The field of video generation has made remarkable advancements, yet there remains a pressing need for a clear, systematic recipe that can guide the development of robust and scalable models. In this work, we present a comprehensive study that systematically explores the interplay of model architectures, training recipes, and data curation strategies, culminating in a simple and scalable text-image-conditioned video generation method, named STIV. Our framework integrates image condition into a Diffusion Transformer (DiT) through frame replacement, while incorporating text conditioning via a joint image-text conditional classifier-free guidance. This design enables STIV to perform both text-to-video (T2V) and text-image-to-video (TI2V) tasks simultaneously. Additionally, STIV can be easily extended to various applications, such as video prediction, frame interpolation, multi-view generation, and long video generation, etc. With comprehensive ablation studies on T2I, T2V, and TI2V, STIV demonstrate strong performance, despite its simple design. An 8.7B model with 512 resolution achieves 83.1 on VBench T2V, surpassing both leading open and closed-source models like CogVideoX-5B, Pika, Kling, and Gen-3. The same-sized model also achieves a state-of-the-art result of 90.1 on VBench I2V task at 512 resolution. By providing a transparent and extensible recipe for building cutting-edge video generation models, we aim to empower future research and accelerate progress toward more versatile and reliable video generation solutions.

In recent years there have been remarkable breakthroughs in image-to-video generation. However, the 3D consistency and camera controllability of generated frames have remained unsolved. Recent studies have attempted to incorporate camera control into the generation process, but their results are often limited to simple trajectories or lack the ability to generate consistent videos from multiple distinct camera paths for the same scene. To address these limitations, we introduce Cavia, a novel framework for camera-controllable, multi-view video generation, capable of converting an input image into multiple spatiotemporally consistent videos. Our framework extends the spatial and temporal attention modules into view-integrated attention modules, improving both viewpoint and temporal consistency. This flexible design allows for joint training with diverse curated data sources, including scene-level static videos, object-level synthetic multi-view dynamic videos, and real-world monocular dynamic videos. To the best of our knowledge, Cavia is the first framework that enables users to generate multiple videos of the same scene with precise control over camera motion, while simultaneously preserving object motion. Extensive experiments demonstrate that Cavia surpasses state-of-the-art methods in terms of geometric consistency and perceptual quality.

Conditional generative modeling aims to learn a conditional data distribution from samples containing data-condition pairs. For this, diffusion and flow-based methods have attained compelling results. These methods use a learned (flow) model to transport an initial standard Gaussian noise that ignores the condition to the conditional data distribution. The model is hence required to learn both mass transport \emph{and} conditional injection. To ease the demand on the model, we propose \emph{Condition-Aware Reparameterization for Flow Matching} (CAR-Flow) -- a lightweight, learned \emph{shift} that conditions the source, the target, or both distributions. By relocating these distributions, CAR-Flow shortens the probability path the model must learn, leading to faster training in practice. On low-dimensional synthetic data, we visualize and quantify the effects of CAR-Flow. On higher-dimensional natural image data (ImageNet-256), equipping SiT-XL/2 with CAR-Flow reduces FID from 2.07 to 1.68, while introducing less than \(0.6\%\) additional parameters.

The task of novel view synthesis aims to generate unseen perspectives of an object or scene from a limited set of input images. Nevertheless, synthesizing novel views from a single image remains a significant challenge. Previous approaches tackle this problem by adopting mesh prediction, multi-plane image construction, or more advanced techniques such as neural radiance fields. Recently, a pre-trained diffusion model that is specifically designed for 2D image synthesis has demonstrated its capability in producing photorealistic novel views, if sufficiently optimized with a 3D finetuning task. Despite greatly improved fidelity and generalizability, training such a powerful diffusion model requires a vast volume of training data and model parameters, resulting in a notoriously long time and high computational costs. To tackle this issue, we propose Efficient-3DiM, a highly efficient yet effective framework to learn a single-image novel-view synthesizer. Motivated by our in-depth analysis of the diffusion model inference process, we propose several pragmatic strategies to reduce training overhead to a manageable scale, including a crafted timestep sampling strategy, a superior 3D feature extractor, and an enhanced training scheme. When combined, our framework can reduce the total training time from 10 days to less than 1 day, significantly accelerating the training process on the same computational platform (an instance with 8 Nvidia A100 GPUs). Comprehensive experiments are conducted to demonstrate the efficiency and generalizability of our proposed method.

