Yuming Du

Wen Guo, Yuming Du, Xi Shen et al. · tencent-ai

This paper tackles the problem of human motion prediction, consisting in forecasting future body poses from historically observed sequences. State-of-the-art approaches provide good results, however, they rely on deep learning architectures of arbitrary complexity, such as Recurrent Neural Networks(RNN), Transformers or Graph Convolutional Networks(GCN), typically requiring multiple training stages and more than 2 million parameters. In this paper, we show that, after combining with a series of standard practices, such as applying Discrete Cosine Transform(DCT), predicting residual displacement of joints and optimizing velocity as an auxiliary loss, a light-weight network based on multi-layer perceptrons(MLPs) with only 0.14 million parameters can surpass the state-of-the-art performance. An exhaustive evaluation on the Human3.6M, AMASS, and 3DPW datasets shows that our method, named siMLPe, consistently outperforms all other approaches. We hope that our simple method could serve as a strong baseline for the community and allow re-thinking of the human motion prediction problem. The code is publicly available at \url{https://github.com/dulucas/siMLPe}.

Yangtao Wang, Xi Shen, Yuan Yuan et al. · mit

In this paper, we describe a graph-based algorithm that uses the features obtained by a self-supervised transformer to detect and segment salient objects in images and videos. With this approach, the image patches that compose an image or video are organised into a fully connected graph, where the edge between each pair of patches is labeled with a similarity score between patches using features learned by the transformer. Detection and segmentation of salient objects is then formulated as a graph-cut problem and solved using the classical Normalized Cut algorithm. Despite the simplicity of this approach, it achieves state-of-the-art results on several common image and video detection and segmentation tasks. For unsupervised object discovery, this approach outperforms the competing approaches by a margin of 6.1%, 5.7%, and 2.6%, respectively, when tested with the VOC07, VOC12, and COCO20K datasets. For the unsupervised saliency detection task in images, this method improves the score for Intersection over Union (IoU) by 4.4%, 5.6% and 5.2%. When tested with the ECSSD, DUTS, and DUT-OMRON datasets, respectively, compared to current state-of-the-art techniques. This method also achieves competitive results for unsupervised video object segmentation tasks with the DAVIS, SegTV2, and FBMS datasets.

Georgy Ponimatkin, Nermin Samet, Yang Xiao et al.

We propose a simple, yet powerful approach for unsupervised object segmentation in videos. We introduce an objective function whose minimum represents the mask of the main salient object over the input sequence. It only relies on independent image features and optical flows, which can be obtained using off-the-shelf self-supervised methods. It scales with the length of the sequence with no need for superpixels or sparsification, and it generalizes to different datasets without any specific training. This objective function can actually be derived from a form of spectral clustering applied to the entire video. Our method achieves on-par performance with the state of the art on standard benchmarks (DAVIS2016, SegTrack-v2, FBMS59), while being conceptually and practically much simpler. Code is available at https://ponimatkin.github.io/ssl-vos.

Van Nguyen Nguyen, Yuming Du, Yang Xiao et al.

Estimating the relative pose of a new object without prior knowledge is a hard problem, while it is an ability very much needed in robotics and Augmented Reality. We present a method for tracking the 6D motion of objects in RGB video sequences when neither the training images nor the 3D geometry of the objects are available. In contrast to previous works, our method can therefore consider unknown objects in open world instantly, without requiring any prior information or a specific training phase. We consider two architectures, one based on two frames, and the other relying on a Transformer Encoder, which can exploit an arbitrary number of past frames. We train our architectures using only synthetic renderings with domain randomization. Our results on challenging datasets are on par with previous works that require much more information (training images of the target objects, 3D models, and/or depth data). Our source code is available at https://github.com/nv-nguyen/pizza

Yuming Du, Robin Kips, Albert Pumarola et al.

