Yihua Cheng

Yuhan Liu, Hanchen Li, Yihua Cheng et al. · stanford

As large language models (LLMs) take on complex tasks, their inputs are supplemented with longer contexts that incorporate domain knowledge. Yet using long contexts is challenging, as nothing can be generated until the whole context is processed by the LLM. While the context-processing delay can be reduced by reusing the KV cache of a context across different inputs, fetching the KV cache, which contains large tensors, over the network can cause high extra network delays. CacheGen is a fast context-loading module for LLM systems. First, CacheGen uses a custom tensor encoder, leveraging KV cache's distributional properties to encode a KV cache into more compact bitstream representations with negligible decoding overhead, to save bandwidth usage. Second, CacheGen adapts the compression level of different parts of a KV cache to cope with changes in available bandwidth, in order to maintain low context-loading delay and high generation quality. % When available bandwidth drops, CacheGen may raise the compression level for a part of the context or recompute its KV cache on the fly. We test CacheGen on popular LLMs and datasets. Compared to the recent systems that reuse the KV cache, CacheGen reduces the KV cache size by 3.5-4.3x and the total delay in fetching and processing contexts by 3.2-3.7x with negligible impact on the LLM response quality. Our code is at: https://github.com/UChi-JCL/CacheGen.

Yihua Cheng, Feng Lu

Gaze estimation methods estimate gaze from facial appearance with a single camera. However, due to the limited view of a single camera, the captured facial appearance cannot provide complete facial information and thus complicate the gaze estimation problem. Recently, camera devices are rapidly updated. Dual cameras are affordable for users and have been integrated in many devices. This development suggests that we can further improve gaze estimation performance with dual-view gaze estimation. In this paper, we propose a dual-view gaze estimation network (DV-Gaze). DV-Gaze estimates dual-view gaze directions from a pair of images. We first propose a dual-view interactive convolution (DIC) block in DV-Gaze. DIC blocks exchange dual-view information during convolution in multiple feature scales. It fuses dual-view features along epipolar lines and compensates for the original feature with the fused feature. We further propose a dual-view transformer to estimate gaze from dual-view features. Camera poses are encoded to indicate the position information in the transformer. We also consider the geometric relation between dual-view gaze directions and propose a dual-view gaze consistency loss for DV-Gaze. DV-Gaze achieves state-of-the-art performance on ETH-XGaze and EVE datasets. Our experiments also prove the potential of dual-view gaze estimation. We release codes in https://github.com/yihuacheng/DVGaze.

Yihua Cheng, Kuntai Du, Jiayi Yao et al.

As the use of large language models (LLMs) expands rapidly, so does the range of knowledge needed to supplement various LLM queries. Thus, enabling flexible and efficient injection of new knowledge in LLM inference is critical. Three high-level options exist: (i) embedding the knowledge in LLM's weights (i.e., fine-tuning), (ii) including the knowledge as a part of LLM's text input (i.e., in-context learning), or (iii) injecting the KV caches of the new knowledge to LLM during prefill. This paper argues that, although fine-tuning and in-context learning are popular, using KV caches as the medium of knowledge could simultaneously enable more modular management of knowledge injection and more efficient LLM serving with low cost and fast response. To realize these benefits, we envision a Knowledge Delivery Network (KDN), a new system component in LLM services that dynamically optimizes the storage, transfer, and composition of KV cache across LLM engines and other compute and storage resources. We believe that, just like content delivery networks (CDNs), such as Akamai, enabled the success of the Internet ecosystem through their efficient data delivery, KDNs will be critical to the success of LLM applications through their efficient knowledge delivery. We have open-sourced a KDN prototype at https://github.com/LMCache/LMCache.

