Diffusion models are successful for synthesizing high-quality videos but are limited to generating short clips (e.g., 2-10 seconds). Synthesizing sustained footage (e.g. over minutes) still remains an open research question. In this paper, we propose MALT Diffusion (using Memory-Augmented Latent Transformers), a new diffusion model specialized for long video generation. MALT Diffusion (or just MALT) handles long videos by subdividing them into short segments and doing segment-level autoregressive generation. To achieve this, we first propose recurrent attention layers that encode multiple segments into a compact memory latent vector; by maintaining this memory vector over time, MALT is able to condition on it and continuously generate new footage based on a long temporal context. We also present several training techniques that enable the model to generate frames over a long horizon with consistent quality and minimal degradation. We validate the effectiveness of MALT through experiments on long video benchmarks. We first perform extensive analysis of MALT in long-contextual understanding capability and stability using popular long video benchmarks. For example, MALT achieves an FVD score of 220.4 on 128-frame video generation on UCF-101, outperforming the previous state-of-the-art of 648.4. Finally, we explore MALT's capabilities in a text-to-video generation setting and show that it can produce long videos compared with recent techniques for long text-to-video generation.

Recent advancements in large language model(LLM) performance on medical multiple choice question (MCQ) benchmarks have stimulated interest from healthcare providers and patients globally. Particularly in low-and middle-income countries (LMICs) facing acute physician shortages and lack of specialists, LLMs offer a potentially scalable pathway to enhance healthcare access and reduce costs. However, their effectiveness in the Global South, especially across the African continent, remains to be established. In this work, we introduce AfriMed-QA, the first large scale Pan-African English multi-specialty medical Question-Answering (QA) dataset, 15,000 questions (open and closed-ended) sourced from over 60 medical schools across 16 countries, covering 32 medical specialties. We further evaluate 30 LLMs across multiple axes including correctness and demographic bias. Our findings show significant performance variation across specialties and geographies, MCQ performance clearly lags USMLE (MedQA). We find that biomedical LLMs underperform general models and smaller edge-friendly LLMs struggle to achieve a passing score. Interestingly, human evaluations show a consistent consumer preference for LLM answers and explanations when compared with clinician answers.

Foundational models are able to generate text outputs given prompt instructions and text, audio, or image inputs. Recently these models have been combined to perform tasks on video, such as video summarization. Such video foundation models perform pre-training by aligning outputs from each modality-specific model into the same embedding space. Then the embeddings from each model are used within a language model, which is fine-tuned on a desired instruction set. Aligning each modality during pre-training is computationally expensive and prevents rapid testing of different base modality models. During fine-tuning, evaluation is carried out within in-domain videos where it is hard to understand the generalizability and data efficiency of these methods. To alleviate these issues we propose a plug-and-play video language model. It directly uses the texts generated from each input modality into the language model, avoiding pre-training alignment overhead. Instead of fine-tuning we leverage few-shot instruction adaptation strategies. We compare the performance versus the computational costs for our plug-and-play style method and baseline tuning methods. Finally, we explore the generalizability of each method during domain shift and present insights on what data is useful when training data is limited. Through this analysis, we present practical insights on how to leverage multi-modal foundational models for effective results given realistic compute and data limitations.

We extend multimodal transformers to include 3D camera motion as a conditioning signal for the task of video generation. Generative video models are becoming increasingly powerful, thus focusing research efforts on methods of controlling the output of such models. We propose to add virtual 3D camera controls to generative video methods by conditioning generated video on an encoding of three-dimensional camera movement over the course of the generated video. Results demonstrate that we are (1) able to successfully control the camera during video generation, starting from a single frame and a camera signal, and (2) we demonstrate the accuracy of the generated 3D camera paths using traditional computer vision methods.

Although the task of anticipating future actions is highly uncertain, information from additional modalities help to narrow down plausible action choices. Each modality provides different environmental context for the model to learn from. While previous multi-modal methods leverage information from modalities such as video and audio, we primarily explore how text inputs for actions and objects can also enable more accurate action anticipation. Therefore, we propose a Multi-modal Anticipative Transformer (MAT), an attention-based video transformer architecture that jointly learns from multi-modal features and text captions. We train our model in two-stages, where the model first learns to predict actions in the video clip by aligning with captions, and during the second stage, we fine-tune the model to predict future actions. Compared to existing methods, MAT has the advantage of learning additional environmental context from two kinds of text inputs: action descriptions during the pre-training stage, and the text inputs for detected objects and actions during modality feature fusion. Through extensive experiments, we evaluate the effectiveness of the pre-training stage, and show that our model outperforms previous methods on all datasets. In addition, we examine the impact of object and action information obtained via text and perform extensive ablations. We evaluate the performance on on three datasets: EpicKitchens-100, EpicKitchens-55 and EGTEA GAZE+; and show that text descriptions do indeed aid in more effective action anticipation.

