Shengze Wang

Koki Nagano, Hongyu Liu, Seonwook Park et al.

We present DyaPlex, a streaming, full-duplex speech-and-motion model designed for dyadic interaction. To capture the continuous and reciprocal nature of human communication, this full-duplex capability empowers the agent to simultaneously perceive and generate both speech and physical motion in a streaming fashion. At its core, our method leverages the strong priors of a foundational full-duplex speech model and integrates a novel motion pathway, thereby achieving fully synchronized multi-modal interaction. Specifically, we design a dual-tower Transformer architecture that preserves the zero-shot conversational reasoning of a frozen base speech model while constructing a deeply coupled, streaming motion pathway. By introducing a unified dyadic token interleaving mechanism and guiding cross-attention via a time-aligned speech-motion RoPE, our model effectively aligns autoregressive motions with rich latent speech features. Trained on the 4,000-hour Seamless Interaction dataset, our model effectively captures cross-speaker dependencies and establishes new state-of-the-art performance across both monadic and dyadic human interaction benchmarks.

Shengze Wang, Alexey Supikov, Joshua Ratcliff et al.

Recent works in spatiotemporal radiance fields can produce photorealistic free-viewpoint videos. However, they are inherently unsuitable for interactive streaming scenarios (e.g. video conferencing, telepresence) because have an inevitable lag even if the training is instantaneous. This is because these approaches consume videos and thus have to buffer chunks of frames (often seconds) before processing. In this work, we take a step towards interactive streaming via a frame-by-frame approach naturally free of lag. Conventional wisdom believes that per-frame NeRFs are impractical due to prohibitive training costs and storage. We break this belief by introducing Incremental Neural Videos (INV), a per-frame NeRF that is efficiently trained and streamable. We designed INV based on two insights: (1) Our main finding is that MLPs naturally partition themselves into Structure and Color Layers, which store structural and color/texture information respectively. (2) We leverage this property to retain and improve upon knowledge from previous frames, thus amortizing training across frames and reducing redundant learning. As a result, with negligible changes to NeRF, INV can achieve good qualities (>28.6db) in 8min/frame. It can also outperform prior SOTA in 19% less training time. Additionally, our Temporal Weight Compression reduces the per-frame size to 0.3MB/frame (6.6% of NeRF). More importantly, INV is free from buffer lag and is naturally fit for streaming. While this work does not achieve real-time training, it shows that incremental approaches like INV present new possibilities in interactive 3D streaming. Moreover, our discovery of natural information partition leads to a better understanding and manipulation of MLPs. Code and dataset will be released soon.

Shengze Wang, YoungJoong Kwon, Yuan Shen et al.

There have been significant advancements in dynamic novel view synthesis in recent years. However, current deep learning models often require (1) prior models (e.g., SMPL human models), (2) heavy pre-processing, or (3) per-scene optimization. We propose to utilize RGBD cameras to remove these limitations and synthesize free-viewpoint videos of dynamic indoor scenes. We generate feature point clouds from RGBD frames and then render them into free-viewpoint videos via a neural renderer. However, the inaccurate, unstable, and incomplete depth measurements induce severe distortions, flickering, and ghosting artifacts. We enforce spatial-temporal consistency via the proposed Cycle Reconstruction Consistency and Temporal Stabilization module to reduce these artifacts. We introduce a simple Regional Depth-Inpainting module that adaptively inpaints missing depth values to render complete novel views. Additionally, we present a Human-Things Interactions dataset to validate our approach and facilitate future research. The dataset consists of 43 multi-view RGBD video sequences of everyday activities, capturing complex interactions between human subjects and their surroundings. Experiments on the HTI dataset show that our method outperforms the baseline per-frame image fidelity and spatial-temporal consistency. We will release our code, and the dataset on the website soon.

Amrita Mazumdar, Seonwook Park, Rajarshi Roy et al.

