Recent advancements in Audio-Video Large Language Models (AV-LLMs) have enhanced their capabilities in tasks like audio-visual question answering and multimodal dialog systems. Video and audio introduce an extended temporal dimension, resulting in a larger key-value (KV) cache compared to static image embedding. A naive optimization strategy is to selectively focus on and retain KV caches of audio or video based on task. However, in the experiment, we observed that the attention of AV-LLMs to various modalities in the high layers is not strictly dependent on the task. In higher layers, the attention of AV-LLMs shifts more towards the video modality. In addition, we also found that directly integrating temporal KV of audio and spatial-temporal KV of video may lead to information confusion and significant performance degradation of AV-LLMs. If audio and video are processed indiscriminately, it may also lead to excessive compression or reservation of a certain modality, thereby disrupting the alignment between modalities. To address these challenges, we propose AccKV, an Adaptive-Focusing and Cross-Calibration KV cache optimization framework designed specifically for efficient AV-LLMs inference. Our method is based on layer adaptive focusing technology, selectively focusing on key modalities according to the characteristics of different layers, and enhances the recognition of heavy hitter tokens through attention redistribution. In addition, we propose a Cross-Calibration technique that first integrates inefficient KV caches within the audio and video modalities, and then aligns low-priority modality with high-priority modality to selectively evict KV cache of low-priority modality. The experimental results show that AccKV can significantly improve the computational efficiency of AV-LLMs while maintaining accuracy.

Predicting a user's next search query from recent interaction behaviors is a critical problem in modern e-commerce systems, particularly in scenarios where user intent evolves rapidly. Large Language Models (LLMs) offer strong semantic reasoning capabilities and have recently been adopted to enhance training data construction for next-query prediction. However, due to resource constraints on mobile devices, existing applications are deployed on cloud servers, resulting in high inference costs. In this paper, we propose RecGPT-Mobile, a framework that designs a lightweight LLM-based intent understanding agent to improve recommendation quality in mobile e-commerce scenarios. By deploying LLM directly on mobile devices, our approach can capture the evolving interests of users more quickly and adjust the recommendation results in real time. Extensive offline analyzes and online experiments demonstrate that our method significantly improves the accuracy of recommendation results, laying a practical path for LLM deployment in production-scale recommendation systems on mobile devices, as well as a scalable solution for integrating LLMs into real-world next-query prediction systems.

To guide consumers to mobile livestreams that involve their interested products, it is necessary to have a good understanding of livestreaming contents, and the key task is to accurately recognize the products being promoted by streamers. The mainstream cloud-based service framework is challenged by the high concurrency of requests, the high overhead of multimodal recognition, and the requirement of low latency. To break these bottlenecks, we propose a new device-cloud collaborative service framework, where each streamer's mobile device holds unimodal model that can process most of video frames and also uploads the extracted unimodal features to facilitate the cloud-side multimodal recognition of the remaining few frames. On-device unimodal model is further incrementally trained over the samples constructed by leveraging the streamers' manual labeling behaviors, leading to personalized version that can adapt to the heterogeneous and dynamic livestreaming contents of different streamers. Nevertheless, the device-side personalized unimodal features are misaligned in feature space and cannot be directly fused into the cloud-side multimodal model. We thus design a pluggable prompt generation module to transform the personalized unimodal features into prompt embeddings and prepend them to the original inputs of the multimodal backbone network, instructing feature fusion in a streamer-specific manner. We extensively evaluate on the public Fashion-Gen dataset and an industrial dataset collected from Taobao Live. We also perform online testing and practical overhead testing in Taobao Live. Evaluation results reveal the effectiveness and efficiency of our design as well as the consistent advantage over existing baselines.

Kunxi Li, Zhonghua Jiang, Zhouzhou Shen, ZhaodeWang ZhaodeWang, Chengfei Lv, Shengyu Zhang, Fan Wu, Fei Wu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Existing work on large language model (LLM) personalization assigned different responding roles to LLMs, but overlooked the diversity of queriers. In this work, we propose a new form of querier-aware LLM personalization, generating different responses even for the same query from different queriers. We design a dual-tower model architecture with a cross-querier general encoder and a querier-specific encoder. We further apply contrastive learning with multi-view augmentation, pulling close the dialogue representations of the same querier, while pulling apart those of different queriers. To mitigate the impact of query diversity on querier-contrastive learning, we cluster the dialogues based on query similarity and restrict the scope of contrastive learning within each cluster. To address the lack of datasets designed for querier-aware personalization, we also build a multi-querier dataset from English and Chinese scripts, as well as WeChat records, called MQDialog, containing 173 queriers and 12 responders. Extensive evaluations demonstrate that our design significantly improves the quality of personalized response generation, achieving relative improvement of 8.4% to 48.7% in ROUGE-L scores and winning rates ranging from 54% to 82% compared with various baseline methods.

