This paper presents a comprehensive study on the role of Classifier-Free Guidance (CFG) in text-conditioned diffusion models from the perspective of inference efficiency. In particular, we relax the default choice of applying CFG in all diffusion steps and instead propose to search for more efficient guidance policies. We formulate the discovery of such policies in the framework of differentiable neural architecture search. Our findings suggest that, as denoising progresses, the updates produced by CFG become increasingly aligned with simple conditional steps, which renders CFG's additional neural network evaluation redundant, especially in the second half of the denoising process. Building upon this insight, we propose "Adaptive Guidance" (AG), an efficient variant of CFG that adaptively omits network evaluations when the denoising process displays convergence. Our experiments demonstrate that AG preserves CFG's image quality while reducing computation by 25%. Thus, AG constitutes a plug-and-play alternative to Guidance Distillation, achieving 50% of the speed-ups of the latter, while being training-free and retaining the capacity to handle negative prompts. We conclude by uncovering further redundancies of CFG in the first half of the diffusion process, showing that entire neural network evaluations can be replaced by simple affine transformations of past score estimates.

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.

Diffusion or flow-based models are powerful generative paradigms that are notoriously hard to sample as samples are defined as solutions to high-dimensional Ordinary or Stochastic Differential Equations (ODEs/SDEs) which require a large Number of Function Evaluations (NFE) to approximate well. Existing methods to alleviate the costly sampling process include model distillation and designing dedicated ODE solvers. However, distillation is costly to train and sometimes can deteriorate quality, while dedicated solvers still require relatively large NFE to produce high quality samples. In this paper we introduce ``Bespoke solvers'', a novel framework for constructing custom ODE solvers tailored to the ODE of a given pre-trained flow model. Our approach optimizes an order consistent and parameter-efficient solver (e.g., with 80 learnable parameters), is trained for roughly 1\% of the GPU time required for training the pre-trained model, and significantly improves approximation and generation quality compared to dedicated solvers. For example, a Bespoke solver for a CIFAR10 model produces samples with Fréchet Inception Distance (FID) of 2.73 with 10 NFE, and gets to 1\% of the Ground Truth (GT) FID (2.59) for this model with only 20 NFE. On the more challenging ImageNet-64$\times$64, Bespoke samples at 2.2 FID with 10 NFE, and gets within 2\% of GT FID (1.71) with 20 NFE.

This paper introduces Bespoke Non-Stationary (BNS) Solvers, a solver distillation approach to improve sample efficiency of Diffusion and Flow models. BNS solvers are based on a family of non-stationary solvers that provably subsumes existing numerical ODE solvers and consequently demonstrate considerable improvement in sample approximation (PSNR) over these baselines. Compared to model distillation, BNS solvers benefit from a tiny parameter space ($<$200 parameters), fast optimization (two orders of magnitude faster), maintain diversity of samples, and in contrast to previous solver distillation approaches nearly close the gap from standard distillation methods such as Progressive Distillation in the low-medium NFE regime. For example, BNS solver achieves 45 PSNR / 1.76 FID using 16 NFE in class-conditional ImageNet-64. We experimented with BNS solvers for conditional image generation, text-to-image generation, and text-2-audio generation showing significant improvement in sample approximation (PSNR) in all.

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.

Convolutional neural networks (CNNs) have been very successful at solving a variety of computer vision tasks such as object classification and detection, semantic segmentation, activity understanding, to name just a few. One key enabling factor for their great performance has been the ability to train very deep networks. Despite their huge success in many tasks, CNNs do not work well with non-euclidean data, which is prevalent in many real-world applications. Graph Convolutional Networks (GCNs) offer an alternative that allows for non-Eucledian data input to a neural network. While GCNs already achieve encouraging results, they are currently limited to architectures with a relatively small number of layers, primarily due to vanishing gradients during training. This work transfers concepts such as residual/dense connections and dilated convolutions from CNNs to GCNs in order to successfully train very deep GCNs. We show the benefit of using deep GCNs (with as many as 112 layers) experimentally across various datasets and tasks. Specifically, we achieve very promising performance in part segmentation and semantic segmentation on point clouds and in node classification of protein functions across biological protein-protein interaction (PPI) graphs. We believe that the insights in this work will open avenues for future research on GCNs and their application to further tasks not explored in this paper. The source code for this work is available at https://github.com/lightaime/deep_gcns_torch and https://github.com/lightaime/deep_gcns for PyTorch and TensorFlow implementations respectively.