Shaoting Feng, Yuhan Liu, Hanchen Li et al. · stanford

Reusing KV cache is essential for high efficiency of Large Language Model (LLM) inference systems. With more LLM users, the KV cache footprint can easily exceed GPU memory capacity, so prior work has proposed to either evict KV cache to lower-tier storage devices, or compress KV cache so that more KV cache can be fit in the fast memory. However, prior work misses an important opportunity: jointly optimizing the eviction and compression decisions across all KV caches to minimize average generation latency without hurting quality. We propose EVICPRESS, a KV-cache management system that applies lossy compression and adaptive eviction to KV cache across multiple storage tiers. Specifically, for each KV cache of a context, EVICPRESS considers the effect of compression and eviction of the KV cache on the average generation quality and delay across all contexts as a whole. To achieve this, EVICPRESS proposes a unified utility function that quantifies the effect of quality and delay of the lossy compression or eviction. To this end, EVICPRESS's profiling module periodically updates the utility function scores on all possible eviction-compression configurations for all contexts and places KV caches using a fast heuristic to rearrange KV caches on all storage tiers, with the goal of maximizing the utility function scores on each storage tier. Compared to the baselines that evict KV cache or compress KV cache, EVICPRESS achieves higher KV-cache hit rates on fast devices, i.e., lower delay, while preserving high generation quality by applying conservative compression to contexts that are sensitive to compression errors. Evaluation on 12 datasets and 5 models demonstrates that EVICPRESS achieves up to 2.19x faster time-to-first-token (TTFT) at equivalent generation quality.

Hengfei Wang, Zhongqun Zhang, Yihua Cheng et al.

Face rendering using neural radiance fields (NeRF) is a rapidly developing research area in computer vision. While recent methods primarily focus on controlling facial attributes such as identity and expression, they often overlook the crucial aspect of modeling eyeball rotation, which holds importance for various downstream tasks. In this paper, we aim to learn a face NeRF model that is sensitive to eye movements from multi-view images. We address two key challenges in eye-aware face NeRF learning: how to effectively capture eyeball rotation for training and how to construct a manifold for representing eyeball rotation. To accomplish this, we first fit FLAME, a well-established parametric face model, to the multi-view images considering multi-view consistency. Subsequently, we introduce a new Dynamic Eye-aware NeRF (DeNeRF). DeNeRF transforms 3D points from different views into a canonical space to learn a unified face NeRF model. We design an eye deformation field for the transformation, including rigid transformation, e.g., eyeball rotation, and non-rigid transformation. Through experiments conducted on the ETH-XGaze dataset, we demonstrate that our model is capable of generating high-fidelity images with accurate eyeball rotation and non-rigid periocular deformation, even under novel viewing angles. Furthermore, we show that utilizing the rendered images can effectively enhance gaze estimation performance.

Shijing Wang, Chaoqun Cui, Yaping Huang et al.

Human gaze provides essential cues for interpreting attention, intention, and social interaction in visual scenes, yet gaze understanding remains largely unexplored in current vision-language models (VLMs). While recent VLMs achieve strong scene-level reasoning across a range of visual tasks, there exists no benchmark that systematically evaluates or trains them for gaze interpretation, leaving open the question of whether gaze understanding can emerge from general-purpose vision-language pre-training. To address this gap, we introduce VL4Gaze, the first large-scale benchmark designed to investigate, evaluate, and unlock the potential of VLMs for gaze understanding. VL4Gaze contains 489K automatically generated question-answer pairs across 124K images and formulates gaze understanding as a unified VQA problem through four complementary tasks: (1) gaze object description, (2) gaze direction description, (3) gaze point location, and (4) ambiguous question recognition. We comprehensively evaluate both commercial and open-source VLMs under in-context learning and fine-tuning settings. The results show that even large-scale VLMs struggle to reliably infer gaze semantics and spatial localization without task-specific supervision. In contrast, training on VL4Gaze brings substantial and consistent improvements across all tasks, highlighting the importance of targeted multi-task supervision for developing gaze understanding capabilities in VLMs. We will release the dataset and code to support further research and development in this direction.

Zhongqun Zhang, Hengfei Wang, Ziwei Yu et al.

