on a wide range of resource-constrained devices, including mobiles and AR/VR headsets. Notably, we find that Re-ReND can achieve over a 2.6-fold increase in rendering speed versus the state-of-the-art without perceptible losses in quality.

Surface reconstruction has been seeing a lot of progress lately by utilizing Implicit Neural Representations (INRs). Despite their success, INRs often introduce hard to control inductive bias (i.e., the solution surface can exhibit unexplainable behaviours), have costly inference, and are slow to train. The goal of this work is to show that replacing neural networks with simple grid functions, along with two novel geometric priors achieve comparable results to INRs, with instant inference, and improved training times. To that end we introduce VisCo Grids: a grid-based surface reconstruction method incorporating Viscosity and Coarea priors. Intuitively, the Viscosity prior replaces the smoothness inductive bias of INRs, while the Coarea favors a minimal area solution. Experimenting with VisCo Grids on a standard reconstruction baseline provided comparable results to the best performing INRs on this dataset.

Deep metric learning (DML) is a cornerstone of many computer vision applications. It aims at learning a mapping from the input domain to an embedding space, where semantically similar objects are located nearby and dissimilar objects far from another. The target similarity on the training data is defined by user in form of ground-truth class labels. However, while the embedding space learns to mimic the user-provided similarity on the training data, it should also generalize to novel categories not seen during training. Besides user-provided groundtruth training labels, a lot of additional visual factors (such as viewpoint changes or shape peculiarities) exist and imply different notions of similarity between objects, affecting the generalization on the images unseen during training. However, existing approaches usually directly learn a single embedding space on all available training data, struggling to encode all different types of relationships, and do not generalize well. We propose to build a more expressive representation by jointly splitting the embedding space and the data hierarchically into smaller sub-parts. We successively focus on smaller subsets of the training data, reducing its variance and learning a different embedding subspace for each data subset. Moreover, the subspaces are learned jointly to cover not only the intricacies, but the breadth of the data as well. Only after that, we build the final embedding from the subspaces in the conquering stage. The proposed algorithm acts as a transparent wrapper that can be placed around arbitrary existing DML methods. Our approach significantly improves upon the state-of-the-art on image retrieval, clustering, and re-identification tasks evaluated using CUB200-2011, CARS196, Stanford Online Products, In-shop Clothes, and PKU VehicleID datasets.

We tackle the problem of learning the geometry of multiple categories of deformable objects jointly. Recent work has shown that it is possible to learn a unified dense pose predictor for several categories of related objects. However, training such models requires to initialize inter-category correspondences by hand. This is suboptimal and the resulting models fail to maintain correct correspondences as individual categories are learned. In this paper, we show that improved correspondences can be learned automatically as a natural byproduct of learning category-specific dense pose predictors. To do this, we express correspondences between different categories and between images and categories using a unified embedding. Then, we use the latter to enforce two constraints: symmetric inter-category cycle consistency and a new asymmetric image-to-category cycle consistency. Without any manual annotations for the intercategory correspondences, we obtain state-of-the-art alignment results, outperforming dedicated methods for matching 3D shapes. Moreover, the new model is also better at the task of dense pose prediction than prior work.

Consumer-level depth cameras and depth sensors embedded in mobile devices enable numerous applications, such as AR games and face identification. However, the quality of the captured depth is sometimes insufficient for 3D reconstruction, tracking and other computer vision tasks. In this paper, we propose a self-supervised depth denoising approach to denoise and refine depth coming from a low quality sensor. We record simultaneous RGB-D sequences with unzynchronized lower- and higher-quality cameras and solve a challenging problem of aligning sequences both temporally and spatially. We then learn a deep neural network to denoise the lower-quality depth using the matched higher-quality data as a source of supervision signal. We experimentally validate our method against state-of-the-art filtering-based and deep denoising techniques and show its application for 3D object reconstruction tasks where our approach leads to more detailed fused surfaces and better tracking.

Recent contributions have demonstrated that it is possible to recognize the pose of humans densely and accurately given a large dataset of poses annotated in detail. In principle, the same approach could be extended to any animal class, but the effort required for collecting new annotations for each case makes this strategy impractical, despite important applications in natural conservation, science and business. We show that, at least for proximal animal classes such as chimpanzees, it is possible to transfer the knowledge existing in dense pose recognition for humans, as well as in more general object detectors and segmenters, to the problem of dense pose recognition in other classes. We do this by (1) establishing a DensePose model for the new animal which is also geometrically aligned to humans (2) introducing a multi-head R-CNN architecture that facilitates transfer of multiple recognition tasks between classes, (3) finding which combination of known classes can be transferred most effectively to the new animal and (4) using self-calibrated uncertainty heads to generate pseudo-labels graded by quality for training a model for this class. We also introduce two benchmark datasets labelled in the manner of DensePose for the class chimpanzee and use them to evaluate our approach, showing excellent transfer learning performance.

