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.

Recent advancements in multimodal models highlight the value of rewritten captions for improving performance, yet key challenges remain. For example, while synthetic captions often provide superior quality and image-text alignment, it is not clear whether they can fully replace AltTexts: the role of synthetic captions and their interaction with original web-crawled AltTexts in pre-training is still not well understood. Moreover, different multimodal foundation models may have unique preferences for specific caption formats, but efforts to identify the optimal captions for each model remain limited. In this work, we propose a novel, controllable, and scalable captioning pipeline designed to generate diverse caption formats tailored to various multimodal models. By examining short synthetic captions (SSC) and descriptive synthetic captions (DSC) as case studies, we systematically explore their effects and interactions with AltTexts across models such as CLIP, multimodal LLMs, and diffusion models. Our findings reveal that a hybrid approach that keeps both synthetic captions and AltTexts can outperform the use of synthetic captions alone, improving both alignment and performance, with each model demonstrating preferences for particular caption formats. This comprehensive analysis provides valuable insights into optimizing captioning strategies, thereby advancing the pre-training of multimodal foundation models.

Instruction-based image editing improves the controllability and flexibility of image manipulation via natural commands without elaborate descriptions or regional masks. However, human instructions are sometimes too brief for current methods to capture and follow. Multimodal large language models (MLLMs) show promising capabilities in cross-modal understanding and visual-aware response generation via LMs. We investigate how MLLMs facilitate edit instructions and present MLLM-Guided Image Editing (MGIE). MGIE learns to derive expressive instructions and provides explicit guidance. The editing model jointly captures this visual imagination and performs manipulation through end-to-end training. We evaluate various aspects of Photoshop-style modification, global photo optimization, and local editing. Extensive experimental results demonstrate that expressive instructions are crucial to instruction-based image editing, and our MGIE can lead to a notable improvement in automatic metrics and human evaluation while maintaining competitive inference efficiency.

With the wide and deep adoption of deep learning models in real applications, there is an increasing need to model and learn the representations of the neural networks themselves. These models can be used to estimate attributes of different neural network architectures such as the accuracy and latency, without running the actual training or inference tasks. In this paper, we propose a neural architecture representation model that can be used to estimate these attributes holistically. Specifically, we first propose a simple and effective tokenizer to encode both the operation and topology information of a neural network into a single sequence. Then, we design a multi-stage fusion transformer to build a compact vector representation from the converted sequence. For efficient model training, we further propose an information flow consistency augmentation and correspondingly design an architecture consistency loss, which brings more benefits with less augmentation samples compared with previous random augmentation strategies. Experiment results on NAS-Bench-101, NAS-Bench-201, DARTS search space and NNLQP show that our proposed framework can be used to predict the aforementioned latency and accuracy attributes of both cell architectures and whole deep neural networks, and achieves promising performance. Code is available at https://github.com/yuny220/NAR-Former.

Transformers have shown great potential in various computer vision tasks. By borrowing design concepts from transformers, many studies revolutionized CNNs and showed remarkable results. This paper falls in this line of studies. Specifically, we propose a new convolutional neural network, ParCNetV2, that extends the research line of ParCNetV1 by bridging the gap between CNN and ViT. It introduces two key designs: 1) Oversized Convolution (OC) with twice the size of the input, and 2) Bifurcate Gate Unit (BGU) to ensure that the model is input adaptive. Fusing OC and BGU in a unified CNN, ParCNetV2 is capable of flexibly extracting global features like ViT, while maintaining lower latency and better accuracy. Extensive experiments demonstrate the superiority of our method over other convolutional neural networks and hybrid models that combine CNNs and transformers. The code are publicly available at https://github.com/XuRuihan/ParCNetV2.

Currently, one main research line in designing a more efficient vision transformer is reducing the computational cost of self attention modules by adopting sparse attention or using local attention windows. In contrast, we propose a different approach that aims to improve the performance of transformer-based architectures by densifying the attention pattern. Specifically, we proposed forward cross attention for hybrid vision transformer (FcaFormer), where tokens from previous blocks in the same stage are secondary used. To achieve this, the FcaFormer leverages two innovative components: learnable scale factors (LSFs) and a token merge and enhancement module (TME). The LSFs enable efficient processing of cross tokens, while the TME generates representative cross tokens. By integrating these components, the proposed FcaFormer enhances the interactions of tokens across blocks with potentially different semantics, and encourages more information flows to the lower levels. Based on the forward cross attention (Fca), we have designed a series of FcaFormer models that achieve the best trade-off between model size, computational cost, memory cost, and accuracy. For example, without the need for knowledge distillation to strengthen training, our FcaFormer achieves 83.1% top-1 accuracy on Imagenet with only 16.3 million parameters and about 3.6 billion MACs. This saves almost half of the parameters and a few computational costs while achieving 0.7% higher accuracy compared to distilled EfficientFormer. Code is available at https://github.com/hkzhang-git/FcaFormer