Modeling the physical contacts between the hand and object is standard for refining inaccurate hand poses and generating novel human grasp in 3D hand-object reconstruction. However, existing methods rely on geometric constraints that cannot be specified or controlled. This paper introduces a novel task of controllable 3D hand-object contact modeling with natural language descriptions. Challenges include i) the complexity of cross-modal modeling from language to contact, and ii) a lack of descriptive text for contact patterns. To address these issues, we propose NL2Contact, a model that generates controllable contacts by leveraging staged diffusion models. Given a language description of the hand and contact, NL2Contact generates realistic and faithful 3D hand-object contacts. To train the model, we build \textit{ContactDescribe}, the first dataset with hand-centered contact descriptions. It contains multi-level and diverse descriptions generated by large language models based on carefully designed prompts (e.g., grasp action, grasp type, contact location, free finger status). We show applications of our model to grasp pose optimization and novel human grasp generation, both based on a textual contact description.

Shijing Wang, Yaping Huang, Chaoqun Cui et al.

Gaze following requires both scene understanding and gaze reasoning to localize the gaze target of an in-scene person. Recently, vision foundation models (VFMs) have demonstrated strong performance on this task, enabling simpler architectures while outperforming prior methods. However, we observe a key limitation of VFM-based approaches: while VFMs substantially improve scene understanding, they contribute little to gaze reasoning. As a result, existing methods often rely on semantically salient objects rather than true gaze cues, leading to degraded performance when targets are not salient. To address this, we propose a novel training mechanism to enhance gaze reasoning in VFMs for gaze following. Our method includes: (1) a head-conditioned local LoRA, which enables localized adaptation to preserve scene token learning while improving head token learning for gaze reasoning; and (2) an out-of-cone penalty, which injects gaze cues into head tokens while aligning them with scene tokens. Experiments on the GazeFollow and VAT datasets demonstrate that our method achieves state-of-the-art performance, with particularly strong improvements when gaze targets are not semantically salient. Our findings offer valuable insights for advancing future gaze following research. We will release the code once the paper is accepted.

Yuqi Hou, Zhongqun Zhang, Nora Horanyi et al.

Gaze following estimates gaze targets of in-scene person by understanding human behavior and scene information. Existing methods usually analyze scene images for gaze following. However, compared with visual images, audio also provides crucial cues for determining human behavior.This suggests that we can further improve gaze following considering audio cues. In this paper, we explore gaze following tasks in conversational scenarios. We propose a novel multi-modal gaze following framework based on our observation ``audiences tend to focus on the speaker''. We first leverage the correlation between audio and lips, and classify speakers and listeners in a scene. We then use the identity information to enhance scene images and propose a gaze candidate estimation network. The network estimates gaze candidates from enhanced scene images and we use MLP to match subjects with candidates as classification tasks. Existing gaze following datasets focus on visual images while ignore audios.To evaluate our method, we collect a conversational dataset, VideoGazeSpeech (VGS), which is the first gaze following dataset including images and audio. Our method significantly outperforms existing methods in VGS datasets. The visualization result also prove the advantage of audio cues in gaze following tasks. Our work will inspire more researches in multi-modal gaze following estimation.

Yihua Cheng, Yuhan Liu, Jiayi Yao et al.

Today's LLM inference systems treat individual engines and queries independently for simplicity, but this causes significant resource inefficiencies. While there are proposals to avoid redundant computation by reusing KV caches across queries and to increase GPU utilization by disaggregating a single query to different engines, their promises cannot be realized without efficiently offloading and communicating KV cache across LLM inference engines and queries. We present LMCache, the first and so far the most efficient open-source KV caching solution, which extracts and stores KV caches generated by modern LLM engines (vLLM and SGLang) and shares the KV caches across engines and queries. LMCache exposes KV caches in the LLM engine interface, effectively transforming LLM engines from individual token processors to a collection of engines with KV cache as the storage and communication medium. In particular, it supports both cache offloading (prefix reuse across queries) and prefill-decode disaggregation (cross-engine cache transfer). LMCache's high performance and wide adoption stem from the following contributions: highly optimized KV cache data movement with performance optimizations including batched data movement operations, compute and I/O pipelining; a modular KV cache connector component, decoupling LMCache from the rapid evolution of inference engines; a first-class control API, such as pinning, lookup, cleanup, movement, and compression, for flexible cache orchestration across GPU, CPU, storage, and network layers. Evaluation shows that combining LMCache with vLLM achieves up to 15x improvement in throughput across diverse workloads. With a growing community, LMCache has seen dramatic growth in adoption by enterprise inference systems, which provides valuable lessons for future KV caching solutions. The source code of LMCache is at: https://github.com/LMCache/LMCache.

