We present Emu Video, a text-to-video generation model that factorizes the generation into two steps: first generating an image conditioned on the text, and then generating a video conditioned on the text and the generated image. We identify critical design decisions--adjusted noise schedules for diffusion, and multi-stage training--that enable us to directly generate high quality and high resolution videos, without requiring a deep cascade of models as in prior work. In human evaluations, our generated videos are strongly preferred in quality compared to all prior work--81% vs. Google's Imagen Video, 90% vs. Nvidia's PYOCO, and 96% vs. Meta's Make-A-Video. Our model outperforms commercial solutions such as RunwayML's Gen2 and Pika Labs. Finally, our factorizing approach naturally lends itself to animating images based on a user's text prompt, where our generations are preferred 96% over prior work.

We introduce Emu Video Edit (EVE), a model that establishes a new state-of-the art in video editing without relying on any supervised video editing data. To develop EVE we separately train an image editing adapter and a video generation adapter, and attach both to the same text-to-image model. Then, to align the adapters towards video editing we introduce a new unsupervised distillation procedure, Factorized Diffusion Distillation. This procedure distills knowledge from one or more teachers simultaneously, without any supervised data. We utilize this procedure to teach EVE to edit videos by jointly distilling knowledge to (i) precisely edit each individual frame from the image editing adapter, and (ii) ensure temporal consistency among the edited frames using the video generation adapter. Finally, to demonstrate the potential of our approach in unlocking other capabilities, we align additional combinations of adapters

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.

Instruction-based image editing holds immense potential for a variety of applications, as it enables users to perform any editing operation using a natural language instruction. However, current models in this domain often struggle with accurately executing user instructions. We present Emu Edit, a multi-task image editing model which sets state-of-the-art results in instruction-based image editing. To develop Emu Edit we train it to multi-task across an unprecedented range of tasks, such as region-based editing, free-form editing, and Computer Vision tasks, all of which are formulated as generative tasks. Additionally, to enhance Emu Edit's multi-task learning abilities, we provide it with learned task embeddings which guide the generation process towards the correct edit type. Both these elements are essential for Emu Edit's outstanding performance. Furthermore, we show that Emu Edit can generalize to new tasks, such as image inpainting, super-resolution, and compositions of editing tasks, with just a few labeled examples. This capability offers a significant advantage in scenarios where high-quality samples are scarce. Lastly, to facilitate a more rigorous and informed assessment of instructable image editing models, we release a new challenging and versatile benchmark that includes seven different image editing tasks. (1)

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.

Recent text-to-image diffusion models are able to generate convincing results of unprecedented quality. However, it is nearly impossible to control the shapes of different regions/objects or their layout in a fine-grained fashion. Previous attempts to provide such controls were hindered by their reliance on a fixed set of labels. To this end, we present SpaText - a new method for text-to-image generation using open-vocabulary scene control. In addition to a global text prompt that describes the entire scene, the user provides a segmentation map where each region of interest is annotated by a free-form natural language description. Due to lack of large-scale datasets that have a detailed textual description for each region in the image, we choose to leverage the current large-scale text-to-image datasets and base our approach on a novel CLIP-based spatio-textual representation, and show its effectiveness on two state-of-the-art diffusion models: pixel-based and latent-based. In addition, we show how to extend the classifier-free guidance method in diffusion models to the multi-conditional case and present an alternative accelerated inference algorithm. Finally, we offer several automatic evaluation metrics and use them, in addition to FID scores and a user study, to evaluate our method and show that it achieves state-of-the-art results on image generation with free-form textual scene control.

We propose Make-A-Video -- an approach for directly translating the tremendous recent progress in Text-to-Image (T2I) generation to Text-to-Video (T2V). Our intuition is simple: learn what the world looks like and how it is described from paired text-image data, and learn how the world moves from unsupervised video footage. Make-A-Video has three advantages: (1) it accelerates training of the T2V model (it does not need to learn visual and multimodal representations from scratch), (2) it does not require paired text-video data, and (3) the generated videos inherit the vastness (diversity in aesthetic, fantastical depictions, etc.) of today's image generation models. We design a simple yet effective way to build on T2I models with novel and effective spatial-temporal modules. First, we decompose the full temporal U-Net and attention tensors and approximate them in space and time. Second, we design a spatial temporal pipeline to generate high resolution and frame rate videos with a video decoder, interpolation model and two super resolution models that can enable various applications besides T2V. In all aspects, spatial and temporal resolution, faithfulness to text, and quality, Make-A-Video sets the new state-of-the-art in text-to-video generation, as determined by both qualitative and quantitative measures.