Graph Neural Networks (GNNs) have been drawing significant attention to representation learning on graphs. Recent works developed frameworks to train very deep GNNs and showed impressive results in tasks like point cloud learning and protein interaction prediction. In this work, we study the performance of such deep models in large-scale graphs. In particular, we look at the effect of adequately choosing an aggregation function on deep models. We find that GNNs are very sensitive to the choice of aggregation functions (e.g. mean, max, and sum) when applied to different datasets. We systematically study and propose to alleviate this issue by introducing a novel class of aggregation functions named Generalized Aggregation Functions. The proposed functions extend beyond commonly used aggregation functions to a wide range of new permutation-invariant functions. Generalized Aggregation Functions are fully differentiable, where their parameters can be learned in an end-to-end fashion to yield a suitable aggregation function for each task. We show that equipped with the proposed aggregation functions, deep residual GNNs outperform state-of-the-art in several benchmarks from Open Graph Benchmark (OGB) across tasks and domains.

Point cloud architecture design has become a crucial problem for deep learning in 3D. Several efforts have been made to manually design architectures targeting high accuracy in point cloud tasks such as classification, segmentation, and detection. Recent progress in automatic Neural Architecture Search (NAS) minimizes the human effort in network design and optimizes architectures for high performance. However, those efforts fail to consider crucial factors such as latency during inference, which is of high importance in time-critical and hardware-bounded applications like self-driving cars, robot navigation, and mobile applications. In this paper, we introduce a new NAS framework, dubbed LC-NAS, that searches for point cloud architectures constrained to a target latency. We implement a novel latency constraint formulation for the trade-off between accuracy and latency in our architecture search. Contrary to previous works, our latency loss enables us to find the best architecture with latency near a specific target value, which is crucial when the end task is to be deployed in a limited hardware setting. Extensive experiments show that LC-NAS is able to find state-of-the-art architectures for point cloud classification in ModelNet40 with a minimal computational cost. We also show how our searched architectures achieve any desired latency with a reasonably low drop in accuracy. Finally, we show how our searched architectures easily transfer to the part segmentation task on PartNet, where we achieve state-of-the-art results with significantly lower latency.

Deep neural networks are vulnerable to small input perturbations known as adversarial attacks. Inspired by the fact that these adversaries are constructed by iteratively minimizing the confidence of a network for the true class label, we propose the anti-adversary layer, aimed at countering this effect. In particular, our layer generates an input perturbation in the opposite direction of the adversarial one and feeds the classifier a perturbed version of the input. Our approach is training-free and theoretically supported. We verify the effectiveness of our approach by combining our layer with both nominally and robustly trained models and conduct large-scale experiments from black-box to adaptive attacks on CIFAR10, CIFAR100, and ImageNet. Our layer significantly enhances model robustness while coming at no cost on clean accuracy.

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.

The effectiveness of learning-based point cloud upsampling pipelines heavily relies on the upsampling modules and feature extractors used therein. For the point upsampling module, we propose a novel model called NodeShuffle, which uses a Graph Convolutional Network (GCN) to better encode local point information from point neighborhoods. NodeShuffle is versatile and can be incorporated into any point cloud upsampling pipeline. Extensive experiments show how NodeShuffle consistently improves state-of-the-art upsampling methods. For feature extraction, we also propose a new multi-scale point feature extractor, called Inception DenseGCN. By aggregating features at multiple scales, this feature extractor enables further performance gain in the final upsampled point clouds. We combine Inception DenseGCN with NodeShuffle into a new point upsampling pipeline called PU-GCN. PU-GCN sets new state-of-art performance with much fewer parameters and more efficient inference. Our code is publicly available at https://github.com/guochengqian/PU-GCN.

Active speaker detection requires a mindful integration of multi-modal cues. Current methods focus on modeling and fusing short-term audiovisual features for individual speakers, often at frame level. We present a novel approach to active speaker detection that directly addresses the multi-modal nature of the problem and provides a straightforward strategy, where independent visual features (speakers) in the scene are assigned to a previously detected speech event. Our experiments show that a small graph data structure built from local information can approximate an instantaneous audio-visual assignment problem. More-over, the temporal extension of this initial graph achieves a new state-of-the-art performance on the AVA-ActiveSpeaker dataset with a mAP of 88.8%.