Style transfer has recently received a lot of attention, since it allows to study fundamental challenges in image understanding and synthesis. Recent work has significantly improved the representation of color and texture and computational speed and image resolution. The explicit transformation of image content has, however, been mostly neglected: while artistic style affects formal characteristics of an image, such as color, shape or texture, it also deforms, adds or removes content details. This paper explicitly focuses on a content-and style-aware stylization of a content image. Therefore, we introduce a content transformation module between the encoder and decoder. Moreover, we utilize similar content appearing in photographs and style samples to learn how style alters content details and we generalize this to other class details. Additionally, this work presents a novel normalization layer critical for high resolution image synthesis. The robustness and speed of our model enables a video stylization in real-time and high definition. We perform extensive qualitative and quantitative evaluations to demonstrate the validity of our approach.

Learning the embedding space, where semantically similar objects are located close together and dissimilar objects far apart, is a cornerstone of many computer vision applications. Existing approaches usually learn a single metric in the embedding space for all available data points, which may have a very complex non-uniform distribution with different notions of similarity between objects, e.g. appearance, shape, color or semantic meaning. Approaches for learning a single distance metric often struggle to encode all different types of relationships and do not generalize well. In this work, we propose a novel easy-to-implement divide and conquer approach for deep metric learning, which significantly improves the state-of-the-art performance of metric learning. Our approach utilizes the embedding space more efficiently by jointly splitting the embedding space and data into K smaller sub-problems. It divides both, the data and the embedding space into K subsets and learns K separate distance metrics in the non-overlapping subspaces of the embedding space, defined by groups of neurons in the embedding layer of the neural network. The proposed approach increases the convergence speed and improves generalization since the complexity of each sub-problem is reduced compared to the original one. We show that our approach outperforms the state-of-the-art by a large margin in retrieval, clustering and re-identification tasks on CUB200-2011, CARS196, Stanford Online Products, In-shop Clothes and PKU VehicleID datasets. Source code: https://bit.ly/dcesml.

Artists rarely paint in a single style throughout their career. More often they change styles or develop variations of it. In addition, artworks in different styles and even within one style depict real content differently: while Picasso's Blue Period displays a vase in a blueish tone but as a whole, his Cubist works deconstruct the object. To produce artistically convincing stylizations, style transfer models must be able to reflect these changes and variations. Recently many works have aimed to improve the style transfer task, but neglected to address the described observations. We present a novel approach which captures particularities of style and the variations within and separates style and content. This is achieved by introducing two novel losses: a fixpoint triplet style loss to learn subtle variations within one style or between different styles and a disentanglement loss to ensure that the stylization is not conditioned on the real input photo. In addition the paper proposes various evaluation methods to measure the importance of both losses on the validity, quality and variability of final stylizations. We provide qualitative results to demonstrate the performance of our approach.

Unsupervised learning of visual similarities is of paramount importance to computer vision, particularly due to lacking training data for fine-grained similarities. Deep learning of similarities is often based on relationships between pairs or triplets of samples. Many of these relations are unreliable and mutually contradicting, implying inconsistencies when trained without supervision information that relates different tuples or triplets to each other. To overcome this problem, we use local estimates of reliable (dis-)similarities to initially group samples into compact surrogate classes and use local partial orders of samples to classes to link classes to each other. Similarity learning is then formulated as a partial ordering task with soft correspondences of all samples to classes. Adopting a strategy of self-supervision, a CNN is trained to optimally represent samples in a mutually consistent manner while updating the classes. The similarity learning and grouping procedure are integrated in a single model and optimized jointly. The proposed unsupervised approach shows competitive performance on detailed pose estimation and object classification.

Exemplar learning is a powerful paradigm for discovering visual similarities in an unsupervised manner. In this context, however, the recent breakthrough in deep learning could not yet unfold its full potential. With only a single positive sample, a great imbalance between one positive and many negatives, and unreliable relationships between most samples, training of convolutional neural networks is impaired. Given weak estimates of local distance we propose a single optimization problem to extract batches of samples with mutually consistent relations. Conflicting relations are distributed over different batches and similar samples are grouped into compact cliques. Learning exemplar similarities is framed as a sequence of clique categorization tasks. The CNN then consolidates transitivity relations within and between cliques and learns a single representation for all samples without the need for labels. The proposed unsupervised approach has shown competitive performance on detailed posture analysis and object classification.