To date, little attention has been given to multi-view 3D human mesh estimation, despite real-life applicability (e.g., motion capture, sport analysis) and robustness to single-view ambiguities. Existing solutions typically suffer from poor generalization performance to new settings, largely due to the limited diversity of image/3D-mesh pairs in multi-view training data. To address this shortcoming, people have explored the use of synthetic images. But besides the usual impact of visual gap between rendered and target data, synthetic-data-driven multi-view estimators also suffer from overfitting to the camera viewpoint distribution sampled during training which usually differs from real-world distributions. Tackling both challenges, we propose a novel simulation-based training pipeline for multi-view human mesh recovery, which (a) relies on intermediate 2D representations which are more robust to synthetic-to-real domain gap; (b) leverages learnable calibration and triangulation to adapt to more diversified camera setups; and (c) progressively aggregates multi-view information in a canonical 3D space to remove ambiguities in 2D representations. Through extensive benchmarking, we demonstrate the superiority of the proposed solution especially for unseen in-the-wild scenarios.

We present a novel type of neural fields that uses general radial bases for signal representation. State-of-the-art neural fields typically rely on grid-based representations for storing local neural features and N-dimensional linear kernels for interpolating features at continuous query points. The spatial positions of their neural features are fixed on grid nodes and cannot well adapt to target signals. Our method instead builds upon general radial bases with flexible kernel position and shape, which have higher spatial adaptivity and can more closely fit target signals. To further improve the channel-wise capacity of radial basis functions, we propose to compose them with multi-frequency sinusoid functions. This technique extends a radial basis to multiple Fourier radial bases of different frequency bands without requiring extra parameters, facilitating the representation of details. Moreover, by marrying adaptive radial bases with grid-based ones, our hybrid combination inherits both adaptivity and interpolation smoothness. We carefully designed weighting schemes to let radial bases adapt to different types of signals effectively. Our experiments on 2D image and 3D signed distance field representation demonstrate the higher accuracy and compactness of our method than prior arts. When applied to neural radiance field reconstruction, our method achieves state-of-the-art rendering quality, with small model size and comparable training speed.

In this paper, we address the problem of simultaneous relighting and novel view synthesis of a complex scene from multi-view images with a limited number of light sources. We propose an analysis-synthesis approach called Relit-NeuLF. Following the recent neural 4D light field network (NeuLF)[22], Relit-NeuLF first leverages a two-plane light field representation to parameterize each ray in a 4D coordinate system, enabling efficient learning and inference. Then, we recover the spatially-varying bidirectional reflectance distribution function (SVBRDF) of a 3D scene in a self-supervised manner. A DecomposeNet learns to map each ray to its SVBRDF components: albedo, normal, and roughness. Based on the decomposed BRDF components and conditioning light directions, a RenderNet learns to synthesize the color of the ray. To self-supervise the SVBRDF decomposition, we encourage the predicted ray color to be close to the physically-based rendering result using the microfacet model. Comprehensive experiments demonstrate that the proposed method is efficient and effective on both synthetic data and real-world human face data, and outperforms the state-of-the-art results.

The techniques for 3D indoor scene capturing are widely used, but the meshes produced leave much to be desired. In this paper, we propose "RoomDreamer", which leverages powerful natural language to synthesize a new room with a different style. Unlike existing image synthesis methods, our work addresses the challenge of synthesizing both geometry and texture aligned to the input scene structure and prompt simultaneously. The key insight is that a scene should be treated as a whole, taking into account both scene texture and geometry. The proposed framework consists of two significant components: Geometry Guided Diffusion and Mesh Optimization. Geometry Guided Diffusion for 3D Scene guarantees the consistency of the scene style by applying the 2D prior to the entire scene simultaneously. Mesh Optimization improves the geometry and texture jointly and eliminates the artifacts in the scanned scene. To validate the proposed method, real indoor scenes scanned with smartphones are used for extensive experiments, through which the effectiveness of our method is demonstrated.