Video content creation keeps growing at an incredible pace; yet, creating engaging stories remains challenging and requires non-trivial video editing expertise. Many video editing components are astonishingly hard to automate primarily due to the lack of raw video materials. This paper focuses on a new task for computational video editing, namely the task of raking cut plausibility. Our key idea is to leverage content that has already been edited to learn fine-grained audiovisual patterns that trigger cuts. To do this, we first collected a data source of more than 10K videos, from which we extract more than 255K cuts. We devise a model that learns to discriminate between real and artificial cuts via contrastive learning. We set up a new task and a set of baselines to benchmark video cut generation. We observe that our proposed model outperforms the baselines by large margins. To demonstrate our model in real-world applications, we conduct human studies in a collection of unedited videos. The results show that our model does a better job at cutting than random and alternative baselines.

Temporal action localization (TAL) in videos is a challenging task, especially due to the large variation in action temporal scales. Short actions usually occupy a major proportion in the datasets, but tend to have the lowest performance. In this paper, we confront the challenge of short actions and propose a multi-level cross-scale solution dubbed as video self-stitching graph network (VSGN). We have two key components in VSGN: video self-stitching (VSS) and cross-scale graph pyramid network (xGPN). In VSS, we focus on a short period of a video and magnify it along the temporal dimension to obtain a larger scale. We stitch the original clip and its magnified counterpart in one input sequence to take advantage of the complementary properties of both scales. The xGPN component further exploits the cross-scale correlations by a pyramid of cross-scale graph networks, each containing a hybrid module to aggregate features from across scales as well as within the same scale. Our VSGN not only enhances the feature representations, but also generates more positive anchors for short actions and more short training samples. Experiments demonstrate that VSGN obviously improves the localization performance of short actions as well as achieving the state-of-the-art overall performance on THUMOS-14 and ActivityNet-v1.3. VSGN code is available at https://github.com/coolbay/VSGN.

Access to 3D point cloud representations has been widely facilitated by LiDAR sensors embedded in various mobile devices. This has led to an emerging need for fast and accurate point cloud processing techniques. In this paper, we revisit and dive deeper into PointNet++, one of the most influential yet under-explored networks, and develop faster and more accurate variants of the model. We first present a novel Separable Set Abstraction (SA) module that disentangles the vanilla SA module used in PointNet++ into two separate learning stages: (1) learning channel correlation and (2) learning spatial correlation. The Separable SA module is significantly faster than the vanilla version, yet it achieves comparable performance. We then introduce a new Anisotropic Reduction function into our Separable SA module and propose an Anisotropic Separable SA (ASSA) module that substantially increases the network's accuracy. We later replace the vanilla SA modules in PointNet++ with the proposed ASSA modules, and denote the modified network as ASSANet. Extensive experiments on point cloud classification, semantic segmentation, and part segmentation show that ASSANet outperforms PointNet++ and other methods, achieving much higher accuracy and faster speeds. In particular, ASSANet outperforms PointNet++ by $7.4$ mIoU on S3DIS Area 5, while maintaining $1.6 \times $ faster inference speed on a single NVIDIA 2080Ti GPU. Our scaled ASSANet variant achieves $66.8$ mIoU and outperforms KPConv, while being more than $54 \times$ faster.

Video action detectors are usually trained using datasets with fully-supervised temporal annotations. Building such datasets is an expensive task. To alleviate this problem, recent methods have tried to leverage weak labeling, where videos are untrimmed and only a video-level label is available. In this paper, we propose RefineLoc, a novel weaklysupervised temporal action localization method. RefineLoc uses an iterative refinement approach by estimating and training on snippet-level pseudo ground truth at every iteration. We show the benefit of this iterative approach and present an extensive analysis of five different pseudo ground truth generators. We show the effectiveness of our model on two standard action datasets, ActivityNet v1.2 and THUMOS14. RefineLoc shows competitive results with the stateof-the-art in weakly-supervised temporal localization. Additionally, our iterative refinement process is able to significantly improve the performance of two state-of-the-art methods, setting a new state-of-the-art on THUMOS14.