Realistic 3D full-body talking avatars hold great potential in AR, with applications ranging from e-commerce live streaming to holographic communication. Despite advances in 3D Gaussian Splatting (3DGS) for lifelike avatar creation, existing methods struggle with fine-grained control of facial expressions and body movements in full-body talking tasks. Additionally, they often lack sufficient details and cannot run in real-time on mobile devices. We present TaoAvatar, a high-fidelity, lightweight, 3DGS-based full-body talking avatar driven by various signals. Our approach starts by creating a personalized clothed human parametric template that binds Gaussians to represent appearances. We then pre-train a Style-Unet-based network to handle complex pose-dependent non-rigid deformation, which can capture high-frequency appearance details but is too resource-intensive for mobile devices. To overcome this, we "bake" the non-rigid deformations into a lightweight MLP-based network using a distillation technique and develop blend shapes to compensate for details. Extensive experiments show that TaoAvatar achieves state-of-the-art rendering quality while running in real-time across various devices, maintaining 90 FPS on high-definition stereo devices such as the Apple Vision Pro.

Stereo video synthesis from monocular input is challenging in spatial computing and virtual reality due to the lack of high-quality stereo video pairs for training and the difficulty of maintaining spatio-temporal consistency between frames. Existing methods primarily address these issues by directly applying novel view synthesis (NVS) techniques to video, while facing limitations such as the inability to effectively represent dynamic scenes and the requirement for extensive training data. In this paper, we introduce a novel self-supervised stereo video synthesis paradigm via a video diffusion model, termed SpatialDreamer, which meets the challenges head-on. Firstly, to address the stereo video data insufficiency, we propose a Depth based Video Generation module DVG, which employs a forward-backward rendering mechanism to generate paired videos with geometric and temporal priors. Leveraging data generated by DVG, we propose RefinerNet along with a self-supervised synthetic framework designed to facilitate efficient and dedicated training. More importantly, we devise a consistency control module, which consists of a metric of stereo deviation strength and a Temporal Interaction Learning module TIL for geometric and temporal consistency ensurance respectively. We evaluated the proposed method against various benchmark methods, with the results showcasing its superior performance. Our project website is at: https://spatialdreamer.github.io.

Generating 360-degree panoramas from narrow field of view (NFoV) image is a promising computer vision task for Virtual Reality (VR) applications. Existing methods mostly assess the generated panoramas with InceptionNet or CLIP based metrics, which tend to perceive the image quality and is not suitable for evaluating the distortion. In this work, we first propose a distortion-specific CLIP, named Distort-CLIP to accurately evaluate the panorama distortion and discover the "visual cheating" phenomenon in previous works (i.e., tending to improve the visual results by sacrificing distortion accuracy). This phenomenon arises because prior methods employ a single network to learn the distinct panorama distortion and content completion at once, which leads the model to prioritize optimizing the latter. To address the phenomenon, we propose PanoDecouple, a decoupled diffusion model framework, which decouples the panorama generation into distortion guidance and content completion, aiming to generate panoramas with both accurate distortion and visual appeal. Specifically, we design a DistortNet for distortion guidance by imposing panorama-specific distortion prior and a modified condition registration mechanism; and a ContentNet for content completion by imposing perspective image information. Additionally, a distortion correction loss function with Distort-CLIP is introduced to constrain the distortion explicitly. The extensive experiments validate that PanoDecouple surpasses existing methods both in distortion and visual metrics.