While 3D object-centered shape-based models are appealing in comparison with 2D viewer-centered appearance-based models for their lower model complexities and potentially better view generalizabilities, the learning and inference of 3D models has been much less studied in the recent literature due to two factors: i) the enormous complexities of 3D shapes in geometric space; and ii) the gap between 3D shapes and their appearances in images. This paper aims at tackling the two problems by studying an And-Or Tree (AoT) representation that consists of two parts: i) a geometry-AoT quantizing the geometry space, i.e. the possible compositions of 3D volumetric parts and 2D surfaces within the volumes; and ii) an appearance-AoT quantizing the appearance space, i.e. the appearance variations of those shapes in different views. In this AoT, an And-node decomposes an entity into constituent parts, and an Or-node represents alternative ways of decompositions. Thus it can express a combinatorial number of geometry and appearance configurations through small dictionaries of 3D shape primitives and 2D image primitives. In the quantized space, the problem of learning a 3D object template is transformed to a structure search problem which can be efficiently solved in a dynamic programming algorithm by maximizing the information gain. We focus on learning 3D car templates from the AoT and collect a new car dataset featuring more diverse views. The learned car templates integrate both the shape-based model and the appearance-based model to combine the benefits of both. In experiments, we show three aspects: 1) the AoT is more efficient than the frequently used octree method in space representation; 2) the learned 3D car template matches the state-of-the art performances on car detection and pose estimation in a public multi-view car dataset; and 3) in our new dataset, the learned 3D template solves the joint task of simultaneous object detection, pose/view estimation, and part localization. It can generalize over unseen views and performs better than the version 5 of the DPM model in terms of object detection and semantic part localization.

This paper proposes an unsupervised method for learning dictionaries of hierarchical compositional models for representing natural images. Each model is in the form of a template that consists of a small group of part templates that are allowed to shift their locations and orientations relative to each other, and each part template is in turn a composition of Gabor wavelets that are also allowed to shift their locations and orientations relative to each other. Given a set of unannotated training images, a dictionary of such hierarchical templates are learned so that each training image can be represented by a small number of templates that are spatially translated, rotated and scaled versions of the templates in the learned dictionary. The learning algorithm iterates between the following two steps: (1) Image encoding by a template matching pursuit process that involves a bottom-up template matching sub-process and a top-down template localization sub-process. (2) Dictionary re-learning by a shared matching pursuit process. Experimental results show that the proposed approach is capable of learning meaningful templates, and the learned templates are useful for tasks such as domain adaption and image cosegmentation.

We investigate an inhomogeneous version of the FRAME (Filters, Random field, And Maximum Entropy) model and apply it to modeling object patterns. The inhomogeneous FRAME is a non-stationary Markov random field model that reproduces the observed marginal distributions or statistics of filter responses at all the different locations, scales and orientations. Our experiments show that the inhomogeneous FRAME model is capable of generating a wide variety of object patterns in natural images. We then propose a sparsified version of the inhomogeneous FRAME model where the model reproduces observed statistical properties of filter responses at a small number of selected locations, scales and orientations. We propose to select these locations, scales and orientations by a shared sparse coding scheme, and we explore the connection between the sparse FRAME model and the linear additive sparse coding model. Our experiments show that it is possible to learn sparse FRAME models in unsupervised fashion and the learned models are useful for object classification.

Occlusion presents a challenge for detecting objects in real world applications. To address this issue, this paper models object occlusion with an AND-OR structure which (i) represents occlusion at semantic part level, and (ii) captures the regularities of different occlusion configurations (i.e., the different combinations of object part visibilities). This paper focuses on car detection on street. Since annotating part occlusion on real images is time-consuming and error-prone, we propose to learn the the AND-OR structure automatically using synthetic images of CAD models placed at different relative positions. The model parameters are learned from real images under the latent structural SVM (LSSVM) framework. In inference, an efficient dynamic programming (DP) algorithm is utilized. In experiments, we test our method on both car detection and car view estimation. Experimental results show that (i) Our CAD simulation strategy is capable of generating occlusion patterns for real scenarios, (ii) The proposed AND-OR structure model is effective for modeling occlusions, which outperforms the deformable part-based model (DPM) DPM, voc5 in car detection on both our self-collected street parking dataset and the Pascal VOC 2007 car dataset pascal-voc-2007}, (iii) The learned model is on-par with the state-of-the-art methods on car view estimation tested on two public datasets.