Response-based Knowledge Distillation refers to the technique of supervising the student network with the teacher networks' predictions. The method is motivated by observing that the predicted probabilities reflect the relation among labels, which is the knowledge to be transferred. This paper explores the transferred knowledge from a novel perspective: comparing the knowledge transferred through different teachers. Two intriguing properties are observed. First, higher confidence scores of teachers' predictions lead to better distillation results, and second, teachers' incorrectly predicted training samples should be kept for distillation. We then analyze the phenomenon by studying teachers' decision boundaries, of which some can help the student generalize while some may not. Based on the observations, we further propose an embarrassingly simple distillation framework named Efficient Distillation, which is effective on ImageNet with different teacher-student pairs: When using ResNet34 as the teacher, the student ResNet18 trained from scratch reaches 74.07% Top-1 accuracy within 98 GPU hours (RTX 3090), outperforming current state-of-the-art result (73.19%) by a large margin. Our code is available at https://github.com/lsongx/EffDstl.

Federated Learning (FL) is a machine learning paradigm where many local nodes collaboratively train a central model while keeping the training data decentralized. This is particularly relevant for clinical applications since patient data are usually not allowed to be transferred out of medical facilities, leading to the need for FL. Existing FL methods typically share model parameters or employ co-distillation to address the issue of unbalanced data distribution. However, they also require numerous rounds of synchronized communication and, more importantly, suffer from a privacy leakage risk. We propose a privacy-preserving FL framework leveraging unlabeled public data for one-way offline knowledge distillation in this work. The central model is learned from local knowledge via ensemble attention distillation. Our technique uses decentralized and heterogeneous local data like existing FL approaches, but more importantly, it significantly reduces the risk of privacy leakage. We demonstrate that our method achieves very competitive performance with more robust privacy preservation based on extensive experiments on image classification, segmentation, and reconstruction tasks.

Visually exploring in a real-world 4D spatiotemporal space freely in VR has been a long-term quest. The task is especially appealing when only a few or even single RGB cameras are used for capturing the dynamic scene. To this end, we present an efficient framework capable of fast reconstruction, compact modeling, and streamable rendering. First, we propose to decompose the 4D spatiotemporal space according to temporal characteristics. Points in the 4D space are associated with probabilities of belonging to three categories: static, deforming, and new areas. Each area is represented and regularized by a separate neural field. Second, we propose a hybrid representations based feature streaming scheme for efficiently modeling the neural fields. Our approach, coined NeRFPlayer, is evaluated on dynamic scenes captured by single hand-held cameras and multi-camera arrays, achieving comparable or superior rendering performance in terms of quality and speed comparable to recent state-of-the-art methods, achieving reconstruction in 10 seconds per frame and interactive rendering. Project website: https://bit.ly/nerfplayer.

In this paper, we present an efficient and robust deep learning solution for novel view synthesis of complex scenes. In our approach, a 3D scene is represented as a light field, i.e., a set of rays, each of which has a corresponding color when reaching the image plane. For efficient novel view rendering, we adopt a two-plane parameterization of the light field, where each ray is characterized by a 4D parameter. We then formulate the light field as a function that indexes rays to corresponding color values. We train a deep fully connected network to optimize this implicit function and memorize the 3D scene. Then, the scene-specific model is used to synthesize novel views. Different from previous light field approaches which require dense view sampling to reliably render novel views, our method can render novel views by sampling rays and querying the color for each ray from the network directly, thus enabling high-quality light field rendering with a sparser set of training images. Per-ray depth can be optionally predicted by the network, thus enabling applications such as auto refocus. Our novel view synthesis results are comparable to the state-of-the-arts, and even superior in some challenging scenes with refraction and reflection. We achieve this while maintaining an interactive frame rate and a small memory footprint.