Hengfei Wang, Zhongqun Zhang, Yihua Cheng et al.

Gaze redirection methods aim to generate realistic human face images with controllable eye movement. However, recent methods often struggle with 3D consistency, efficiency, or quality, limiting their practical applications. In this work, we propose RTGaze, a real-time and high-quality gaze redirection method. Our approach learns a gaze-controllable facial representation from face images and gaze prompts, then decodes this representation via neural rendering for gaze redirection. Additionally, we distill face geometric priors from a pretrained 3D portrait generator to enhance generation quality. We evaluate RTGaze both qualitatively and quantitatively, demonstrating state-of-the-art performance in efficiency, redirection accuracy, and image quality across multiple datasets. Our system achieves real-time, 3D-aware gaze redirection with a feedforward network (~0.06 sec/image), making it 800x faster than the previous state-of-the-art 3D-aware methods.

Yizhou Huang, Yihua Cheng, Kezhi Wang

Deep learning architectures with powerful reasoning capabilities have driven significant advancements in autonomous driving technology. Large language models (LLMs) applied in this field can describe driving scenes and behaviors with a level of accuracy similar to human perception, particularly in visual tasks. Meanwhile, the rapid development of edge computing, with its advantage of proximity to data sources, has made edge devices increasingly important in autonomous driving. Edge devices process data locally, reducing transmission delays and bandwidth usage, and achieving faster response times. In this work, we propose a driving behavior narration and reasoning framework that applies LLMs to edge devices. The framework consists of multiple roadside units, with LLMs deployed on each unit. These roadside units collect road data and communicate via 5G NSR/NR networks. Our experiments show that LLMs deployed on edge devices can achieve satisfactory response speeds. Additionally, we propose a prompt strategy to enhance the narration and reasoning performance of the system. This strategy integrates multi-modal information, including environmental, agent, and motion data. Experiments conducted on the OpenDV-Youtube dataset demonstrate that our approach significantly improves performance across both tasks.

Yihua Cheng, Feng Lu

Recent work has proven the effectiveness of transformers in many computer vision tasks. However, the performance of transformers in gaze estimation is still unexplored. In this paper, we employ transformers and assess their effectiveness for gaze estimation. We consider two forms of vision transformer which are pure transformers and hybrid transformers. We first follow the popular ViT and employ a pure transformer to estimate gaze from images. On the other hand, we preserve the convolutional layers and integrate CNNs as well as transformers. The transformer serves as a component to complement CNNs. We compare the performance of the two transformers in gaze estimation. The Hybrid transformer significantly outperforms the pure transformer in all evaluation datasets with less parameters. We further conduct experiments to assess the effectiveness of the hybrid transformer and explore the advantage of self-attention mechanism. Experiments show the hybrid transformer can achieve state-of-the-art performance in all benchmarks with pre-training.To facilitate further research, we release codes and models in https://github.com/yihuacheng/GazeTR.

Yihua Cheng, Yiwei Bao, Feng Lu

Gaze estimation methods learn eye gaze from facial features. However, among rich information in the facial image, real gaze-relevant features only correspond to subtle changes in eye region, while other gaze-irrelevant features like illumination, personal appearance and even facial expression may affect the learning in an unexpected way. This is a major reason why existing methods show significant performance degradation in cross-domain/dataset evaluation. In this paper, we tackle the cross-domain problem in gaze estimation. Different from common domain adaption methods, we propose a domain generalization method to improve the cross-domain performance without touching target samples. The domain generalization is realized by gaze feature purification. We eliminate gaze-irrelevant factors such as illumination and identity to improve the cross-domain performance. We design a plug-and-play self-adversarial framework for the gaze feature purification. The framework enhances not only our baseline but also existing gaze estimation methods directly and significantly. To the best of our knowledge, we are the first to propose domain generalization methods in gaze estimation. Our method achieves not only state-of-the-art performance among typical gaze estimation methods but also competitive results among domain adaption methods. The code is released in https://github.com/yihuacheng/PureGaze.

