Guan Pang

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Thomas Hayes, Songyang Zhang, Xi Yin et al.

Multimodal video-audio-text understanding and generation can benefit from datasets that are narrow but rich. The narrowness allows bite-sized challenges that the research community can make progress on. The richness ensures we are making progress along the core challenges. To this end, we present a large-scale video-audio-text dataset MUGEN, collected using the open-sourced platform game CoinRun [11]. We made substantial modifications to make the game richer by introducing audio and enabling new interactions. We trained RL agents with different objectives to navigate the game and interact with 13 objects and characters. This allows us to automatically extract a large collection of diverse videos and associated audio. We sample 375K video clips (3.2s each) and collect text descriptions from human annotators. Each video has additional annotations that are extracted automatically from the game engine, such as accurate semantic maps for each frame and templated textual descriptions. Altogether, MUGEN can help progress research in many tasks in multimodal understanding and generation. We benchmark representative approaches on tasks involving video-audio-text retrieval and generation. Our dataset and code are released at: https://mugen-org.github.io/.

Songwei Ge, Thomas Hayes, Harry Yang et al.

Videos are created to express emotion, exchange information, and share experiences. Video synthesis has intrigued researchers for a long time. Despite the rapid progress driven by advances in visual synthesis, most existing studies focus on improving the frames' quality and the transitions between them, while little progress has been made in generating longer videos. In this paper, we present a method that builds on 3D-VQGAN and transformers to generate videos with thousands of frames. Our evaluation shows that our model trained on 16-frame video clips from standard benchmarks such as UCF-101, Sky Time-lapse, and Taichi-HD datasets can generate diverse, coherent, and high-quality long videos. We also showcase conditional extensions of our approach for generating meaningful long videos by incorporating temporal information with text and audio. Videos and code can be found at https://songweige.github.io/projects/tats/index.html.

Samaneh Azadi, Thomas Hayes, Akbar Shah et al.

Recent large-scale text-to-image generation models have made significant improvements in the quality, realism, and diversity of the synthesized images and enable users to control the created content through language. However, the personalization aspect of these generative models is still challenging and under-explored. In this work, we propose a pipeline that enables personalization of image generation with avatars capturing a user's identity in a delightful way. Our pipeline is zero-shot, avatar texture and style agnostic, and does not require training on the avatar at all - it is scalable to millions of users who can generate a scene with their avatar. To render the avatar in a pose faithful to the given text prompt, we propose a novel text-to-3D pose diffusion model trained on a curated large-scale dataset of in-the-wild human poses improving the performance of the SOTA text-to-motion models significantly. We show, for the first time, how to leverage large-scale image datasets to learn human 3D pose parameters and overcome the limitations of motion capture datasets.

Mei-Yuh Hwang, Yangyang Shi, Ankit Ramchandani et al.

This paper discusses the challenges of optical character recognition (OCR) on natural scenes, which is harder than OCR on documents due to the wild content and various image backgrounds. We propose to uniformly use word error rates (WER) as a new measurement for evaluating scene-text OCR, both end-to-end (e2e) performance and individual system component performances. Particularly for the e2e metric, we name it DISGO WER as it considers Deletion, Insertion, Substitution, and Grouping/Ordering errors. Finally we propose to utilize the concept of super blocks to automatically compute BLEU scores for e2e OCR machine translation. The small SCUT public test set is used to demonstrate WER performance by a modularized OCR system.

Siyan Zhao, Zhihui Xie, Mengchen Liu et al.

Knowledge distillation improves large language model (LLM) reasoning by compressing the knowledge of a teacher LLM to train smaller LLMs. On-policy distillation advances this approach by having the student sample its own trajectories while a teacher LLM provides dense token-level supervision, addressing the distribution mismatch between training and inference in off-policy distillation methods. However, on-policy distillation typically requires a separate, often larger, teacher LLM and does not explicitly leverage ground-truth solutions available in reasoning datasets. Inspired by the intuition that a sufficiently capable LLM can rationalize external privileged reasoning traces and teach its weaker self (i.e., the version without access to privileged information), we introduce On-Policy Self-Distillation (OPSD), a framework where a single model acts as both teacher and student by conditioning on different contexts. The teacher policy conditions on privileged information (e.g., verified reasoning traces) while the student policy sees only the question; training minimizes the per-token divergence between these distributions over the student's own rollouts. We demonstrate the efficacy of our method on multiple mathematical reasoning benchmarks, achieving 4-8x token efficiency compared to reinforcement learning methods such as GRPO and superior performance over off-policy distillation methods.