Object detection and instance segmentation with a large number of object categories and long-tailed data distribution are challenging for most existing deep learning models. As the number of classes increases, the outputs of a classifier become sensitive to likely noisy logits, which can easily result in an incorrect recognition. To alleviate the large-vocabulary problem, we cluster fine-grained classes into coarser parent classes and then build a classification tree to classify an object into a fine-grained class via its parent class. Because the number of parent class is much fewer, their logits are more stable to suppress the wrong/noisy logits existed in the fine-grained class nodes. Due to a variety of ways for clustering fine-grained classes into parent classes, we can further construct multiple trees to build a classification forest where each single tree contributes its vote to the fine-grained classification. Moreover, a simple yet effective resampling method, termed as NMS Resampling, is proposed aiming at solving the long tail (data imbalance) problem. Our method, coined as ForestDet, serves as a plug-and-play module, which can be readily employed in both one-stage and two-stage object recognition models for recognizing more than 1000 categories. Extensive experiments are conducted on the large vocabulary dataset LVIS. Compared to the Mask R-CNN baseline, our two-stage counterpart Forest R-CNN significantly boosts the performance by 11.5% and 3.9% AP improvements on the rare categories and overall categories, respectively. Compared to the RetinaNet baseline, our one-stage counterpart Forest RetinaNet improves 2.1% AP on overall categories. Moreover, we achieve state-of-the-art results on the LVIS dataset.Code and models are available athttps://github.com/JialianW/Forest_RCNN.

When adopting deep neural networks for a new vision task, a common practice is to start with fine-tuning some off-the-shelf well-trained network models from the community. Since a new task may require training a different network architecture with new domain data, taking advantage of off-the-shelf models is not trivial and generally requires considerable try-and-error and parameter tuning. In this paper, we denote a well-trained model as a teacher network and a model for the new task as a student network. We aim to ease the efforts of transferring knowledge from the teacher to the student network, robust to the gaps between their network architectures, domain data, and task definitions. Specifically, we propose a hybrid forward scheme in training the teacher-student models, alternately updating layer weights of the student model. The key merit of our hybrid forward scheme is on the dynamical balance between the knowledge transfer loss and task specific loss in training. We demonstrate the effectiveness of our method on a variety of tasks, e.g., model compression, segmentation, and detection, under a variety of knowledge transfer settings.

Most online multi-object trackers perform object detection stand-alone in a neural net without any input from tracking. In this paper, we present a new online joint detection and tracking model, TraDeS (TRAck to DEtect and Segment), exploiting tracking clues to assist detection end-to-end. TraDeS infers object tracking offset by a cost volume, which is used to propagate previous object features for improving current object detection and segmentation. Effectiveness and superiority of TraDeS are shown on 4 datasets, including MOT (2D tracking), nuScenes (3D tracking), MOTS and Youtube-VIS (instance segmentation tracking). Project page: https://jialianwu.com/projects/TraDeS.html.

Multi-view pedestrian detection aims to predict a bird’s eye view (BEV) occupancy map from multiple camera views. This task is confronted with two challenges: how to establish the 3D correspondences from views to the BEV map and how to assemble occupancy information across views. In this paper, we propose a novel Stacked HOmography Transformations (SHOT) approach, which is motivated by approximating projections in 3D world coordinates via a stack of homographies. We first construct a stack of transformations for projecting views to the ground plane at different height levels. Then we design a soft selection module so that the network learns to predict the likelihood of the stack of transformations. Moreover, we provide an in-depth theoretical analysis on constructing SHOT and how well SHOT approximates projections in 3D world coordinates. SHOT is empirically verified to be capable of estimating accurate correspondences from individual views to the BEV map, leading to new state-of-the-art performance on standard evaluation benchmarks.