Abstract High dynamic range (HDR) image acquisition from a single image capture, also known as snapshot HDR imaging, is challenging because the bit depths of camera sensors are far from sufficient to cover the full dynamic range of the scene. Existing HDR techniques focus either on algorithmic reconstruction or hardware modification to extend the dynamic range. In this paper we propose a joint design for snapshot HDR imaging by devising a spatially varying modulation mask in the hardware and building a deep learning algorithm to reconstruct the HDR image. We leverage transfer learning to overcome the lack of sufficiently large HDR datasets available. We show how transferring from a different large‐scale task (image classification on ImageNet) leads to considerable improvements in HDR reconstruction. We achieve a reconfigurable HDR camera design that does not require custom sensors, and instead can be reconfigured between HDR and conventional mode with very simple calibration steps. We demonstrate that the proposed hardware–software so lution offers a flexible yet robust way to modulate per‐pixel exposures, and the network requires little knowledge of the hardware to faithfully reconstruct the HDR image. Comparison results show that our method outperforms the state of the art in terms of visual perception quality.

Architecture design has become a crucial component of successful deep learning. Recent progress in automatic neural architecture search (NAS) shows a lot of promise. However, discovered architectures often fail to generalize in the final evaluation. Architectures with a higher validation accuracy during the search phase may perform worse in the evaluation. Aiming to alleviate this common issue, we introduce sequential greedy architecture search (SGAS), an efficient method for neural architecture search. By dividing the search procedure into sub-problems, SGAS chooses and prunes candidate operations in a greedy fashion. We apply SGAS to search architectures for Convolutional Neural Networks (CNN) and Graph Convolutional Networks (GCN). Extensive experiments show that SGAS is able to find state-of-the-art architectures for tasks such as image classification, point cloud classification and node classification in protein-protein interaction graphs with minimal computational cost.

Temporal action detection is a fundamental yet challenging task in video understanding. Video context is a critical cue to effectively detect actions, but current works mainly focus on temporal context, while neglecting semantic context as well as other important context properties. In this work, we propose a graph convolutional network (GCN) model to adaptively incorporate multi-level semantic context into video features and cast temporal action detection as a sub-graph localization problem. Specifically, we formulate video snippets as graph nodes, snippet-snippet correlations as edges, and actions associated with context as target sub-graphs. With graph convolution as the basic operation, we design a GCN block called GCNeXt, which learns the features of each node by aggregating its context and dynamically updates the edges in the graph. To localize each sub-graph, we also design an SGAlign layer to embed each sub-graph into the Euclidean space. Extensive experiments show that G-TAD is capable of finding effective video context without extra supervision and achieves state-of-the-art performance on two detection benchmarks. On ActivityNet-1.3 it obtains an average mAP of 34.09%; on THUMOS14 it reaches 51.6% at [email protected] when combined with a proposal processing method. The code has been made available at https://github.com/frostinassiky/gtad.

Convolutional Neural Networks (CNNs) achieve impressive performance in a wide variety of fields. Their success benefited from a massive boost when very deep CNN models were able to be reliably trained. Despite their merits, CNNs fail to properly address problems with non-Euclidean data. To overcome this challenge, Graph Convolutional Networks (GCNs) build graphs to represent non-Euclidean data, borrow concepts from CNNs, and apply them in training. GCNs show promising results, but they are usually limited to very shallow models due to the vanishing gradient problem. As a result, most state-of-the-art GCN models are no deeper than 3 or 4 layers. In this work, we present new ways to successfully train very deep GCNs. We do this by borrowing concepts from CNNs, specifically residual/dense connections and dilated convolutions, and adapting them to GCN architectures. Extensive experiments show the positive effect of these deep GCN frameworks. Finally, we use these new concepts to build a very deep 56-layer GCN, and show how it significantly boosts performance (+3.7% mIoU over state-of-the-art) in the task of point cloud semantic segmentation. We believe that the community can greatly benefit from this work, as it opens up many opportunities for advancing GCN-based research.

This paper proposes a new framework for estimating the Manhattan Frame (MF) of an indoor scene from a single RGB-D image. Our technique formulates this problem as the estimation of a rotation matrix that best aligns the normals of the captured scene to a canonical world axes. By introducing sparsity constraints, our method can simultaneously estimate the scene MF, the surfaces in the scene that are best aligned to one of three coordinate axes, and the outlier surfaces that do not align with any of the axes. To test our approach, we contribute a new set of annotations to determine ground truth MFs in each image of the popular NYUv2 dataset. We use this new benchmark to experimentally demonstrate that our method is more accurate, faster, more reliable and more robust than the methods used in the literature. We further motivate our technique by showing how it can be used to address the RGB-D SLAM problem in indoor scenes by incorporating it into and improving the performance of a popular RGB-D SLAM method.