Natural human conversation is full-duplex and audio-visual: people simultaneously speak and listen while continuously interpreting and producing nonverbal cues, such as nods, smiles, and gestures. To support successful human-agent interaction, agents must model full-duplex audiovisual conversation; however, existing full-duplex benchmarks evaluate only speech. In this work, we present VideoFDB, the first benchmark to evaluate full-duplex audio-visual-to-audio-visual (AV2AV) conversational agents. VideoFDB contributes (i) 237 dyadic clips spanning 11 nonverbal conversational dynamics from real-world video calls, (ii) a taxonomy separating perception from generation behaviors, and (iii) a rubric-based LM-as-judge evaluation framework with interpretable axes for assessing conversational quality with respect to nonverbal conversational dynamics. Across open- and closed-source vision-speech agents, we find systematic failure modes: captioning collapse and visual-stream ignorance, and we show that current systems exploit vision for explicit visual question answering but not for the streaming joint audiovisual grounding required in natural conversation. We further evaluate cascaded speech-to-avatar systems and find that their architecture fundamentally precludes the production of full-duplex nonverbal cues. As the first benchmark for full-duplex AV2AV interaction, VideoFDB establishes a foundation for systematic evaluation and, we hope, will accelerate the advancement and development of next-generation multimodal conversational agents.

Shengze Wang, Ziheng Wang, Ryan Schmelzle et al.

In this paper, we work to bring telepresence to every desktop. Unlike commercial systems, personal 3D video conferencing systems must render high-quality videos while remaining financially and computationally viable for the average consumer. To this end, we introduce a capturing and rendering system that only requires 4 consumer-grade RGBD cameras and synthesizes high-quality free-viewpoint videos of users as well as their environments. Experimental results show that our system renders high-quality free-viewpoint videos without using object templates or heavy pre-processing. While not real-time, our system is fast and does not require per-video optimizations. Moreover, our system is robust to complex hand gestures and clothing, and it can generalize to new users. This work provides a strong basis for further optimization, and it will help bring telepresence to every desk in the near future. The code and dataset will be made available on our website https://mcmvmc.github.io/PersonalTelepresence/.

Luchao Qi, Jiaye Wu, Annie N. Wang et al.

In recent years, generative 3D face models (e.g., EG3D) have been developed to tackle the problem of synthesizing photo-realistic faces. However, these models are often unable to capture facial features unique to each individual, highlighting the importance of personalization. Some prior works have shown promise in personalizing generative face models, but these studies primarily focus on 2D settings. Also, these methods require both fine-tuning and storing a large number of parameters for each user, posing a hindrance to achieving scalable personalization. Another challenge of personalization is the limited number of training images available for each individual, which often leads to overfitting when using full fine-tuning methods. Our proposed approach, My3DGen, generates a personalized 3D prior of an individual using as few as 50 training images. My3DGen allows for novel view synthesis, semantic editing of a given face (e.g. adding a smile), and synthesizing novel appearances, all while preserving the original person's identity. We decouple the 3D facial features into global features and personalized features by freezing the pre-trained EG3D and training additional personalized weights through low-rank decomposition. As a result, My3DGen introduces only $\textbf{240K}$ personalized parameters per individual, leading to a $\textbf{127}\times$ reduction in trainable parameters compared to the $\textbf{30.6M}$ required for fine-tuning the entire parameter space. Despite this significant reduction in storage, our model preserves identity features without compromising the quality of downstream applications.

Fei Fang, Yifan Hua, Shengze Wang et al.