With the recent surge in popularity of AR/VR applications, realistic and accurate control of 3D full-body avatars has become a highly demanded feature. A particular challenge is that only a sparse tracking signal is available from standalone HMDs (Head Mounted Devices), often limited to tracking the user's head and wrists. While this signal is resourceful for reconstructing the upper body motion, the lower body is not tracked and must be synthesized from the limited information provided by the upper body joints. In this paper, we present AGRoL, a novel conditional diffusion model specifically designed to track full bodies given sparse upper-body tracking signals. Our model is based on a simple multi-layer perceptron (MLP) architecture and a novel conditioning scheme for motion data. It can predict accurate and smooth full-body motion, particularly the challenging lower body movement. Unlike common diffusion architectures, our compact architecture can run in real-time, making it suitable for online body-tracking applications. We train and evaluate our model on AMASS motion capture dataset, and demonstrate that our approach outperforms state-of-the-art methods in generated motion accuracy and smoothness. We further justify our design choices through extensive experiments and ablation studies.

Boxun Xu, Yuming Du, Zichang Liu et al.

We introduce Sparse Forcing, a training-and-inference paradigm for autoregressive video diffusion models that improves long-horizon generation quality while reducing decoding latency. Sparse Forcing is motivated by an empirical observation in autoregressive diffusion rollouts: attention concentrates on a persistent subset of salient visual blocks, forming an implicit spatiotemporal memory in the KV cache, and exhibits a locally structured block-sparse pattern within sliding windows. Building on this observation, we propose a trainable native sparsity mechanism that learns to compress, preserve, and update these persistent blocks while restricting computation within each local window to a dynamically selected local neighborhood. To make the approach practical at scale for both training and inference, we further propose Persistent Block-Sparse Attention (PBSA), an efficient GPU kernel that accelerates sparse attention and memory updates for low-latency, memory-efficient decoding. Experiments show that Sparse Forcing improves the VBench score by +0.26 over Self-Forcing on 5-second text-to-video generation while delivering a 1.11-1.17x decoding speedup and 42% lower peak KV-cache footprint. The gains are more pronounced on longer-horizon rollouts, delivering improved visual quality with +0.68 and +2.74 VBench improvements, and 1.22x and 1.27x speedups on 20-second and 1-minute generations, respectively.

Cheeun Hong, German Barquero, Fadime Sener et al.

Instructional video generation is an emerging task that aims to synthesize coherent demonstrations of procedural activities from textual descriptions. Such capability has broad implications for content creation, education, and human-AI interaction, yet existing video diffusion models struggle to maintain temporal consistency and controllability across long sequences of multiple action steps. We introduce a pipeline for future-driven streaming instructional video generation, dubbed SneakPeek, a diffusion-based autoregressive framework designed to generate precise, stepwise instructional videos conditioned on an initial image and structured textual prompts. Our approach introduces three key innovations to enhance consistency and controllability: (1) predictive causal adaptation, where a causal model learns to perform next-frame prediction and anticipate future keyframes; (2) future-guided self-forcing with a dual-region KV caching scheme to address the exposure bias issue at inference time; (3) multi-prompt conditioning, which provides fine-grained and procedural control over multi-step instructions. Together, these components mitigate temporal drift, preserve motion consistency, and enable interactive video generation where future prompt updates dynamically influence ongoing streaming video generation. Experimental results demonstrate that our method produces temporally coherent and semantically faithful instructional videos that accurately follow complex, multi-step task descriptions.

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.

Yeongmin Kim, Sotiris Anagnostidis, Yuming Du et al.