Adam Polyak, Amit Zohar, Andrew Brown et al. · meta-ai

We present Movie Gen, a cast of foundation models that generates high-quality, 1080p HD videos with different aspect ratios and synchronized audio. We also show additional capabilities such as precise instruction-based video editing and generation of personalized videos based on a user's image. Our models set a new state-of-the-art on multiple tasks: text-to-video synthesis, video personalization, video editing, video-to-audio generation, and text-to-audio generation. Our largest video generation model is a 30B parameter transformer trained with a maximum context length of 73K video tokens, corresponding to a generated video of 16 seconds at 16 frames-per-second. We show multiple technical innovations and simplifications on the architecture, latent spaces, training objectives and recipes, data curation, evaluation protocols, parallelization techniques, and inference optimizations that allow us to reap the benefits of scaling pre-training data, model size, and training compute for training large scale media generation models. We hope this paper helps the research community to accelerate progress and innovation in media generation models. All videos from this paper are available at https://go.fb.me/MovieGenResearchVideos.

Minghui Liao, Guan Pang, Jing Huang et al.

Recent end-to-end trainable methods for scene text spotting, integrating detection and recognition, showed much progress. However, most of the current arbitrary-shape scene text spotters use region proposal networks (RPN) to produce proposals. RPN relies heavily on manually designed anchors and its proposals are represented with axis-aligned rectangles. The former presents difficulties in handling text instances of extreme aspect ratios or irregular shapes, and the latter often includes multiple neighboring instances into a single proposal, in cases of densely oriented text. To tackle these problems, we propose Mask TextSpotter v3, an end-to-end trainable scene text spotter that adopts a Segmentation Proposal Network (SPN) instead of an RPN. Our SPN is anchor-free and gives accurate representations of arbitrary-shape proposals. It is therefore superior to RPN in detecting text instances of extreme aspect ratios or irregular shapes. Furthermore, the accurate proposals produced by SPN allow masked RoI features to be used for decoupling neighboring text instances. As a result, our Mask TextSpotter v3 can handle text instances of extreme aspect ratios or irregular shapes, and its recognition accuracy won't be affected by nearby text or background noise. Specifically, we outperform state-of-the-art methods by 21.9 percent on the Rotated ICDAR 2013 dataset (rotation robustness), 5.9 percent on the Total-Text dataset (shape robustness), and achieve state-of-the-art performance on the MSRA-TD500 dataset (aspect ratio robustness). Code is available at: https://github.com/MhLiao/MaskTextSpotterV3

Bolin Lai, Xiaoliang Dai, Lawrence Chen et al.

Generating instructional images of human daily actions from an egocentric viewpoint serves as a key step towards efficient skill transfer. In this paper, we introduce a novel problem -- egocentric action frame generation. The goal is to synthesize an image depicting an action in the user's context (i.e., action frame) by conditioning on a user prompt and an input egocentric image. Notably, existing egocentric action datasets lack the detailed annotations that describe the execution of actions. Additionally, existing diffusion-based image manipulation models are sub-optimal in controlling the state transition of an action in egocentric image pixel space because of the domain gap. To this end, we propose to Learn EGOcentric (LEGO) action frame generation via visual instruction tuning. First, we introduce a prompt enhancement scheme to generate enriched action descriptions from a visual large language model (VLLM) by visual instruction tuning. Then we propose a novel method to leverage image and text embeddings from the VLLM as additional conditioning to improve the performance of a diffusion model. We validate our model on two egocentric datasets -- Ego4D and Epic-Kitchens. Our experiments show substantial improvement over prior image manipulation models in both quantitative and qualitative evaluation. We also conduct detailed ablation studies and analysis to provide insights in our method. More details of the dataset and code are available on the website (https://bolinlai.github.io/Lego_EgoActGen/).

Siyan Zhao, Mengchen Liu, Jing Huang et al.

Masked diffusion large language models (dLLMs) are emerging as promising alternatives to autoregressive LLMs, offering competitive performance while supporting unique generation capabilities such as inpainting. We explore how inpainting can inform RL algorithm design for dLLMs. Aligning LLMs with reinforcement learning faces an exploration challenge: sparse reward signals and sample waste when models fail to discover correct solutions. While this inefficiency affects LLMs broadly, dLLMs offer a distinctive opportunity--their inpainting ability can guide exploration. We introduce IGPO (Inpainting Guided Policy Optimization), an RL framework that strategically inserts partial ground-truth reasoning traces during online sampling. Unlike providing full solutions, inpainting steers exploration toward promising trajectory spaces while preserving self-generated reasoning, bridging supervised fine-tuning and reinforcement learning. We apply IGPO to group-based optimization methods such as GRPO, where exploration failures cause zero advantages and gradients. IGPO restores meaningful gradients while improving sample efficiency. We also propose supervised fine-tuning on synthetically rewritten concise traces that better align with dLLM generation patterns. With additional techniques including entropy-based filtering, our training recipe yields substantial gains across three mathematical benchmarks--GSM8K, Math500, and AMC--achieving new state-of-the-art results for full-attention masked dLLMs.