Yizhou Huang, Gengze Jiang, Yihua Cheng et al.

Accurate trajectory prediction is vital for safe autonomous driving, yet existing approaches struggle to balance modeling power and computational efficiency. Attention-based architectures incur quadratic complexity with increasing agents, while recurrent models struggle to capture long-range dependencies and fine-grained local dynamics. Building upon this, we present FoSS, a dual-branch framework that unifies frequency-domain reasoning with linear-time sequence modeling. The frequency-domain branch performs a discrete Fourier transform to decompose trajectories into amplitude components encoding global intent and phase components capturing local variations, followed by a progressive helix reordering module that preserves spectral order; two selective state-space submodules, Coarse2Fine-SSM and SpecEvolve-SSM, refine spectral features with O(N) complexity. In parallel, a time-domain dynamic selective SSM reconstructs self-attention behavior in linear time to retain long-range temporal context. A cross-attention layer fuses temporal and spectral representations, while learnable queries generate multiple candidate trajectories, and a weighted fusion head expresses motion uncertainty. Experiments on Argoverse 1 and Argoverse 2 benchmarks demonstrate that FoSS achieves state-of-the-art accuracy while reducing computation by 22.5% and parameters by over 40%. Comprehensive ablations confirm the necessity of each component.

Yihua Cheng, Yaning Zhu, Zongji Wang et al.

Driver's eye gaze holds a wealth of cognitive and intentional cues crucial for intelligent vehicles. Despite its significance, research on in-vehicle gaze estimation remains limited due to the scarcity of comprehensive and well-annotated datasets in real driving scenarios. In this paper, we present three novel elements to advance in-vehicle gaze research. Firstly, we introduce IVGaze, a pioneering dataset capturing in-vehicle gaze, collected from 125 subjects and covering a large range of gaze and head poses within vehicles. Conventional gaze collection systems are inadequate for in-vehicle use. In this dataset, we propose a new vision-based solution for in-vehicle gaze collection, introducing a refined gaze target calibration method to tackle annotation challenges. Second, our research focuses on in-vehicle gaze estimation leveraging the IVGaze. In-vehicle face images often suffer from low resolution, prompting our introduction of a gaze pyramid transformer that leverages transformer-based multilevel features integration. Expanding upon this, we introduce the dual-stream gaze pyramid transformer (GazeDPTR). Employing perspective transformation, we rotate virtual cameras to normalize images, utilizing camera pose to merge normalized and original images for accurate gaze estimation. GazeDPTR shows state-of-the-art performance on the IVGaze dataset. Thirdly, we explore a novel strategy for gaze zone classification by extending the GazeDPTR. A foundational tri-plane and project gaze onto these planes are newly defined. Leveraging both positional features from the projection points and visual attributes from images, we achieve superior performance compared to relying solely on visual features, substantiating the advantage of gaze estimation. Our project is available at https://yihua.zone/work/ivgaze.

Yuhan Liu, Yuyang Huang, Jiayi Yao et al.

Compound AI systems, such as agentic systems, are an emerging trend in large-scale enterprise settings, with multiple LLMs specialized for different users, tasks, and/or roles working together. In these scenarios, different models often process inputs that share the same context prefix. Although much work was done in the past to enable the reuse of prefix KV caches across inputs for a single model, how to enable one model to reuse the prefix KV caches of a different model remains an open question. We introduce DroidSpeak, the first distributed LLM inference system that enables KV cache reuse across distributed nodes running inference of different LLMs, so long as the LLMs have the same architecture. We present the first study that aims at understanding the impact of sharing KV caches across different LLMs, and if/when such sharing affects quality. Inspired by the findings, we present DroidSpeak, which selectively recomputes a few layers of the KV cache produced by another LLM and reuses the remaining layers, with negligible quality loss. Moreover, carefully pipelining the layer-wise re-computation and the loading of reused KV cache further improves the inference performance. Experiments on diverse datasets and model pairs demonstrate that DroidSpeak achieves up to 4x throughput improvement and about 3.1x faster prefill (time to first token), with negligible loss of quality in F1 scores, Rouge-L or code similarity score, compared to the baseline which does not allow any sharing across models.

