This work introduces Video Diffusion Transformer (VDT), which pioneers the use of transformers in diffusion-based video generation. It features transformer blocks with modularized temporal and spatial attention modules to leverage the rich spatial-temporal representation inherited in transformers. We also propose a unified spatial-temporal mask modeling mechanism, seamlessly integrated with the model, to cater to diverse video generation scenarios. VDT offers several appealing benefits. 1) It excels at capturing temporal dependencies to produce temporally consistent video frames and even simulate the physics and dynamics of 3D objects over time. 2) It facilitates flexible conditioning information, \eg, simple concatenation in the token space, effectively unifying different token lengths and modalities. 3) Pairing with our proposed spatial-temporal mask modeling mechanism, it becomes a general-purpose video diffuser for harnessing a range of tasks, including unconditional generation, video prediction, interpolation, animation, and completion, etc. Extensive experiments on these tasks spanning various scenarios, including autonomous driving, natural weather, human action, and physics-based simulation, demonstrate the effectiveness of VDT. Additionally, we present comprehensive studies on how \model handles conditioning information with the mask modeling mechanism, which we believe will benefit future research and advance the field. Project page: https:VDT-2023.github.io
In this paper, we investigate the challenge of spatio-temporal video prediction task, which involves generating future video frames based on historical spatio-temporal observation streams. Existing approaches typically utilize external information such as semantic maps to improve video prediction accuracy, which often neglect the inherent physical knowledge embedded within videos. Worse still, their high computational costs could impede their applications for high-resolution videos. To address these constraints, we introduce a novel framework called \underline{P}hysics-\underline{a}ssisted \underline{S}patio-\underline{t}emporal \underline{Net}work (PastNet) for high-quality video prediction. The core of PastNet lies in incorporating a spectral convolution operator in the Fourier domain, which efficiently introduces inductive biases from the underlying physical laws. Additionally, we employ a memory bank with the estimated intrinsic dimensionality to discretize local features during the processing of complex spatio-temporal signals, thereby reducing computational costs and facilitating efficient high-resolution video prediction. Extensive experiments on various widely-used spatio-temporal video benchmarks demonstrate the effectiveness and efficiency of the proposed PastNet compared with a range of state-of-the-art methods, particularly in high-resolution scenarios.
Video is a promising source of knowledge for embodied agents to learn models of the world's dynamics. Large deep networks have become increasingly effective at modeling complex video data in a self-supervised manner, as evaluated by metrics based on human perceptual similarity or pixel-wise comparison. However, it remains unclear whether current metrics are accurate indicators of performance on downstream tasks. We find empirically that for planning robotic manipulation, existing metrics can be unreliable at predicting execution success. To address this, we propose a benchmark for action-conditioned video prediction in the form of a control benchmark that evaluates a given model for simulated robotic manipulation through sampling-based planning. Our benchmark, Video Prediction for Visual Planning ($VP^2$), includes simulated environments with 11 task categories and 310 task instance definitions, a full planning implementation, and training datasets containing scripted interaction trajectories for each task category. A central design goal of our benchmark is to expose a simple interface -- a single forward prediction call -- so it is straightforward to evaluate almost any action-conditioned video prediction model. We then leverage our benchmark to study the effects of scaling model size, quantity of training data, and model ensembling by analyzing five highly-performant video prediction models, finding that while scale can improve perceptual quality when modeling visually diverse settings, other attributes such as uncertainty awareness can also aid planning performance.
In this paper, we explore the impact of adding tactile sensation to video prediction models for physical robot interactions. Predicting the impact of robotic actions on the environment is a fundamental challenge in robotics. Current methods leverage visual and robot action data to generate video predictions over a given time period, which can then be used to adjust robot actions. However, humans rely on both visual and tactile feedback to develop and maintain a mental model of their physical surroundings. In this paper, we investigate the impact of integrating tactile feedback into video prediction models for physical robot interactions. We propose three multi-modal integration approaches and compare the performance of these tactile-enhanced video prediction models. Additionally, we introduce two new datasets of robot pushing that use a magnetic-based tactile sensor for unsupervised learning. The first dataset contains visually identical objects with different physical properties, while the second dataset mimics existing robot-pushing datasets of household object clusters. Our results demonstrate that incorporating tactile feedback into video prediction models improves scene prediction accuracy and enhances the agent's perception of physical interactions and understanding of cause-effect relationships during physical robot interactions.
