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.

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.

We present MOFI, Manifold OF Images, a new vision foundation model designed to learn image representations from noisy entity annotated images. MOFI differs from previous work in two key aspects: 1. pre-training data, and 2. training recipe. Regarding data, we introduce a new approach to automatically assign entity labels to images from noisy image-text pairs. Our approach involves employing a named entity recognition model to extract entities from the alt-text, and then using a CLIP model to select the correct entities as labels of the paired image. It's a simple, cost-effective method that can scale to handle billions of web-mined image-text pairs. Through this method, we have created Image-to-Entities (I2E), a new dataset with 1 billion images and 2 million distinct entities, covering rich visual concepts in the wild. Building upon the I2E dataset, we study different training recipes like supervised pre-training, contrastive pre-training, and multi-task learning. For constrastive pre-training, we treat entity names as free-form text, and further enrich them with entity descriptions. Experiments show that supervised pre-training with large-scale fine-grained entity labels is highly effective for image retrieval tasks, and multi-task training further improves the performance. The final MOFI model achieves 86.66\% mAP on the challenging GPR1200 dataset, surpassing the previous state-of-the-art performance of 72.19% from OpenAI's CLIP model. Further experiments on zero-shot and linear probe image classification also show that MOFI outperforms a CLIP model trained on the original image-text data, demonstrating the effectiveness of the I2E dataset in learning strong image representations. We release our code and model weights at https://github.com/apple/ml-mofi.

Chen Chen, Bowen Zhang, Liangliang Cao, Jiguang Shen, Tom Gunter, Albin Jose, Alexander Toshev, Yantao Zheng, Jonathon Shlens, Ruoming Pang, Yinfei Yang. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Xiaotao Gu, Liyuan Liu, Hongkun Yu, Jing Li, Chen Chen, Jiawei Han. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Databases and data warehouse systems have been evolving from handling normalized spreadsheets stored in relational databases, to managing and analyzing diverse application-oriented data with complex interconnecting structures. Responding to this emerging trend, graphs have been growing rapidly and showing their critical importance in many applications, such as the analysis of XML, social networks, Web, biological data, multimedia data and spatiotemporal data. Can we extend useful functions of databases and data warehouse systems to handle graph structured data? In particular, OLAP (On-Line Analytical Processing) has been a popular tool for fast and user-friendly multi-dimensional analysis of data warehouses. Can we OLAP graphs? Unfortunately, to our best knowledge, there are no OLAP tools available that can interactively view and analyze graph data from different perspectives and with multiple granularities. In this paper, we argue that it is critically important to OLAP graph structured data and propose a novel Graph OLAP framework. According to this framework, given a graph dataset with its nodes and edges associated with respective attributes, a multi-dimensional model can be built to enable efficient on-line analytical processing so that any portions of the graphs can be generalized/specialized dynamically, offering multiple, versatile views of the data. The contributions of this work are three-fold. First, starting from basic definitions, i.e ., what are dimensions and measures in the Graph OLAP scenario, we develop a conceptual framework for data cubes on graphs. We also look into different semantics of OLAP operations, and classify the framework into two major subcases: informational OLAP and topological OLAP . Second, we show how a graph cube can be materialized by calculating a special kind of measure called aggregated graph and how to implement it efficiently. This includes both full materialization and partial materialization where constraints are enforced to obtain an iceberg cube . As we can see, due to the increased structural complexity of data, aggregated graphs that depend on the underlying “network” properties of the graph dataset are much harder to compute than their traditional OLAP counterparts. Third, to provide more flexible, interesting and informative OLAP of graphs, we further propose a discovery-driven multi-dimensional analysis model to ensure that OLAP is performed in an intelligent manner, guided by expert rules and knowledge discovery processes. We outline such a framework and discuss some challenging research issues for discovery-driven Graph OLAP.

Graphs are prevalent in many domains such as Bioinformatics, social networks, Web and cyber-security. Graph pattern mining has become an important tool in the management and analysis of complexly structured data, where example applications include indexing, clustering and classification. Existing graph mining algorithms have achieved great success by exploiting various properties in the pattern space . Unfortunately, due to the fundamental role subgraph isomorphism plays in these methods, they may all enter into a pitfall when the cost to enumerate a huge set of isomorphic embeddings blows up, especially in large graphs. The solution we propose for this problem resorts to reduction on the data space . For each graph, we build a summary of it and mine this shrunk graph instead. Compared to other data reduction techniques that either reduce the number of transactions or compress between transactions, this new framework, called Summarize-Mine, suggests a third path by compressing within transactions . Summarize-Mine is effective in cutting down the size of graphs, thus decreasing the embedding enumeration cost. However, compression might lose patterns at the same time. We address this issue by generating randomized summaries and repeating the process for multiple rounds, where the main idea is that true patterns are unlikely to miss from all rounds. We provide strict probabilistic guarantees on pattern loss likelihood. Experiments on real malware trace data show that Summarize-Mine is very efficient, which can find interesting malware fingerprints that were not revealed previously.