Hanchen Li, Yuhan Liu, Yihua Cheng et al.

To render each generated token in real-time for users, the Large Language Model (LLM) server generates tokens one by one and streams each token (or group of a few tokens) through the network to the user right after generation, which we refer to as LLM token streaming. However, under unstable network conditions, the LLM token streaming experience could suffer greatly from stalls since one packet loss could block the rendering of later tokens even if the packets containing them arrive on time. With a measurement study, we show that current applications suffer from increased stalls under unstable networks. For this emerging token streaming problem in LLM Chatbots that differs from previous multimedia and text applications, we propose a novel transmission scheme, called Eloquent, which puts newly generated tokens as well as currently unacknowledged tokens in the next outgoing packet. This ensures that each packet contains some new tokens and, in the meantime, is independently rendered when received, avoiding the aforementioned stalls caused by missing packets. Through simulation under various networks, we show Eloquent reduces stall ratio (proportion of token rendering wait time) by 71.0% compared to the retransmission method commonly used by real chatbot applications and by 31.6% compared to the baseline packet duplication scheme. By tailoring Eloquent to fit the token-by-token generation of LLM, we enable the Chatbots to respond like an eloquent speaker for users to better enjoy pervasive AI.

Yuwei An, Yihua Cheng, Seo Jin Park et al.

Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm for enhancing the performance of large language models (LLMs) by integrating external knowledge into the generation process. A key component of RAG pipelines is the reranker, which selects the most relevant documents from a pool of retrieved candidates and significantly improves the quality of the generated responses. While rerankers refine the selection of retrieved documents in RAG pipelines, they introduce computational challenges that hinder high throughput and low latency. To address this problem, we propose HyperRAG, a system that optimizes the trade-off between quality and efficiency in RAG pipelines by leveraging KV-cache reuse for efficient reranker inference. By reusing document-side KV-cache, HyperRAG achieves both high-quality generation and system-level efficiency. To fully realize the benefits of KV-cache reuse, HyperRAG incorporates a range of system-level optimizations designed to enhance efficiency and scalability. Experiments show that HyperRAG achieves a 2 - 3 throughput improvement with decoder-only rerankers while also delivering higher downstream performance compared with traditional RAG service.

Yizhou Huang, Yihua Cheng, Kezhi Wang

Motion prediction is crucial for autonomous driving, as it enables accurate forecasting of future vehicle trajectories based on historical inputs. This paper introduces Trajectory Mamba, a novel efficient trajectory prediction framework based on the selective state-space model (SSM). Conventional attention-based models face the challenge of computational costs that grow quadratically with the number of targets, hindering their application in highly dynamic environments. In response, we leverage the SSM to redesign the self-attention mechanism in the encoder-decoder architecture, thereby achieving linear time complexity. To address the potential reduction in prediction accuracy resulting from modifications to the attention mechanism, we propose a joint polyline encoding strategy to better capture the associations between static and dynamic contexts, ultimately enhancing prediction accuracy. Additionally, to balance prediction accuracy and inference speed, we adopted the decoder that differs entirely from the encoder. Through cross-state space attention, all target agents share the scene context, allowing the SSM to interact with the shared scene representation during decoding, thus inferring different trajectories over the next prediction steps. Our model achieves state-of-the-art results in terms of inference speed and parameter efficiency on both the Argoverse 1 and Argoverse 2 datasets. It demonstrates a four-fold reduction in FLOPs compared to existing methods and reduces parameter count by over 40% while surpassing the performance of the vast majority of previous methods. These findings validate the effectiveness of Trajectory Mamba in trajectory prediction tasks.