We tackle the problem of generating audio samples conditioned on descriptive text captions. In this work, we propose AaudioGen, an auto-regressive generative model that generates audio samples conditioned on text inputs. AudioGen operates on a learnt discrete audio representation. The task of text-to-audio generation poses multiple challenges. Due to the way audio travels through a medium, differentiating ``objects'' can be a difficult task (e.g., separating multiple people simultaneously speaking). This is further complicated by real-world recording conditions (e.g., background noise, reverberation, etc.). Scarce text annotations impose another constraint, limiting the ability to scale models. Finally, modeling high-fidelity audio requires encoding audio at high sampling rate, leading to extremely long sequences. To alleviate the aforementioned challenges we propose an augmentation technique that mixes different audio samples, driving the model to internally learn to separate multiple sources. We curated 10 datasets containing different types of audio and text annotations to handle the scarcity of text-audio data points. For faster inference, we explore the use of multi-stream modeling, allowing the use of shorter sequences while maintaining a similar bitrate and perceptual quality. We apply classifier-free guidance to improve adherence to text. Comparing to the evaluated baselines, AudioGen outperforms over both objective and subjective metrics. Finally, we explore the ability of the proposed method to generate audio continuation conditionally and unconditionally. Samples: https://felixkreuk.github.io/audiogen

Text-guided human motion generation has drawn significant interest because of its impactful applications spanning animation and robotics. Recently, application of diffusion models for motion generation has enabled improvements in the quality of generated motions. However, existing approaches are limited by their reliance on relatively small-scale motion capture data, leading to poor performance on more diverse, in-the-wild prompts. In this paper, we introduce Make-An-Animation, a text-conditioned human motion generation model which learns more diverse poses and prompts from large-scale image-text datasets, enabling significant improvement in performance over prior works. Make-An-Animation is trained in two stages. First, we train on a curated large-scale dataset of (text, static pseudo-pose) pairs extracted from image-text datasets. Second, we fine-tune on motion capture data, adding additional layers to model the temporal dimension. Unlike prior diffusion models for motion generation, Make-An-Animation uses a U-Net architecture similar to recent text-to-video generation models. Human evaluation of motion realism and alignment with input text shows that our model reaches state-of-the-art performance on text-to-motion generation.

We present MAV3D (Make-A-Video3D), a method for generating three-dimensional dynamic scenes from text descriptions. Our approach uses a 4D dynamic Neural Radiance Field (NeRF), which is optimized for scene appearance, density, and motion consistency by querying a Text-to-Video (T2V) diffusion-based model. The dynamic video output generated from the provided text can be viewed from any camera location and angle, and can be composited into any 3D environment. MAV3D does not require any 3D or 4D data and the T2V model is trained only on Text-Image pairs and unlabeled videos. We demonstrate the effectiveness of our approach using comprehensive quantitative and qualitative experiments and show an improvement over previously established internal baselines. To the best of our knowledge, our method is the first to generate 3D dynamic scenes given a text description.

Egocentric augmented reality devices such as wearable glasses passively capture visual data as a human wearer tours a home environment. We envision a scenario wherein the human communicates with an AI agent powering such a device by asking questions (e.g., "where did you last see my keys?"). In order to succeed at this task, the egocentric AI assistant must (1) construct semantically rich and efficient scene memories that encode spatio-temporal information about objects seen during the tour and (2) possess the ability to understand the question and ground its answer into the semantic memory representation. Towards that end, we introduce (1) a new task - Episodic Memory Question Answering (EMQA) wherein an egocentric AI assistant is provided with a video sequence (the tour) and a question as an input and is asked to localize its answer to the question within the tour, (2) a dataset of grounded questions designed to probe the agent's spatio-temporal understanding of the tour, and (3) a model for the task that encodes the scene as an allocentric, top-down semantic feature map and grounds the question into the map to localize the answer. We show that our choice of episodic scene memory outperforms naive, off-the-shelf solutions for the task as well as a host of very competitive baselines and is robust to noise in depth, pose as well as camera jitter.

Current pre-trained language models have enabled remarkable improvements in downstream tasks, but it remains difficult to distinguish effects of statistical correlation from more systematic logical reasoning grounded on understanding of the real world. In this paper we tease these factors apart by leveraging counterfactual conditionals , which force language models to predict unusual consequences based on hypothetical propositions. We introduce a set of tests drawn from psycholinguistic experiments, as well as larger-scale controlled datasets, to probe counterfactual predictions from a variety of popular pre-trained language models. We find that models are consistently able to override real-world knowledge in counterfactual scenarios, and that this effect is more robust in case of stronger baseline world knowledge—however, we also find that for most models this effect appears largely to be driven by simple lexical cues. When we mitigate effects of both world knowledge and lexical cues to test knowledge of linguistic nuances of counterfactuals, we find that only GPT-3 shows sensitivity to these nuances, though this sensitivity is also non-trivially impacted by lexical associative factors.