This paper presents a method for learning 3D object templates from view labeled object images. The 3D template is defined in a joint appearance and geometry space composed of deformable planar part templates placed at different 3D positions and orientations. Appearance of each part template is represented by Gabor filters, which are hierarchically grouped into line segments and geometric shapes. AND-OR trees are further used to quantize the possible geometry and appearance of part templates, so that learning can be done on a subsampled discrete space. Using information gain as a criterion, the best 3D template can be searched through the AND-OR trees using one bottom-up pass and one top-down pass. Experiments on a new car dataset with diverse views show that the proposed method can learn meaningful 3D car templates, and give satisfactory detection and view estimation performance. Experiments are also performed on a public car dataset, which show comparable performance with recent methods.

This paper proposes a sparse image representation using deformable templates of simple geometric structures that are commonly observed in images of natural scenes. These deformable templates include active curve templates and active corner templates. An active curve template is a composition of Gabor wavelet elements placed with equal spacing on a straight line segment or a circular arc segment of constant curvature, where each Gabor wavelet element is allowed to locally shift its location and orientation, so that the original line and arc segment of the active curve template can be deformed to fit the observed image. An active corner or angle template is a composition of two active curve templates that share a common end point, and the active curve templates are allowed to vary their overall lengths and curvatures, so that the original corner template can deform to match the observed image. This paper then proposes a hierarchical computational architecture of summax maps that pursues a sparse representation of an image by selecting a small number of active curve and corner templates from a dictionary of all such templates. Experiments show that the proposed method is capable of finding sparse representations of natural images. It is also shown that object templates can be learned by selecting and composing active curve and corner templates.

This paper presents a method learning mixed templates for view invariant object recognition. The template is composed of 3D and 2D primitives which are stick-like elements defined in 3D and 2D spaces respectively. The primitives are allowed to perturb within a local range to account for instance variations of an object category. When projected onto images, the appearance of these primitives are represented by Gabor filters. Both 3D and 2D primitives have parameters describing their visible range in a viewing hemisphere. Our algorithm sequentially selects primitives and builds a probabilistic model using the selected primitives. The order of this sequential selection is decided by the information gains of primitives, which can be estimated together with the visible range parameter efficiently. In experiments, we evaluate performance of the learned 3D templates on car recognition and pose estimation. We also show that the algorithm can learn intuitive mixed templates on various object categories, which suggests that our method could be used as a numerical method to justify the debate over viewer-centered and object-centered representations.

This paper proposes an algorithm for abandoned object detection based on generative model of low level features. First, suspected blobs are detected by foreground detection and pixel variance thresholding. Then several low level features on blobs are calculated to remove false alarms, which include pixel variance over time, edge intensity score, edge variance over time, foreground completeness, histogram contrast with respect to surrounding background, bag color model and priors of height and width. The last two features are used as threshold. For other features, their log probability ratio between distributions on positive examples and negative ones are tted by sigmoid function. At last, the fitted log probability ratios are weighted to construct a classifier. The algorithm has been verified in 29 challenging scenes and produces very low false alarms and missing detection.

This paper presents a novel integrated background model for video surveillance. Our model uses a primal sketch representation for image appearance and 3D scene geometry to capture the ground plane and major surfaces in the scene. The primal sketch model divides the background image into three types of regions - flat, sketchable and textured. The three types of regions are modeled respectively by mixture of Gaussians, image primitives and LBP histograms. We calibrate the camera and recover important planes such as ground, horizontal surfaces, walls, stairs in the 3D scene, and use geometric information to predict the sizes and locations of foreground blobs to further reduce false alarms. Compared with the state-of-the-art background modeling methods, our approach is more effective, especially for indoor scenes where shadows, highlights and reflections of moving objects and camera exposure adjusting usually cause problems. Experiment results demonstrate that our approach improves the performance of background/foreground separation at pixel level, and the integrated video surveillance system at the object and trajectory level.

An interactive scene annotation tool for video surveillance is presented in this paper. The annotation process is divided into three stages. (1) camera rough calibration;(2) calibration refinement; (3) major surfaces annotation. Inputs are then rendered in a 3D environment, which again help users check calibration accuracy and annotation correctness. Experiments show that this tool is easy to use and attains acceptable annotation accuracy. The interactive procedure helps users without knowledge in computer vision to complete camera calibration as well as surface annotation.