The drastic variation of motion in spatial and temporal dimensions makes the video prediction task extremely challenging. Existing RNN models obtain higher performance by deepening or widening the model. They obtain the multi-scale features of the video only by stacking layers, which is inefficient and brings unbearable training costs (such as memory, FLOPs, and training time). Different from them, this paper proposes a spatiotemporal multi-scale model called MS-LSTM wholly from a multi-scale perspective. On the basis of stacked layers, MS-LSTM incorporates two additional efficient multi-scale designs to fully capture spatiotemporal context information. Concretely, we employ LSTMs with mirrored pyramid structures to construct spatial multi-scale representations and LSTMs with different convolution kernels to construct temporal multi-scale representations. We theoretically analyze the training cost and performance of MS-LSTM and its components. Detailed comparison experiments with twelve baseline models on four video datasets show that MS-LSTM has better performance but lower training costs.
The innovative application of precise geospatial vegetation forecasting holds immense potential across diverse sectors, including agriculture, forestry, humanitarian aid, and carbon accounting. To leverage the vast availability of satellite imagery for this task, various works have applied deep neural networks for predicting multispectral images in photorealistic quality. However, the important area of vegetation dynamics has not been thoroughly explored. Our study breaks new ground by introducing GreenEarthNet, the first dataset specifically designed for high-resolution vegetation forecasting, and Contextformer, a novel deep learning approach for predicting vegetation greenness from Sentinel 2 satellite images with fine resolution across Europe. Our multi-modal transformer model Contextformer leverages spatial context through a vision backbone and predicts the temporal dynamics on local context patches incorporating meteorological time series in a parameter-efficient manner. The GreenEarthNet dataset features a learned cloud mask and an appropriate evaluation scheme for vegetation modeling. It also maintains compatibility with the existing satellite imagery forecasting dataset EarthNet2021, enabling cross-dataset model comparisons. Our extensive qualitative and quantitative analyses reveal that our methods outperform a broad range of baseline techniques. This includes surpassing previous state-of-the-art models on EarthNet2021, as well as adapted models from time series forecasting and video prediction. To the best of our knowledge, this work presents the first models for continental-scale vegetation modeling at fine resolution able to capture anomalies beyond the seasonal cycle, thereby paving the way for predicting vegetation health and behaviour in response to climate variability and extremes.
Imagining the future trajectory is the key for robots to make sound planning and successfully reach their goals. Therefore, text-conditioned video prediction (TVP) is an essential task to facilitate general robot policy learning. To tackle this task and empower robots with the ability to foresee the future, we propose a sample and computation-efficient model, named \textbf{Seer}, by inflating the pretrained text-to-image (T2I) stable diffusion models along the temporal axis. We enhance the U-Net and language conditioning model by incorporating computation-efficient spatial-temporal attention. Furthermore, we introduce a novel Frame Sequential Text Decomposer module that dissects a sentence's global instruction into temporally aligned sub-instructions, ensuring precise integration into each frame of generation. Our framework allows us to effectively leverage the extensive prior knowledge embedded in pretrained T2I models across the frames. With the adaptable-designed architecture, Seer makes it possible to generate high-fidelity, coherent, and instruction-aligned video frames by fine-tuning a few layers on a small amount of data. The experimental results on Something Something V2 (SSv2), Bridgedata and EpicKitchens-100 datasets demonstrate our superior video prediction performance with around 480-GPU hours versus CogVideo with over 12,480-GPU hours: achieving the 31% FVD improvement compared to the current SOTA model on SSv2 and 83.7% average preference in the human evaluation.