Evaluating the response quality of large language models (LLMs) for open-ended questions poses a significant challenge, especially given the subjectivity and multi-dimensionality of "quality" in natural language generation. Existing LLM evaluators often neglect that different scenarios require distinct evaluation criteria. In this work, we propose **SaMer**, a scenario-aware multi-dimensional evaluator designed to provide both overall and fine-grained assessments of LLM-generated responses. Unlike fixed-dimension evaluation approaches, SaMer adapts to different scenarios by automatically identifying and prioritizing relevant evaluation dimensions tailored to the given query. To achieve this, we construct a large-scale fine-grained preference dataset spanning multiple real-world scenarios, each with distinct evaluation dimensions. We then leverage a text embedding model combined with three specialized heads to predict the appropriate evaluation dimensions and corresponding scores, as well as the respective weights that contribute to the overall score. The resulting model offers fine-grained and interpretable evaluations and shows robust adaptability across diverse scenarios. Extensive experiments on eight single rating and pairwise comparison datasets demonstrate that SaMer outperforms existing baselines in a variety of evaluation tasks, showcasing its robustness, versatility, and generalizability.

Speech-driven 3D facial animation has attracted significant attention due to its wide range of applications in animation production and virtual reality. Recent research has explored speech-emotion disentanglement to enhance facial expressions rather than manually assigning emotions. However, this approach face issues such as feature confusion, emotions weakening and mean-face. To address these issues, we present EcoFace, a framework that (1) proposes a novel collaboration objective to provide a explicit signal for emotion representation learning from the speaker's expressive movements and produced sounds, constructing an audio-visual joint and coordinated emotion space that is independent of speech content. (2) constructs a universal facial motion distribution space determined by speech features and implement speaker-specific generation. Extensive experiments show that our method achieves more generalized and emotionally realistic talking face generation compared to previous methods.

Speech-driven 3D facial animation aims to create lifelike facial expressions that synchronize accurately with speech. Despite significant progress, many existing methods may focus on generating facial animation with a fixed emotional state, neglecting the diverse transformations of facial emotions under a given speech input. To solve this issue, we focus on exploring the refined alignment between speech representations and multiple domains in facial expression information. We aim to disentangle the spoken language and emotion facial priors from speech expressions, to guide the refinement of the facial vertices based on speech. To accomplish this objective, we propose ExpTalk, which first applies an Adaptive Disentanglement Variational Autoencoder (AD-VAE) to decouple facial expression aligned with spoken language and emotions of speech through contrastive learning. Then a Refined Alignment Diffusion (RAD) is employed to iteratively refine the decoupled facial expression priors through diffusion-based perturbations, producing facial animations that align with the emotional variations of the given speech. Extensive experiments prove the effectiveness of our ExpTalk by surpassing state-of-the-arts by a large margin.

We present HRM2Avatar, a novel framework for creating high-fidelity avatars from monocular phone scans, which can be rendered and animated in real-time on mobile devices. Monocular capture with commodity smartphones provides a low-cost, pervasive alternative to studio-grade multi-camera rigs, making avatar digitization accessible to non-expert users. Reconstructing high-fidelity avatars from single-view video sequences poses significant challenges due to deficient visual and geometric data relative to multi-camera setups. To address these limitations, at the data level, our method leverages two types of data captured with smartphones: static pose sequences for detailed texture reconstruction and dynamic motion sequences for learning pose-dependent deformations and lighting changes. At the representation level, we employ a lightweight yet expressive representation to reconstruct high-fidelity digital humans from sparse monocular data. First, we extract explicit garment meshes from monocular data to model clothing deformations more effectively. Second, we attach illumination-aware Gaussians to the mesh surface, enabling high-fidelity rendering and capturing pose-dependent lighting changes. This representation efficiently learns high-resolution and dynamic information from our tailored monocular data, enabling the creation of detailed avatars. At the rendering level, real-time performance is critical for rendering and animating high-fidelity avatars in AR/VR, social gaming, and on-device creation, demanding sub-frame responsiveness. Our fully GPU-driven rendering pipeline delivers 120 FPS on mobile devices and 90 FPS on standalone VR devices at 2K resolution, over 2.7 × faster than representative mobile-engine baselines. Experiments show that HRM2Avatar delivers superior visual realism and real-time interactivity at high resolutions, outperforming state-of-the-art monocular methods.

Augmented reality (AR) offers users immersive experiences to interact with digital contents in their physical space. However, practical AR applications are challenged by the tight coupling of algorithm and engineering during the development and deployment phases as well as the execution requirements of hybrid AR subtasks on heterogeneous and resource-constraint mobile devices. In this work, we build an end-to-end, cross-platform, and efficient AR system, called ARSys. The infrastructure in ARSys adopts the new principle of integrated design, unifies and refines AR fundamental capabilities, supports streaming media processing, model inference, and real-time rendering by exposing high-performance tensor compute engine to top, and constructs a Python multi-instance virtual machine as the cross-platform AR task execution container. The runtime mechanism of ARSys schedules AR tasks in a pipeline parallelism way and allocates subtasks to hardware backends by optimizing the slowest node. The development workbench and the deployment platform in ARSys allow the decoupling of algorithms written in Python from engineering components in C/C++ and further support remote debugging and quick validation of AR algorithms. We extensively evaluate ARSys in practical AR applications across high-end, mid-end, and low-end Android and iOS devices, demonstrating higher development, deployment, and runtime efficiency than existing MediaPipe-oriented framework. ARSys has been integrated into Mobile Taobao for production use.