Large-scale pre-trained vision-and-language 001 (V+L) transformers have propelled the state 002 of the art (SOTA) on Visual Question Answer-003 ing (VQA) task. Despite impressive perfor-004 mance on the standard VQA benchmark, it re-005 mains unclear how robust these models are. To 006 investigate, we conduct a host of evaluations 007 over 4 different types of robust VQA datasets: 008 ( i ) Linguistic Variation; ( ii ) Logical Reason-009 ing; ( iii ) Visual Content Manipulation; and 010 ( iv ) Answer Distribution Shift. Experiments 011 show that pre-trained V+L models already ex-012 hibit better robustness than many task-specific 013 SOTA methods via standard model finetun-014 ing. To further enhance model robustness, we 015 propose M ANGO , a generic and efficient ap-016 proach that learns a M ultimodal A dversarial 017 N oise G enerat O r in the embedding space to 018 fool V+L models. Differing from previous 019 studies focused on one specific type of robust-020 ness, M ANGO is agnostic to robustness types, 021 and enables universal performance lift for both 022 task-specific and pre-trained models over di-023 verse robust VQA datasets designed to evaluate 024 broad aspects of robustness. Comprehensive 025 experiments demonstrate that M ANGO outper-026 forms previous task-specific SOTAs on 7 out 027 of 9 robustness benchmarks. 028

Interactive robots navigating photo-realistic environments need to be trained to effectively leverage and handle the dynamic nature of dialogue in addition to the challenges underlying vision-and-language navigation (VLN). In this paper, we present VISITRON, a multi-modal Transformer-based navigator better suited to the interactive regime inherent to Cooperative Vision-and-Dialog Navigation (CVDN). VISITRON is trained to: i) identify and associate object-level concepts and semantics between the environment and dialogue history, ii) identify when to interact vs. navigate via imitation learning of a binary classification head. We perform extensive pre-training and fine-tuning ablations with VISITRON to gain empirical insights and improve performance on CVDN. VISITRON’s ability to identify when to interact leads to a natural generalization of the game-play mode introduced by Roman et al. (2020) for enabling the use of such models in different environments. VISITRON is competitive with models on the static CVDN leaderboard and attains state-of-the-art performance on the Success weighted by Path Length (SPL) metric.

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 .

Recent text-to-image generation methods provide a simple yet exciting conversion capability between text and image domains. While these methods have incrementally improved the generated image fidelity and text relevancy, several pivotal gaps remain unanswered, limiting applicability and quality. We propose a novel text-to-image method that addresses these gaps by (i) enabling a simple control mechanism complementary to text in the form of a scene, (ii) introducing elements that substantially improve the tokenization process by employing domain-specific knowledge over key image regions (faces and salient objects), and (iii) adapting classifier-free guidance for the transformer use case. Our model achieves state-of-the-art FID and human evaluation results, unlocking the ability to generate high fidelity images in a resolution of 512x512 pixels, significantly improving visual quality. Through scene controllability, we introduce several new capabilities: (i) Scene editing, (ii) text editing with anchor scenes, (iii) overcoming out-of-distribution text prompts, and (iv) story illustration generation, as demonstrated in the story we wrote.

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. 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 (3.2 s 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/ .

We investigate the problem of generating utter-001 ances from pairs of images showing a before 002 and an after state of a change in a visual scene. 003 We present a transformer model with differ-004 ence attention heads that learns to attend to vi-005 sual changes in consecutive images via a differ-006 ence key. We test our approach in instruction 007 generation, change captioning and difference 008 spotting and compare these tasks in terms of 009 their linguistic phenomena and reasoning abil-010 ities. Our model outperforms the state-of-the-011 art for instruction generation on the BLOCKS 012 and difference spotting on the Spot-the-diff 013 dataset and generates accurate referential and 014 compositional spatial expressions. Finally, we 015 identify linguistic phenomena that pose chal-016 lenges for generation in changing scenes. 017

Abstract This letter is a retrospective analysis of our team's research for the Defense Advanced Research Projects Agency Explainable Artificial Intelligence project. Our initial approach was to use salience maps, English sentences, and lists of feature names to explain the behavior of deep‐learning‐based discriminative systems, with particular focus on visual question answering systems. We found that presenting static explanations along with answers led to limited positive effects. By exploring various combinations of machine and human explanation production and consumption, we evolved a notion of explanation as an interactive process that takes place usually between humans and artificial intelligence systems but sometimes within the software system. We realized that by interacting via explanations people could task and adapt machine learning (ML) agents. We added affordances for editing explanations and modified the ML system to act in accordance with the edits to produce an interpretable interface to the agent. Through this interface, editing an explanation can adapt a system's performance to new, modified purposes. This deep tasking, wherein the agent knows its objective and the explanation for that objective, will be critical to enable higher levels of autonomy.

Recent Visual Question Answering (VQA) models have shown impressive performance on the VQA benchmark but remain sensitive to small linguistic variations in input questions. Existing approaches address this by augmenting the dataset with question paraphrases from visual question generation models or adversarial perturbations. These approaches use the combined data to learn an answer classifier by minimizing the standard cross-entropy loss. To more effectively leverage augmented data, we build on the recent success in contrastive learning. We propose a novel training paradigm (ConClaT) that optimizes both cross-entropy and contrastive losses. The contrastive loss encourages representations to be robust to linguistic variations in questions while the cross-entropy loss preserves the discriminative power of representations for answer prediction.We find that optimizing both losses – either alternately or jointly – is key to effective training. On the VQA-Rephrasings [44] benchmark, which measures the VQA model’s answer consistency across human paraphrases of a question, ConClaT improves Consensus Score by 1.63% over an improved baseline. In addition, on the standard VQA 2.0 benchmark, we improve the VQA accuracy by 0.78% overall. We also show that ConClaT is agnostic to the type of data-augmentation strategy used.