Diffusion models with transformer architectures have demonstrated promising capabilities in generating high-fidelity images and scalability for high resolution. However, iterative sampling process required for synthesis is very resource-intensive. A line of work has focused on distilling solutions to probability flow ODEs into few-step student models. Nevertheless, existing methods have been limited by their reliance on the most recent denoised samples as input, rendering them susceptible to exposure bias. To address this limitation, we propose AutoRegressive Distillation (ARD), a novel approach that leverages the historical trajectory of the ODE to predict future steps. ARD offers two key benefits: 1) it mitigates exposure bias by utilizing a predicted historical trajectory that is less susceptible to accumulated errors, and 2) it leverages the previous history of the ODE trajectory as a more effective source of coarse-grained information. ARD modifies the teacher transformer architecture by adding token-wise time embedding to mark each input from the trajectory history and employs a block-wise causal attention mask for training. Furthermore, incorporating historical inputs only in lower transformer layers enhances performance and efficiency. We validate the effectiveness of ARD in a class-conditioned generation on ImageNet and T2I synthesis. Our model achieves a $5\times$ reduction in FID degradation compared to the baseline methods while requiring only 1.1\% extra FLOPs on ImageNet-256. Moreover, ARD reaches FID of 1.84 on ImageNet-256 in merely 4 steps and outperforms the publicly available 1024p text-to-image distilled models in prompt adherence score with a minimal drop in FID compared to the teacher. Project page: https://github.com/alsdudrla10/ARD.

Gregor Bachmann, Sotiris Anagnostidis, Albert Pumarola et al.

The performance of large language models (LLMs) is closely linked to their underlying size, leading to ever-growing networks and hence slower inference. Speculative decoding has been proposed as a technique to accelerate autoregressive generation, leveraging a fast draft model to propose candidate tokens, which are then verified in parallel based on their likelihood under the target model. While this approach guarantees to reproduce the target output, it incurs a substantial penalty: many high-quality draft tokens are rejected, even when they represent objectively valid continuations. Indeed, we show that even powerful draft models such as GPT-4o, as well as human text cannot achieve high acceptance rates under the standard verification scheme. This severely limits the speedup potential of current speculative decoding methods, as an early rejection becomes overwhelmingly likely when solely relying on alignment of draft and target. We thus ask the following question: Can we adapt verification to recognize correct, but non-aligned replies? To this end, we draw inspiration from the LLM-as-a-judge framework, which demonstrated that LLMs are able to rate answers in a versatile way. We carefully design a dataset to elicit the same capability in the target model by training a compact module on top of the embeddings to produce ``judgements" of the current continuation. We showcase our strategy on the Llama-3.1 family, where our 8b/405B-Judge achieves a speedup of 9x over Llama-405B, while maintaining its quality on a large range of benchmarks. These benefits remain present even in optimized inference frameworks, where our method reaches up to 141 tokens/s for 8B/70B-Judge and 129 tokens/s for 8B/405B on 2 and 8 H100s respectively.

Sotiris Anagnostidis, Gregor Bachmann, Yeongmin Kim et al.

Despite their remarkable performance, modern Diffusion Transformers are hindered by substantial resource requirements during inference, stemming from the fixed and large amount of compute needed for each denoising step. In this work, we revisit the conventional static paradigm that allocates a fixed compute budget per denoising iteration and propose a dynamic strategy instead. Our simple and sample-efficient framework enables pre-trained DiT models to be converted into \emph{flexible} ones -- dubbed FlexiDiT -- allowing them to process inputs at varying compute budgets. We demonstrate how a single \emph{flexible} model can generate images without any drop in quality, while reducing the required FLOPs by more than $40$\% compared to their static counterparts, for both class-conditioned and text-conditioned image generation. Our method is general and agnostic to input and conditioning modalities. We show how our approach can be readily extended for video generation, where FlexiDiT models generate samples with up to $75$\% less compute without compromising performance.

Yuming Du, Wen Guo, Yang Xiao et al.

In this report, we introduce our (pretty straightforard) two-step "detect-then-match" video instance segmentation method. The first step performs instance segmentation for each frame to get a large number of instance mask proposals. The second step is to do inter-frame instance mask matching with the help of optical flow. We demonstrate that with high quality mask proposals, a simple matching mechanism is good enough for tracking. Our approach achieves the first place in the UVO 2021 Video-based Open-World Segmentation Challenge.

Yuming Du, Wen Guo, Yang Xiao et al.