David Yan, Winnie Zhang, Luxin Zhang et al.

We introduce animated stickers, a video diffusion model which generates an animation conditioned on a text prompt and static sticker image. Our model is built on top of the state-of-the-art Emu text-to-image model, with the addition of temporal layers to model motion. Due to the domain gap, i.e. differences in visual and motion style, a model which performed well on generating natural videos can no longer generate vivid videos when applied to stickers. To bridge this gap, we employ a two-stage finetuning pipeline: first with weakly in-domain data, followed by human-in-the-loop (HITL) strategy which we term ensemble-of-teachers. It distills the best qualities of multiple teachers into a smaller student model. We show that this strategy allows us to specifically target improvements to motion quality while maintaining the style from the static image. With inference optimizations, our model is able to generate an eight-frame video with high-quality, interesting, and relevant motion in under one second.

Qilong Zhangli, Jindong Jiang, Di Liu et al.

While diffusion models have significantly advanced the quality of image generation their capability to accurately and coherently render text within these images remains a substantial challenge. Conventional diffusion-based methods for scene text generation are typically limited by their reliance on an intermediate layout output. This dependency often results in a constrained diversity of text styles and fonts an inherent limitation stemming from the deterministic nature of the layout generation phase. To address these challenges this paper introduces SceneTextGen a novel diffusion-based model specifically designed to circumvent the need for a predefined layout stage. By doing so SceneTextGen facilitates a more natural and varied representation of text. The novelty of SceneTextGen lies in its integration of three key components: a character-level encoder for capturing detailed typographic properties coupled with a character-level instance segmentation model and a word-level spotting model to address the issues of unwanted text generation and minor character inaccuracies. We validate the performance of our method by demonstrating improved character recognition rates on generated images across different public visual text datasets in comparison to both standard diffusion based methods and text specific methods.

Praveen Krishnan, Rama Kovvuri, Guan Pang et al.

We present a novel approach for disentangling the content of a text image from all aspects of its appearance. The appearance representation we derive can then be applied to new content, for one-shot transfer of the source style to new content. We learn this disentanglement in a self-supervised manner. Our method processes entire word boxes, without requiring segmentation of text from background, per-character processing, or making assumptions on string lengths. We show results in different text domains which were previously handled by specialized methods, e.g., scene text, handwritten text. To these ends, we make a number of technical contributions: (1) We disentangle the style and content of a textual image into a non-parametric, fixed-dimensional vector. (2) We propose a novel approach inspired by StyleGAN but conditioned over the example style at different resolution and content. (3) We present novel self-supervised training criteria which preserve both source style and target content using a pre-trained font classifier and text recognizer. Finally, (4) we also introduce Imgur5K, a new challenging dataset for handwritten word images. We offer numerous qualitative photo-realistic results of our method. We further show that our method surpasses previous work in quantitative tests on scene text and handwriting datasets, as well as in a user study.

Amanpreet Singh, Guan Pang, Mandy Toh et al.

A crucial component for the scene text based reasoning required for TextVQA and TextCaps datasets involve detecting and recognizing text present in the images using an optical character recognition (OCR) system. The current systems are crippled by the unavailability of ground truth text annotations for these datasets as well as lack of scene text detection and recognition datasets on real images disallowing the progress in the field of OCR and evaluation of scene text based reasoning in isolation from OCR systems. In this work, we propose TextOCR, an arbitrary-shaped scene text detection and recognition with 900k annotated words collected on real images from TextVQA dataset. We show that current state-of-the-art text-recognition (OCR) models fail to perform well on TextOCR and that training on TextOCR helps achieve state-of-the-art performance on multiple other OCR datasets as well. We use a TextOCR trained OCR model to create PixelM4C model which can do scene text based reasoning on an image in an end-to-end fashion, allowing us to revisit several design choices to achieve new state-of-the-art performance on TextVQA dataset.

Jing Huang, Guan Pang, Rama Kovvuri et al.

Recent advances in OCR have shown that an end-to-end (E2E) training pipeline that includes both detection and recognition leads to the best results. However, many existing methods focus primarily on Latin-alphabet languages, often even only case-insensitive English characters. In this paper, we propose an E2E approach, Multiplexed Multilingual Mask TextSpotter, that performs script identification at the word level and handles different scripts with different recognition heads, all while maintaining a unified loss that simultaneously optimizes script identification and multiple recognition heads. Experiments show that our method outperforms the single-head model with similar number of parameters in end-to-end recognition tasks, and achieves state-of-the-art results on MLT17 and MLT19 joint text detection and script identification benchmarks. We believe that our work is a step towards the end-to-end trainable and scalable multilingual multi-purpose OCR system. Our code and model will be released.