While significant progress has been made in research and development on open-source and cost-efficient large-language models (LLMs), serving scalability remains a critical challenge, particularly for small organizations and individuals seeking to deploy and test their LLM innovations. Inspired by peer-to-peer networks that leverage decentralized overlay nodes to increase throughput and availability, we propose GenTorrent, an LLM serving overlay that harnesses computing resources from decentralized contributors. We identify four key research problems inherent to enabling such a decentralized infrastructure: 1) overlay network organization; 2) LLM communication privacy; 3) overlay forwarding for resource efficiency; and 4) verification of serving quality. This work presents the first systematic study of these fundamental problems in the context of decentralized LLM serving. Evaluation results from a prototype implemented on a set of decentralized nodes demonstrate that GenTorrent achieves a latency reduction of over 50% compared to the baseline design without overlay forwarding. Furthermore, the security features introduce minimal overhead to serving latency and throughput. We believe this work pioneers a new direction for democratizing and scaling future AI serving capabilities.

Shengze Wang, Jiefeng Li, Tianye Li et al.

Single-image human mesh recovery is a challenging task due to the ill-posed nature of simultaneous body shape, pose, and camera estimation. Existing estimators work well on images taken from afar, but they break down as the person moves close to the camera. Moreover, current methods fail to achieve both accurate 3D pose and 2D alignment at the same time. Error is mainly introduced by inaccurate perspective projection heuristically derived from orthographic parameters. To resolve this long-standing challenge, we present our method BLADE which accurately recovers perspective parameters from a single image without heuristic assumptions. We start from the inverse relationship between perspective distortion and the person's Z-translation Tz, and we show that Tz can be reliably estimated from the image. We then discuss the important role of Tz for accurate human mesh recovery estimated from close-range images. Finally, we show that, once Tz and the 3D human mesh are estimated, one can accurately recover the focal length and full 3D translation. Extensive experiments on standard benchmarks and real-world close-range images show that our method is the first to accurately recover projection parameters from a single image, and consequently attain state-of-the-art accuracy on 3D pose estimation and 2D alignment for a wide range of images. https://research.nvidia.com/labs/amri/projects/blade/

Shengze Wang, Xueting Li, Chao Liu et al.

Recent breakthroughs in single-image 3D portrait reconstruction have enabled telepresence systems to stream 3D portrait videos from a single camera in real-time, democratizing telepresence. However, per-frame 3D reconstruction exhibits temporal inconsistency and forgets the user's appearance. On the other hand, self-reenactment methods can render coherent 3D portraits by driving a 3D avatar built from a single reference image, but fail to faithfully preserve the user's per-frame appearance (e.g., instantaneous facial expression and lighting). As a result, none of these two frameworks is an ideal solution for democratized 3D telepresence. In this work, we address this dilemma and propose a novel solution that maintains both coherent identity and dynamic per-frame appearance to enable the best possible realism. To this end, we propose a new fusion-based method that takes the best of both worlds by fusing a canonical 3D prior from a reference view with dynamic appearance from per-frame input views, producing temporally stable 3D videos with faithful reconstruction of the user's per-frame appearance. Trained only using synthetic data produced by an expression-conditioned 3D GAN, our encoder-based method achieves both state-of-the-art 3D reconstruction and temporal consistency on in-studio and in-the-wild datasets. https://research.nvidia.com/labs/amri/projects/coherent3d

Shengze Wang, Xueting Li, Chao Liu et al.

Recent breakthroughs in single-image 3D portrait reconstruction have enabled telepresence systems to stream 3D portrait videos from a single camera in real-time, potentially democratizing telepresence. However, per-frame 3D reconstruction exhibits temporal inconsistency and forgets the user's appearance. On the other hand, self-reenactment methods can render coherent 3D portraits by driving a personalized 3D prior, but fail to faithfully reconstruct the user's per-frame appearance (e.g., facial expressions and lighting). In this work, we recognize the need to maintain both coherent identity and dynamic per-frame appearance to enable the best possible realism. To this end, we propose a new fusion-based method that fuses a personalized 3D subject prior with per-frame information, producing temporally stable 3D videos with faithful reconstruction of the user's per-frame appearances. Trained only using synthetic data produced by an expression-conditioned 3D GAN, our encoder-based method achieves both state-of-the-art 3D reconstruction accuracy and temporal consistency on in-studio and in-the-wild datasets.