We describe our two-stage instance segmentation framework we use to compete in the challenge. The first stage of our framework consists of an object detector, which generates object proposals in the format of bounding boxes. Then, the images and the detected bounding boxes are fed to the second stage, where a segmentation network is applied to segment the objects in the bounding boxes. We train all our networks in a class-agnostic way. Our approach achieves the first place in the UVO 2021 Image-based Open-World Segmentation Challenge.

Yang Xiao, Yuming Du, Renaud Marlet

Motivated by the need for estimating the 3D pose of arbitrary objects, we consider the challenging problem of class-agnostic object viewpoint estimation from images only, without CAD model knowledge. The idea is to leverage features learned on seen classes to estimate the pose for classes that are unseen, yet that share similar geometries and canonical frames with seen classes. We train a direct pose estimator in a class-agnostic way by sharing weights across all object classes, and we introduce a contrastive learning method that has three main ingredients: (i) the use of pre-trained, self-supervised, contrast-based features; (ii) pose-aware data augmentations; (iii) a pose-aware contrastive loss. We experimented on Pascal3D+, ObjectNet3D and Pix3D in a cross-dataset fashion, with both seen and unseen classes. We report state-of-the-art results, including against methods that additionally use CAD models as input.

Yuming Du, Yang Xiao, Vincent Lepetit

The ability to localize and segment objects from unseen classes would open the door to new applications, such as autonomous object learning in active vision. Nonetheless, improving the performance on unseen classes requires additional training data, while manually annotating the objects of the unseen classes can be labor-extensive and expensive. In this paper, we explore the use of unlabeled video sequences to automatically generate training data for objects of unseen classes. It is in principle possible to apply existing video segmentation methods to unlabeled videos and automatically obtain object masks, which can then be used as a training set even for classes with no manual labels available. However, our experiments show that these methods do not perform well enough for this purpose. We therefore introduce a Bayesian method that is specifically designed to automatically create such a training set: Our method starts from a set of object proposals and relies on (non-realistic) analysis-by-synthesis to select the correct ones by performing an efficient optimization over all the frames simultaneously. Through extensive experiments, we show that our method can generate a high-quality training set which significantly boosts the performance of segmenting objects of unseen classes. We thus believe that our method could open the door for open-world instance segmentation using abundant Internet videos.

Michael Ramamonjisoa, Yuming Du, Vincent Lepetit

Current methods for depth map prediction from monocular images tend to predict smooth, poorly localized contours for the occlusion boundaries in the input image. This is unfortunate as occlusion boundaries are important cues to recognize objects, and as we show, may lead to a way to discover new objects from scene reconstruction. To improve predicted depth maps, recent methods rely on various forms of filtering or predict an additive residual depth map to refine a first estimate. We instead learn to predict, given a depth map predicted by some reconstruction method, a 2D displacement field able to re-sample pixels around the occlusion boundaries into sharper reconstructions. Our method can be applied to the output of any depth estimation method, in an end-to-end trainable fashion. For evaluation, we manually annotated the occlusion boundaries in all the images in the test split of popular NYUv2-Depth dataset. We show that our approach improves the localization of occlusion boundaries for all state-of-the-art monocular depth estimation methods that we could evaluate, without degrading the depth accuracy for the rest of the images.

Wenbo Li, Zhicheng Wang, Binyi Yin et al.

Existing pose estimation approaches fall into two categories: single-stage and multi-stage methods. While multi-stage methods are seemingly more suited for the task, their performance in current practice is not as good as single-stage methods. This work studies this issue. We argue that the current multi-stage methods' unsatisfactory performance comes from the insufficiency in various design choices. We propose several improvements, including the single-stage module design, cross stage feature aggregation, and coarse-to-fine supervision. The resulting method establishes the new state-of-the-art on both MS COCO and MPII Human Pose dataset, justifying the effectiveness of a multi-stage architecture. The source code is publicly available for further research.