Vítor Albiero, Xingyu Chen, Xi Yin et al.

We propose real-time, six degrees of freedom (6DoF), 3D face pose estimation without face detection or landmark localization. We observe that estimating the 6DoF rigid transformation of a face is a simpler problem than facial landmark detection, often used for 3D face alignment. In addition, 6DoF offers more information than face bounding box labels. We leverage these observations to make multiple contributions: (a) We describe an easily trained, efficient, Faster R-CNN--based model which regresses 6DoF pose for all faces in the photo, without preliminary face detection. (b) We explain how pose is converted and kept consistent between the input photo and arbitrary crops created while training and evaluating our model. (c) Finally, we show how face poses can replace detection bounding box training labels. Tests on AFLW2000-3D and BIWI show that our method runs at real-time and outperforms state of the art (SotA) face pose estimators. Remarkably, our method also surpasses SotA models of comparable complexity on the WIDER FACE detection benchmark, despite not been optimized on bounding box labels.

Jigar Doshi, Saikat Basu, Guan Pang

The use of satellite imagery has become increasingly popular for disaster monitoring and response. After a disaster, it is important to prioritize rescue operations, disaster response and coordinate relief efforts. These have to be carried out in a fast and efficient manner since resources are often limited in disaster-affected areas and it's extremely important to identify the areas of maximum damage. However, most of the existing disaster mapping efforts are manual which is time-consuming and often leads to erroneous results. In order to address these issues, we propose a framework for change detection using Convolutional Neural Networks (CNN) on satellite images which can then be thresholded and clustered together into grids to find areas which have been most severely affected by a disaster. We also present a novel metric called Disaster Impact Index (DII) and use it to quantify the impact of two natural disasters - the Hurricane Harvey flood and the Santa Rosa fire. Our framework achieves a top F1 score of 81.2% on the gridded flood dataset and 83.5% on the gridded fire dataset.

Jing Huang, Viswanath Sivakumar, Mher Mnatsakanyan et al.

A significant number of images shared on social media platforms such as Facebook and Instagram contain text in various forms. It's increasingly becoming commonplace for bad actors to share misinformation, hate speech or other kinds of harmful content as text overlaid on images on such platforms. A scene-text understanding system should hence be able to handle text in various orientations that the adversary might use. Moreover, such a system can be incorporated into screen readers used to aid the visually impaired. In this work, we extend the scene-text extraction system at Facebook, Rosetta, to efficiently handle text in various orientations. Specifically, we incorporate the Rotation Region Proposal Networks (RRPN) in our text extraction pipeline and offer practical suggestions for building and deploying a model for detecting and recognizing text in arbitrary orientations efficiently. Experimental results show a significant improvement on detecting rotated text.

Ilke Demir, Krzysztof Koperski, David Lindenbaum et al.

We present the DeepGlobe 2018 Satellite Image Understanding Challenge, which includes three public competitions for segmentation, detection, and classification tasks on satellite images. Similar to other challenges in computer vision domain such as DAVIS and COCO, DeepGlobe proposes three datasets and corresponding evaluation methodologies, coherently bundled in three competitions with a dedicated workshop co-located with CVPR 2018. We observed that satellite imagery is a rich and structured source of information, yet it is less investigated than everyday images by computer vision researchers. However, bridging modern computer vision with remote sensing data analysis could have critical impact to the way we understand our environment and lead to major breakthroughs in global urban planning or climate change research. Keeping such bridging objective in mind, DeepGlobe aims to bring together researchers from different domains to raise awareness of remote sensing in the computer vision community and vice-versa. We aim to improve and evaluate state-of-the-art satellite image understanding approaches, which can hopefully serve as reference benchmarks for future research in the same topic. In this paper, we analyze characteristics of each dataset, define the evaluation criteria of the competitions, and provide baselines for each task.

Tim Danford, Onur Filiz, Jing Huang et al.

This article discusses a framework to support the design and end-to-end planning of fixed millimeter-wave networks. Compared to traditional techniques, the framework allows an organization to quickly plan a deployment in a cost-effective way. We start by using LiDAR data---basically, a 3D point cloud captured from a city---to estimate potential sites to deploy antennas and whether there is line-of-sight between them. With that data on hand, we use combinatorial optimization techniques to determine the optimal set of locations and how they should communicate with each other, to satisfy engineering (e.g., latency, polarity), design (e.g., reliability) and financial (e.g., total cost of operation) constraints. The primary goal is to connect as many people as possible to the network. Our methodology can be used for strategic planning when an organization is in the process of deciding whether to adopt a millimeter-wave technology or choosing between locations, or for operational planning when conducting a detailed design of the actual network to be deployed in a selected location.