Sicong Huang, Qianqi Yan, Shengze Wang et al.

Abstractive summarization using large language models (LLMs) has become an essential tool for condensing information. However, despite their ability to generate fluent summaries, these models sometimes produce unfaithful summaries, introducing hallucinations at the word, phrase, or concept level. Existing mitigation strategies, such as post-processing corrections or contrastive learning with synthetically generated negative samples, fail to fully address the diverse errors that can occur in LLM-generated summaries. In this paper, we investigate fine-tuning strategies to reduce the occurrence of unfaithful spans in generated summaries. First, we automatically generate summaries for the set of source documents in the training set with a variety of LLMs and then use GPT-4o to annotate any hallucinations it detects at the span-level. Leveraging these annotations, we fine-tune LLMs with both hallucination-free summaries and annotated unfaithful spans to enhance model faithfulness. In this paper, we introduce a new dataset that contains both faithful and unfaithful summaries with span-level labels and we evaluate three techniques to fine-tuning a LLM to improve the faithfulness of the resulting summarization: gradient ascent, unlikelihood training, and task vector negation. Experimental results show that all three approaches successfully leverage span-level annotations to improve faithfulness, with unlikelihood training being the most effective.

Lipu Zhou, Shengze Wang, Jiamin Ye et al.

Estimating pose from given 3D correspondences, including point-to-point, point-to-line and point-to-plane correspondences, is a fundamental task in computer vision with many applications. We present a complete solution for this task, including a solution for the minimal problem and the least-squares problem of this task. Previous works mainly focused on finding the global minimizer to address the least-squares problem. However, existing works that show the ability to achieve global minimizer are still unsuitable for real-time applications. Furthermore, as one of contributions of this paper, we prove that there exist ambiguous configurations for any number of lines and planes. These configurations have several solutions in theory, which makes the correct solution may come from a local minimizer. Our algorithm is efficient and able to reveal local minimizers. We employ the Cayley-Gibbs-Rodriguez (CGR) parameterization of the rotation to derive a general rational cost for the three cases of 3D correspondences. The main contribution of this paper is to solve the resulting equation system of the minimal problem and the first-order optimality conditions of the least-squares problem, both of which are of complicated rational forms. The central idea of our algorithm is to introduce intermediate unknowns to simplify the problem. Extensive experimental results show that our algorithm significantly outperforms previous algorithms when the number of correspondences is small. Besides, when the global minimizer is the solution, our algorithm achieves the same accuracy as previous algorithms that have guaranteed global optimality, but our algorithm is applicable to real-time applications.

Lipu Zhou, Jiamin Ye, Montiel Abello et al.

We present a novel unsupervised learning framework for single view depth estimation using monocular videos. It is well known in 3D vision that enlarging the baseline can increase the depth estimation accuracy, and jointly optimizing a set of camera poses and landmarks is essential. In previous monocular unsupervised learning frameworks, only part of the photometric and geometric constraints within a sequence are used as supervisory signals. This may result in a short baseline and overfitting. Besides, previous works generally estimate a low resolution depth from a low resolution impute image. The low resolution depth is then interpolated to recover the original resolution. This strategy may generate large errors on object boundaries, as the depth of background and foreground are mixed to yield the high resolution depth. In this paper, we introduce a bundle adjustment framework and a super-resolution network to solve the above two problems. In bundle adjustment, depths and poses of an image sequence are jointly optimized, which increases the baseline by establishing the relationship between farther frames. The super resolution network learns to estimate a high resolution depth from a low resolution image. Additionally, we introduce the clip loss to deal with moving objects and occlusion. Experimental results on the KITTI dataset show that the proposed algorithm outperforms the state-of-the-art unsupervised methods using monocular sequences, and achieves comparable or even better result compared to unsupervised methods using stereo sequences.