The performance of video prediction has been greatly boosted by advanced deep neural networks. However, most of the current methods suffer from large model sizes and require extra inputs, e.g., semantic/depth maps, for promising performance. For efficiency consideration, in this paper, we propose a Dynamic Multi-scale Voxel Flow Network (DMVFN) to achieve better video prediction performance at lower computational costs with only RGB images, than previous methods. The core of our DMVFN is a differentiable routing module that can effectively perceive the motion scales of video frames. Once trained, our DMVFN selects adaptive sub-networks for different inputs at the inference stage. Experiments on several benchmarks demonstrate that our DMVFN is an order of magnitude faster than Deep Voxel Flow and surpasses the state-of-the-art iterative-based OPT on generated image quality. Our code and demo are available at https://huxiaotaostasy.github.io/DMVFN/.
Video prediction is a complex time-series forecasting task with great potential in many use cases. However, conventional methods overemphasize accuracy while ignoring the slow prediction speed caused by complicated model structures that learn too much redundant information with excessive GPU memory consumption. Furthermore, conventional methods mostly predict frames sequentially (frame-by-frame) and thus are hard to accelerate. Consequently, valuable use cases such as real-time danger prediction and warning cannot achieve fast enough inference speed to be applicable in reality. Therefore, we propose a transformer-based keypoint prediction neural network (TKN), an unsupervised learning method that boost the prediction process via constrained information extraction and parallel prediction scheme. TKN is the first real-time video prediction solution to our best knowledge, while significantly reducing computation costs and maintaining other performance. Extensive experiments on KTH and Human3.6 datasets demonstrate that TKN predicts 11 times faster than existing methods while reducing memory consumption by 17.4% and achieving state-of-the-art prediction performance on average.
Future frame prediction has been approached through two primary methods: autoregressive and non-autoregressive. Autoregressive methods rely on the Markov assumption and can achieve high accuracy in the early stages of prediction when errors are not yet accumulated. However, their performance tends to decline as the number of time steps increases. In contrast, non-autoregressive methods can achieve relatively high performance but lack correlation between predictions for each time step. In this paper, we propose an Implicit Stacked Autoregressive Model for Video Prediction (IAM4VP), which is an implicit video prediction model that applies a stacked autoregressive method. Like non-autoregressive methods, stacked autoregressive methods use the same observed frame to estimate all future frames. However, they use their own predictions as input, similar to autoregressive methods. As the number of time steps increases, predictions are sequentially stacked in the queue. To evaluate the effectiveness of IAM4VP, we conducted experiments on three common future frame prediction benchmark datasets and weather\&climate prediction benchmark datasets. The results demonstrate that our proposed model achieves state-of-the-art performance.
Learning physical dynamics in a series of non-stationary environments is a challenging but essential task for model-based reinforcement learning (MBRL) with visual inputs. It requires the agent to consistently adapt to novel tasks without forgetting previous knowledge. In this paper, we present a new continual learning approach for visual dynamics modeling and explore its efficacy in visual control and forecasting. The key assumption is that an ideal world model can provide a non-forgetting environment simulator, which enables the agent to optimize the policy in a multi-task learning manner based on the imagined trajectories from the world model. To this end, we first propose the mixture world model that learns task-specific dynamics priors with a mixture of Gaussians, and then introduce a new training strategy to overcome catastrophic forgetting, which we call predictive experience replay. Finally, we extend these methods to continual RL and further address the value estimation problems with the exploratory-conservative behavior learning approach. Our model remarkably outperforms the naive combinations of existing continual learning and visual RL algorithms on DeepMind Control and Meta-World benchmarks with continual visual control tasks. It is also shown to effectively alleviate the forgetting of spatiotemporal dynamics in video prediction datasets with evolving domains.