Reconstructing objects from posed images is a crucial and complex task in computer graphics and computer vision. While NeRF-based neural reconstruction methods have exhibited impressive reconstruction ability, they tend to be time-comsuming. Recent strategies have adopted 3D Gaussian Splatting (3D-GS) for inverse rendering, which have led to quick and effective outcomes. However, these techniques generally have difficulty in producing believable geometries and materials for glossy objects, a challenge that stems from the inherent ambiguities of inverse rendering. To address this, we introduce GlossyGS, an innovative 3D-GS-based inverse rendering framework that aims to precisely reconstruct the geometry and materials of glossy objects by integrating material priors. The key idea is the use of micro-facet geometry segmentation prior, which helps to reduce the intrinsic ambiguities and improve the decomposition of geometries and materials. Additionally, we introduce a normal map prefiltering strategy to more accurately simulate the normal distribution of reflective surfaces. These strategies are integrated into a hybrid geometry and material representation that employs both explicit and implicit methods to depict glossy objects. We demonstrate through quantitative analysis and qualitative visualization that the proposed method is effective to reconstruct high-fidelity geometries and materials of glossy objects, and performs favorably against State-of-the-Arts.

Shuofei Qiao, Ningyu Zhang, Runnan Fang, Yujie Luo, Wangchunshu Zhou, Yuchen Jiang, Chengfei Lv, Huajun Chen. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Recent advancements in single image driven 3D content generation have been propelled by leveraging prior knowledge from pretrained 2D diffusion models. However, the 3D content generated by existing methods often exhibits distorted outline shapes and inadequate details. To solve this problem, we propose a novel framework called Mask-enhanced Progressive Outline-to-Detail optimization (aka. MPOD123), which consists of two stages. Specifically, in the first stage, MPOD123 utilizes the pretrained view-conditioned diffusion model to guide the outline shape optimization of the 3D content. Given certain viewpoint, we estimate outline shape priors in the form of 2D mask from the 3D content by leveraging opacity calculation. In the second stage, MPOD123 incorporates Detail Appearance Inpainting (DAI) to guide the refinement on local geometry and texture with the shape priors. The essence of DAI lies in the Mask Rectified Cross-Attention (MRCA), which can be conveniently plugged in the stable diffusion model. The MRCA module utilizes the mask to rectify the attention map from each cross-attention layer. Accompanied with this new module, DAI is capable of guiding the detail refinement of the 3D content, while better preserves the outline shape. To assess the applicability in practical scenarios, we contribute a new dataset modeled on real-world e-commerce environments. Extensive quantitative and qualitative experiments on this dataset and open benchmarks demonstrate the effectiveness of MPOD123 over the state-of-the-arts.

Recent works on audio-driven talking head synthesis using Neural Radiance Fields (NeRF) have achieved impressive results. However, due to inadequate pose and expression control caused by NeRF implicit representation, these methods still have some limitations, such as unsynchronized or unnatural lip movements, and visual jitter and artifacts. In this paper, we propose GaussianTalker, a novel method for audio-driven talking head synthesis based on 3D Gaussian Splatting. With the explicit representation property of 3D Gaussians, intuitive control of the facial motion is achieved by binding Gaussians to 3D facial models. GaussianTalker consists of two modules, Speaker-specific Motion Translator and Dynamic Gaussian Renderer. Speaker-specific Motion Translator achieves accurate lip movements specific to the target speaker through universalized audio feature extraction and customized lip motion generation. Dynamic Gaussian Renderer introduces Speaker-specific BlendShapes to enhance facial detail representation via a latent pose, delivering stable and realistic rendered videos. Extensive experimental results suggest that GaussianTalker outperforms existing state-of-the-art methods in talking head synthesis, delivering precise lip synchronization and exceptional visual quality. Our method achieves rendering speeds of 130 FPS on NVIDIA RTX4090 GPU, significantly exceeding the threshold for real-time rendering performance, and can potentially be deployed on other hardware platforms.