Given a set of multiple view videos, which records the motion trajectory of an object, we propose to find out the objects' kinematic formulas with neural rendering techniques. For example, if the input multiple view videos record the free fall motion of an object with different initial speed v, the network aims to learn its kinematics: Δ=vt-1over 2 gt2, where Δ, g and t are displacement, gravitational acceleration and time. To achieve this goal, we design a novel framework consisting of a motion network and a differentiable renderer. For the differentiable renderer, we employ Neural Radiance Field (NeRF) since the geometry is implicitly modeled by querying coordinates in the space. The motion network is composed of a series of blending functions and linear weights, enabling us to analytically derive the kinematic formulas after training. The proposed framework is trained end to end and only requires knowledge of cameras' intrinsic and extrinsic parameters. To validate the proposed framework, we design three experiments to demonstrate its effectiveness and extensibility. The first experiment is the video of free fall and the framework can be easily combined with the principle of parsimony, resulting in the correct free fall kinematics. The second experiment is on the large angle pendulum which does not have analytical kinematics. We use the differential equation controlling pendulum dynamics as a physical prior in the framework and demonstrate that the convergence speed becomes much faster. Finally, we study the explosion animation and demonstrate that our framework can well handle such black-box-generated motions.

Multi-label image classification aims to predict multiple labels for a single image. However, the difficulties of predicting different labels may vary dramatically due to semantic variations of the label as well as the image context. Direct learning of multi-label classification models has the risk of being biased and overfitting those difficult labels, e.g., deep network based classifiers are over-trained on the difficult labels, therefore, lead to false-positive errors of those difficult labels during testing. To handle difficult labels of multi-label image classification, we propose to calibrate the model, which not only predicts the labels but also estimates the uncertainty of the prediction. With the new calibration branch of the network, the classification model is trained with the pick-all-labels normalized loss and optimized pertaining to the number of positive labels. Moreover, to improve performance on difficult labels, instead of annotating them, we leverage the calibrated model as the teacher network and teach the student network about handling difficult labels via uncertainty distillation. Our proposed uncertainty distillation teaches the student network which labels are highly uncertain through prediction distribution distillation, and locates the image regions that cause such uncertain predictions through uncertainty attention distillation. Conducting extensive evaluations on benchmark datasets, we demonstrate that our proposed uncertainty distillation is valuable to handle difficult labels of multi-label image classification.

Existing human models, e.g., SMPL and STAR, represent 3D geometry of a human body in the form of a polygon mesh obtained by deforming a template mesh according to a set of shape and pose parameters. The appearance, however, is not directly modeled by most existing human models. We present a novel 3D human model that faithfully models both the 3D geometry and the appearance of a clothed human body with a continuous volumetric representation, i.e., volume densities and emitted colors of continuous 3D locations in the volume encompassing the human body. In contrast to the mesh-based representation whose resolution is limited by a mesh's fixed number of polygons, our volumetric representation does not limit the resolution of our model. Moreover, our volumetric represen-tation can be rendered via differentiable volume rendering, thus enabling us to train the model only using 2D images (without using ground truth 3D geometries of human bodies) by minimizing a loss function which measures the differences between rendered images and ground truth images. On the contrary, existing human models are trained using ground truth 3D geometries of human bodies. Thanks to the ability of our model to jointly model both the geometries and the appearances of clothed people, our model can benefit applications including human image synthesis, gaming and 3D television and telepresence.

In this paper, we propose a novel nonlocal patch tensor-based visual data completion algorithm and analyze its potential problems. Our algorithm consists of two steps: the first step is initializing the image with triangulation-based linear interpolation and the second step is grouping similar nonlocal patches as a tensor then applying the proposed tensor completion technique. Specifically, with treating a group of patch matrices as a tensor, we impose the low-rank constraint on the tensor through the recently proposed tensor nuclear norm. Moreover, we observe that after the first interpolation step, the image gets blurred and, thus, the similar patches we have found may not exactly match the reference. We name the problem "Patch Mismatch," and then in order to avoid the error caused by it, we further decompose the patch tensor into a low-rank tensor and a sparse tensor, which means the accepted horizontal strips in mismatched patches. Furthermore, our theoretical analysis shows that the error caused by Patch Mismatch can be decomposed into two components, one of which can be bounded by a reasonable assumption named local patch similarity, and the other part is lower than that using matrix completion. Extensive experimental results on real-world datasets verify our method's superiority to the state-of-the-art tensor-based image inpainting methods.