Motion, scene and object are three primary visual components of a video. In particular, objects represent the foreground, scenes represent the background, and motion traces their dynamics. Based on this insight, we propose a two-stage MOtion, Scene and Object decomposition framework (MOSO) for video prediction, consisting of MOSO-VQVAE and MOSO-Transformer. In the first stage, MOSO-VQVAE decomposes a previous video clip into the motion, scene and object components, and represents them as distinct groups of discrete tokens. Then, in the second stage, MOSO-Transformer predicts the object and scene tokens of the subsequent video clip based on the previous tokens and adds dynamic motion at the token level to the generated object and scene tokens. Our framework can be easily extended to unconditional video generation and video frame interpolation tasks. Experimental results demonstrate that our method achieves new state-of-the-art performance on five challenging benchmarks for video prediction and unconditional video generation: BAIR, RoboNet, KTH, KITTI and UCF101. In addition, MOSO can produce realistic videos by combining objects and scenes from different videos.
We propose a novel framework for the task of object-centric video prediction, i.e., extracting the compositional structure of a video sequence, as well as modeling objects dynamics and interactions from visual observations in order to predict the future object states, from which we can then generate subsequent video frames. With the goal of learning meaningful spatio-temporal object representations and accurately forecasting object states, we propose two novel object-centric video predictor (OCVP) transformer modules, which decouple the processing of temporal dynamics and object interactions, thus presenting an improved prediction performance. In our experiments, we show how our object-centric prediction framework utilizing our OCVP predictors outperforms object-agnostic video prediction models on two different datasets, while maintaining consistent and accurate object representations.
Prediction of dynamic environment is crucial to safe navigation of an autonomous vehicle. Urban traffic scenes are particularly challenging to forecast due to complex interactions between various dynamic agents, such as vehicles and vulnerable road users. Previous approaches have used egocentric occupancy grid maps to represent and predict dynamic environments. However, these predictions suffer from blurriness, loss of scene structure at turns, and vanishing of agents over longer prediction horizon. In this work, we propose a novel framework to make long-term predictions by representing the traffic scene in a fixed frame, referred as allo-centric occupancy grid. This allows for the static scene to remain fixed and to represent motion of the ego-vehicle on the grid like other agents'. We study the allo-centric grid prediction with different video prediction networks and validate the approach on the real-world Nuscenes dataset. The results demonstrate that the allo-centric grid representation significantly improves scene prediction, in comparison to the conventional ego-centric grid approach.
The task of video prediction and generation is known to be notoriously difficult, with the research in this area largely limited to short-term predictions. Though plagued with noise and stochasticity, videos consist of features that are organised in a spatiotemporal hierarchy, different features possessing different temporal dynamics. In this paper, we introduce Dynamic Latent Hierarchy (DLH) -- a deep hierarchical latent model that represents videos as a hierarchy of latent states that evolve over separate and fluid timescales. Each latent state is a mixture distribution with two components, representing the immediate past and the predicted future, causing the model to learn transitions only between sufficiently dissimilar states, while clustering temporally persistent states closer together. Using this unique property, DLH naturally discovers the spatiotemporal structure of a dataset and learns disentangled representations across its hierarchy. We hypothesise that this simplifies the task of modeling temporal dynamics of a video, improves the learning of long-term dependencies, and reduces error accumulation. As evidence, we demonstrate that DLH outperforms state-of-the-art benchmarks in video prediction, is able to better represent stochasticity, as well as to dynamically adjust its hierarchical and temporal structure. Our paper shows, among other things, how progress in representation learning can translate into progress in prediction tasks.
Natural data is redundant yet predominant architectures tile computation uniformly across their input and output space. We propose the Recurrent Interface Networks (RINs), an attention-based architecture that decouples its core computation from the dimensionality of the data, enabling adaptive computation for more scalable generation of high-dimensional data. RINs focus the bulk of computation (i.e. global self-attention) on a set of latent tokens, using cross-attention to read and write (i.e. route) information between latent and data tokens. Stacking RIN blocks allows bottom-up (data to latent) and top-down (latent to data) feedback, leading to deeper and more expressive routing. While this routing introduces challenges, this is less problematic in recurrent computation settings where the task (and routing problem) changes gradually, such as iterative generation with diffusion models. We show how to leverage recurrence by conditioning the latent tokens at each forward pass of the reverse diffusion process with those from prior computation, i.e. latent self-conditioning. RINs yield state-of-the-art pixel diffusion models for image and video generation, scaling to 1024X1024 images without cascades or guidance, while being domain-agnostic and up to 10X more efficient than 2D and 3D U-Nets.