To provide full-body 3D videos of performers showcasing diverse clothes with dynamic movements for e-commerce platforms, we develop an end-to-end, low-cost, and high-fidelity production and deployment pipeline. We first set up a low-cost capture studio with only 24 RGB cameras and embrace fast neural surface reconstruction to produce high-quality meshes without depth information. We then quickly group all the frames with local motion priors, select a keyframe for each group, and accurately register any other frame to the keyframe under the guidance of semantic labels, thereby avoiding transmitting all the frames to mobile devices and loading them into memory. For real-time rendering, we propose an on-device sparse computation method for efficient deformation from keyframes to the other frames. Evaluation over 2 self-captured performances and 8 public performances reveals that the pipeline achieves the reconstruction time of 28 minutes per frame, the average PSNR of 30.4, the average bandwidth requirement of 4.2MB/s, and the on-device frame rate of 60 fps, demonstrating superiority over existing baselines.

Shuofei Qiao, Honghao Gui, Chengfei Lv, Qianghuai Jia, Huajun Chen, Ningyu Zhang. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

This paper studies implicit surface reconstruction leveraging differentiable ray casting. Previous works such as IDR [34] and NeuS [27] overlook the spatial context in 3D space when predicting and rendering the surface, thereby may fail to capture sharp local topologies such as small holes and structures. To mitigate the limitation, we propose a flexible neural implicit representation leveraging hierarchical voxel grids, namely Neural Deformable Anchor (NeuDA), for high-fidelity surface reconstruction. NeuDA maintains the hierarchical anchor grids where each vertex stores a 3D position (or anchor) instead of the direct embedding (or feature). We optimize the anchor grids such that different local geometry structures can be adaptively encoded. Besides, we dig into the frequency encoding strategies and introduce a simple hierarchical positional encoding method for the hierarchical anchor structure to flexibly exploit the properties of high-frequency and low-frequency geometry and appearance. Experiments on both the DTU [8] and BlendedMVS [32] datasets demonstrate that NeuDA can produce promising mesh surfaces.

Federated learning (FL) trains a machine learning model on mobile devices in a distributed manner using each device’s private data and computing resources. A critical issues is to evaluate individual users’ contributions so that (1) users’ effort in model training can be compensated with proper incentives and (2) malicious and low-quality users can be detected and removed. The state-of-the-art solutions require a representative test dataset for the evaluation purpose, but such a dataset is often unavailable and hard to synthesize. In this paper, we propose a method called Pairwise Correlated Agreement (PCA) based on the idea of peer prediction to evaluate user contribution in FL without a test dataset. PCA achieves this using the statistical correlation of the model parameters uploaded by users. We then apply PCA to designing (1) a new federated learning algorithm called Fed-PCA, and (2) a new incentive mechanism that guarantees truthfulness. We evaluate the performance of PCA and Fed-PCA using the MNIST dataset and a large industrial product recommendation dataset. The results demonstrate that our Fed-PCA outperforms the canonical FedAvg algorithm and other baseline methods in accuracy, and at the same time, PCA effectively incentivizes users to behave truthfully.

Federated learning was proposed with an intriguing vision of achieving collaborative machine learning among numerous clients without uploading their private data to a cloud server. However, the conventional framework requires each client to leverage the full model for learning, which can be prohibitively inefficient for large-scale learning tasks and resource-constrained mobile devices. Thus, we proposed a submodel framework, where clients download only the needed parts of the full model, namely, submodels, and then upload the submodel updates. Nevertheless, the "position" of a client's truly required submodel corresponds to its private data, while the disclosure of the true position to the cloud server during interactions inevitably breaks the tenet of federated learning. To integrate efficiency and privacy, we designed a secure federated submodel learning scheme coupled with a private set union protocol as a cornerstone. The secure scheme features the properties of randomized response, secure aggregation, and Bloom filter, and endows each client with customized plausible deniability (in terms of local differential privacy) against the position of its desired submodel, thereby protecting private data. We further instantiated the scheme with Alibaba's e-commerce recommendation, implemented a prototype system, and extensively evaluated over 30-day Taobao user data. Empirical results demonstrate the feasibility and scalability of the proposed scheme as well as its remarkable advantages over the conventional federated learning framework, from model accuracy and convergency, practical communication, computation, and storage overhead.