We present VideoPoet, a language model capable of synthesizing high-quality video, with matching audio, from a large variety of conditioning signals. VideoPoet employs a decoder-only transformer architecture that processes multimodal inputs -- including images, videos, text, and audio. The training protocol follows that of Large Language Models (LLMs), consisting of two stages: pretraining and task-specific adaptation. During pretraining, VideoPoet incorporates a mixture of multimodal generative objectives within an autoregressive Transformer framework. The pretrained LLM serves as a foundation that can be adapted for a range of video generation tasks. We present empirical results demonstrating the model's state-of-the-art capabilities in zero-shot video generation, specifically highlighting VideoPoet's ability to generate high-fidelity motions. Project page: http://sites.research.google/videopoet/

We present W.A.L.T, a transformer-based approach for photorealistic video generation via diffusion modeling. Our approach has two key design decisions. First, we use a causal encoder to jointly compress images and videos within a unified latent space, enabling training and generation across modalities. Second, for memory and training efficiency, we use a window attention architecture tailored for joint spatial and spatiotemporal generative modeling. Taken together these design decisions enable us to achieve state-of-the-art performance on established video (UCF-101 and Kinetics-600) and image (ImageNet) generation benchmarks without using classifier free guidance. Finally, we also train a cascade of three models for the task of text-to-video generation consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of $512 \times 896$ resolution at $8$ frames per second.

In this work, we introduce Semantic Pyramid AutoEncoder (SPAE) for enabling frozen LLMs to perform both understanding and generation tasks involving non-linguistic modalities such as images or videos. SPAE converts between raw pixels and interpretable lexical tokens (or words) extracted from the LLM's vocabulary. The resulting tokens capture both the semantic meaning and the fine-grained details needed for visual reconstruction, effectively translating the visual content into a language comprehensible to the LLM, and empowering it to perform a wide array of multimodal tasks. Our approach is validated through in-context learning experiments with frozen PaLM 2 and GPT 3.5 on a diverse set of image understanding and generation tasks. Our method marks the first successful attempt to enable a frozen LLM to generate image content while surpassing state-of-the-art performance in image understanding tasks, under the same setting, by over 25%.

Domain randomization (DR), which entails training a policy with randomized dynamics, has proven to be a simple yet effective algorithm for reducing the gap between simulation and the real world. However, DR often requires careful tuning of randomization parameters. Methods like Bayesian Domain Randomization (Bayesian DR) and Active Domain Randomization (Adaptive DR) address this issue by automating parameter range selection using real-world experience. While effective, these algorithms often require long computation time, as a new policy is trained from scratch every iteration. In this work, we propose Adaptive Bayesian Domain Randomization via Strategic Fine-tuning (BayRnTune), which inherits the spirit of BayRn but aims to significantly accelerate the learning processes by fine-tuning from previously learned policy. This idea leads to a critical question: which previous policy should we use as a prior during fine-tuning? We investigated four different fine-tuning strategies and compared them against baseline algorithms in five simulated environments, ranging from simple benchmark tasks to more complex legged robot environments. Our analysis demonstrates that our method yields better rewards in the same amount of timesteps compared to vanilla domain randomization or Bayesian DR.

Humanoid robots are well suited for human habitats due to their morphological similarity, but developing controllers for them is a challenging task that involves multiple sub-problems, such as control, planning and perception. In this paper, we introduce a method to simplify controller design by enabling users to train and fine-tune robot control policies using natural language commands. We first learn a neural network policy that generates behaviors given a natural language command, such as "walk forward", by combining Large Language Models (LLMs), motion retargeting, and motion imitation. Based on the synthesized motion, we iteratively fine-tune by updating the text prompt and querying LLMs to find the best checkpoint associated with the closest motion in history. We validate our approach using a simulated Digit humanoid robot and demonstrate learning of diverse motions, such as walking, hopping, and kicking, without the burden of complex reward engineering. In addition, we show that our iterative refinement enables us to learn 3x times faster than a naive formulation that learns from scratch.