We present a multi-scale predictive coding model for future video frames prediction. Drawing inspiration on the ``Predictive Coding" theories in cognitive science, it is updated by a combination of bottom-up and top-down information flows, which can enhance the interaction between different network levels. However, traditional predictive coding models only predict what is happening hierarchically rather than predicting the future. To address the problem, our model employs a multi-scale approach (Coarse to Fine), where the higher level neurons generate coarser predictions (lower resolution), while the lower level generate finer predictions (higher resolution). In terms of network architecture, we directly incorporate the encoder-decoder network within the LSTM module and share the final encoded high-level semantic information across different network levels. This enables comprehensive interaction between the current input and the historical states of LSTM compared with the traditional Encoder-LSTM-Decoder architecture, thus learning more believable temporal and spatial dependencies. Furthermore, to tackle the instability in adversarial training and mitigate the accumulation of prediction errors in long-term prediction, we propose several improvements to the training strategy. Our approach achieves good performance on datasets such as KTH, Moving MNIST and Caltech Pedestrian. Code is available at https://github.com/Ling-CF/MSPN.
We introduce the MAsked Generative VIdeo Transformer, MAGVIT, to tackle various video synthesis tasks with a single model. We introduce a 3D tokenizer to quantize a video into spatial-temporal visual tokens and propose an embedding method for masked video token modeling to facilitate multi-task learning. We conduct extensive experiments to demonstrate the quality, efficiency, and flexibility of MAGVIT. Our experiments show that (i) MAGVIT performs favorably against state-of-the-art approaches and establishes the best-published FVD on three video generation benchmarks, including the challenging Kinetics-600. (ii) MAGVIT outperforms existing methods in inference time by two orders of magnitude against diffusion models and by 60x against autoregressive models. (iii) A single MAGVIT model supports ten diverse generation tasks and generalizes across videos from different visual domains. The source code and trained models will be released to the public at https://magvit.cs.cmu.edu.
Video prediction is a challenging computer vision task that has a wide range of applications. In this work, we present a new family of Transformer-based models for video prediction. Firstly, an efficient local spatial-temporal separation attention mechanism is proposed to reduce the complexity of standard Transformers. Then, a full autoregressive model, a partial autoregressive model and a non-autoregressive model are developed based on the new efficient Transformer. The partial autoregressive model has a similar performance with the full autoregressive model but a faster inference speed. The non-autoregressive model not only achieves a faster inference speed but also mitigates the quality degradation problem of the autoregressive counterparts, but it requires additional parameters and loss function for learning. Given the same attention mechanism, we conducted a comprehensive study to compare the proposed three video prediction variants. Experiments show that the proposed video prediction models are competitive with more complex state-of-the-art convolutional-LSTM based models. The source code is available at https://github.com/XiYe20/VPTR.
The existing state-of-the-art method for audio-visual conditioned video prediction uses the latent codes of the audio-visual frames from a multimodal stochastic network and a frame encoder to predict the next visual frame. However, a direct inference of per-pixel intensity for the next visual frame is extremely challenging because of the high-dimensional image space. To this end, we decouple the audio-visual conditioned video prediction into motion and appearance modeling. The multimodal motion estimation predicts future optical flow based on the audio-motion correlation. The visual branch recalls from the motion memory built from the audio features to enable better long term prediction. We further propose context-aware refinement to address the diminishing of the global appearance context in the long-term continuous warping. The global appearance context is extracted by the context encoder and manipulated by motion-conditioned affine transformation before fusion with features of warped frames. Experimental results show that our method achieves competitive results on existing benchmarks.