Nowadays, most business, administration, and/or scientific databases contain both structured attributes and text attributes. We call a database that consists of both multidimensional structured data and narrative text data as multidimensional text database . Searching, OLAP, and mining such databases pose many research challenges. To enhance the power of data analysis, interesting entities and relationships can be extracted from such databases to derive heterogeneous information networks, which in turn will substantially increase the power and flexibility of data exploration in such databases. Based on our previous studies on TextCube [1], TopicCube [2], and information network analysis, such as RankClus [3] and NetClus [4], we construct iNextCube , an i nformation- N etwork- e nhanced te xt Cube . In this demo, we show the power of iNextCube in the search and analysis of two multidimensional text databases: (i) a DBLP-based CS bibliographic database, and (ii) an online news database.

In the past, quite a few fast algorithms have been developed to mine frequent patterns over graph data, with the large spectrum covering many variants of the problem. However, the real bottleneck for knowledge discovery on graphs is neither efficiency nor scalability, but the usability of patterns that are mined out. Currently, what the state-of-art techniques give is a lengthy list of exact patterns, which are undesirable in the following two aspects: (1) on the micro side, due to various inherent noises or data diversity, exact patterns are usually not too useful in many real applications; and (2) on the macro side, the rigid structural requirement being posed often generates an excessive amount of patterns that are only slightly different from each other, which easily overwhelm the users.

OLAP (On-Line Analytical Processing) is an important notion in data analysis. Recently, more and more graph or networked data sources come into being. There exists a similar need to deploy graph analysis from different perspectives and with multiple granularities. However, traditional OLAP technology cannot handle such demands because it does not consider the links among individual data tuples. In this paper, we develop a novel graph OLAP framework, which presents a multi-dimensional and multi-level view over graphs. The contributions of this work are two-fold. First, starting from basic definitions, i.e., what are dimensions and measures in the graph OLAP scenario, we develop a conceptual framework for data cubes on graphs. We also look into different semantics of OLAP operations, and classify the framework into two major subcases: informational OLAP and topological OLAP. Then, with more emphasis on informational OLAP (topological OLAP will be covered in a future study due to the lack of space), we show how a graph cube can be materialized by calculating a special kind of measure called aggregated graph and how to implement it efficiently. This includes both full materialization and partial materialization where constraints are enforced to obtain an iceberg cube. We can see that the aggregated graphs, which depend on the graph properties of underlying networks, are much harder to compute than their traditional OLAP counterparts, due to the increased structural complexity of data. Empirical studies show insightful results on real datasets and demonstrate the efficiency of our proposed optimizations.

Recently, there arise a large number of graphs with massive sizes and complex structures in many new applications, such as biological networks, social networks, and the Web, demanding powerful data mining methods. Due to inherent noise or data diversity, it is crucial to address the issue of approximation, if one wants to mine patterns that are potentially interesting with tolerable variations. In this paper, we investigate the problem of mining frequent approximate patterns from a massive network and propose a method called gApprox. gApprox not only finds approximate network patterns, which is the key for many knowledge discovery applications on structural data, but also enriches the library of graph mining methodologies by introducing several novel techniques such as: (1) a complete and redundancy-free strategy to explore the new pattern space faced by gApprox; and (2) transform "frequent in an approximate sense" into an anti-monotonic constraint so that it can be pushed deep into the mining process. Systematic empirical studies on both real and synthetic data sets show that frequent approximate patterns mined from the worm protein-protein interaction network are biologically interesting and gApprox is both effective and efficient.

In this paper, we discuss the inverse classification problem, in which we desire to define the features of an incomplete record in such a way that will result in a desired class label. Such an approach is useful in applications in which it is an objective to determine a set of actions to be taken in order to guide the data mining application towards a desired solution. This system can be used for a variety of decision support applications which have pre-determined task criteria.

Along with the blossom of open source projects comes the convenience for software plagiarism. A company, if less self-disciplined, may be tempted to plagiarize some open source projects for its own products. Although current plagiarism detection tools appear sufficient for academic use, they are nevertheless short for fighting against serious plagiarists. For example, disguises like statement reordering and code insertion can effectively confuse these tools. In this paper, we develop a new plagiarism detection tool, called GPLAG, which detects plagiarism by mining program dependence graphs (PDGs). A PDG is a graphic representation of the data and control dependencies within a procedure. Because PDGs are nearly invariant during plagiarism, GPLAG is more effective than state-of-the-art tools for plagiarism detection. In order to make GPLAG scalable to large programs, a statistical lossy filter is proposed to prune the plagiarism search space. Experiment study shows that GPLAG is both effective and efficient: It detects plagiarism that easily slips over existing tools, and it usually takes a few seconds to find (simulated) plagiarism in programs having thousands of lines of code.