Despite the previous success of object analysis, detecting and segmenting a large number of object categories with a long-tailed data distribution remains a challenging problem and is less investigated. For a large-vocabulary classifier, the chance of obtaining noisy logits is much higher, which can easily lead to a wrong recognition. In this paper, we exploit prior knowledge of the relations among object categories to cluster fine-grained classes into coarser parent classes, and construct a classification tree that is responsible for parsing an object instance into a fine-grained category via its parent class. In the classification tree, as the number of parent class nodes are significantly less, their logits are less noisy and can be utilized to suppress the wrong/noisy logits existed in the fine-grained class nodes. As the way to construct the parent class is not unique, we further build multiple trees to form a classification forest where each tree contributes its vote to the fine-grained classification. To alleviate the imbalanced learning caused by the long-tail phenomena, we propose a simple yet effective resampling method, NMS Resampling, to re-balance the data distribution. Our method, termed as Forest R-CNN, can serve as a plug-and-play module being applied to most object recognition models for recognizing more than 1000 categories. Extensive experiments are performed on the large vocabulary dataset LVIS. Compared with the Mask R-CNN baseline, the Forest R-CNN significantly boosts the performance with 11.5% and 3.9% AP improvements on the rare categories and overall categories, respectively. Moreover, we achieve state-of-the-art results on the LVIS dataset. Code is available at https://github.com/JialianW/Forest_RCNN.

Synthetic visual data refers to the data automatically rendered by the mature computer graphic algorithms. With the rapid development of these techniques, we can now collect photo-realistic synthetic images with accurate pixel-level annotations without much effort. However, due to the domain gaps between synthetic data and real data, in terms of not only visual appearance but also label distribution, directly applying models trained on synthetic images to real ones can hardly yield satisfactory performance. Since the collection of accurate labels for real images is very laborious and time-consuming, developing algorithms which can learn from synthetic images is of great significance. In this paper, we propose a novel framework, namely Active Pseudo-Labeling (APL), to reduce the domain gaps between synthetic images and real images. In APL framework, we first predict pseudo-labels for the unlabeled real images in the target domain by actively adapting the style of the real images to source domain. Specifically, the style of real images is adjusted via a novel task guided generative model, and then pseudo-labels are predicted for these actively adapted images. Lastly, we fine-tune the source-trained model in the pseudo-labeled target domain, which helps to fit the distribution of the real data. Experiments on both semantic segmentation and object detection tasks with several challenging benchmark data sets demonstrate the priority of our proposed method compared to the existing state-of-the-art approaches.

Deep supervised hashing has become an active topic in information retrieval. It generates hashing bits by the output neurons of a deep hashing network. During binary discretization, there often exists much redundancy between hashing bits that degenerates retrieval performance in terms of both storage and accuracy. This paper proposes a simple yet effective Neurons Merging Layer (NMLayer) for deep supervised hashing. A graph is constructed to represent the redundancy relationship between hashing bits that is used to guide the learning of a hashing network. Specifically, it is dynamically learned by a novel mechanism defined in our active and frozen phases. According to the learned relationship, the NMLayer merges the redundant neurons together to balance the importance of each output neuron. Moreover, multiple NMLayers are progressively trained for a deep hashing network to learn a more compact hashing code from a long redundant code. Extensive experiments on four datasets demonstrate that our proposed method outperforms state-of-the-art hashing methods.

In this paper, we propose a novel image inpainting framework consisting of an interpolation step and a low-rank tensor completion step. More specifically, we first initial the image with triangulation-based linear interpolation, and then we find similar patches for each missing-entry centered patch. Treating a group of patch matrices as a tensor, we employ the recently proposed effective t-SVD tensor completion algorithm with a warm start strategy to inpaint it. We observe that the interpolation step is such a rough initialization that the similar patch we found may not exactly match with the reference, so we name the problem as Patch Mismatch and analyse the error caused by it thoroughly. Our theoretical analysis shows that the error caused by Patch Mismatch can be decomposed into two components, one of which can be bounded by a reasonable assumption named local patch similarity, and another part is lower than that using matrix. Experiments on real images verify our method's superiority to the state-of-the-art inpainting methods.