Hengfei Wang, Zhongqun Zhang, Yihua Cheng et al.

Generating face image with specific gaze information has attracted considerable attention. Existing approaches typically input gaze values directly for face generation, which is unnatural and requires annotated gaze datasets for training, thereby limiting its application. In this paper, we present a novel gaze-controllable face generation task. Our approach inputs textual descriptions that describe human gaze and head behavior and generates corresponding face images. Our work first introduces a text-of-gaze dataset containing over 90k text descriptions spanning a dense distribution of gaze and head poses. We further propose a gaze-controllable text-to-face method. Our method contains a sketch-conditioned face diffusion module and a model-based sketch diffusion module. We define a face sketch based on facial landmarks and eye segmentation map. The face diffusion module generates face images from the face sketch, and the sketch diffusion module employs a 3D face model to generate face sketch from text description. Experiments on the FFHQ dataset show the effectiveness of our method. We will release our dataset and code for future research.

Yihua Cheng, Hengfei Wang, Zhongqun Zhang et al.

3D and 2D gaze estimation share the fundamental objective of capturing eye movements but are traditionally treated as two distinct research domains. In this paper, we introduce a novel cross-task few-shot 2D gaze estimation approach, aiming to adapt a pre-trained 3D gaze estimation network for 2D gaze prediction on unseen devices using only a few training images. This task is highly challenging due to the domain gap between 3D and 2D gaze, unknown screen poses, and limited training data. To address these challenges, we propose a novel framework that bridges the gap between 3D and 2D gaze. Our framework contains a physics-based differentiable projection module with learnable parameters to model screen poses and project 3D gaze into 2D gaze. The framework is fully differentiable and can integrate into existing 3D gaze networks without modifying their original architecture. Additionally, we introduce a dynamic pseudo-labelling strategy for flipped images, which is particularly challenging for 2D labels due to unknown screen poses. To overcome this, we reverse the projection process by converting 2D labels to 3D space, where flipping is performed. Notably, this 3D space is not aligned with the camera coordinate system, so we learn a dynamic transformation matrix to compensate for this misalignment. We evaluate our method on MPIIGaze, EVE, and GazeCapture datasets, collected respectively on laptops, desktop computers, and mobile devices. The superior performance highlights the effectiveness of our approach, and demonstrates its strong potential for real-world applications.

Shaoting Feng, Hanchen Li, Kuntai Du et al.

Large language model (LLM) applications often reuse previously processed context, such as chat history and documents, which introduces significant redundant computation. Existing LLM serving systems address such redundant computation by storing the KV caches of processed context and loading the corresponding KV cache when a new request reuses the context. Further, as these LLM applications scale, the total size of KV caches becomes excessively large and requires both DRAM and SSD for full storage. However, prior work that stores KV caches in DRAM and SSD suffers from high loading delays, as most KV cache hits come from SSD, which is slow to load. To increase the KV cache hit rate on DRAM, we identify lossy KV cache compression as a promising approach. We design a lossy compression system that decides the compression algorithm, compression rate and device placement for each KV cache entry to maximise DRAM hits and minimise loading delay without significantly degrading generation quality. Compared to various static compression baselines across three tasks, our system AdaptCache achieves 1.43--2.4 x delay savings at the same quality and 6--55% quality improvements at the same delay.

YoungChan Choi, HengFei Wang, YiHua Cheng et al.

We propose a novel 3D gaze redirection framework that leverages an explicit 3D eyeball structure. Existing gaze redirection methods are typically based on neural radiance fields, which employ implicit neural representations via volume rendering. Unlike these NeRF-based approaches, where the rotation and translation of 3D representations are not explicitly modeled, we introduce a dedicated 3D eyeball structure to represent the eyeballs with 3D Gaussian Splatting (3DGS). Our method generates photorealistic images that faithfully reproduce the desired gaze direction by explicitly rotating and translating the 3D eyeball structure. In addition, we propose an adaptive deformation module that enables the replication of subtle muscle movements around the eyes. Through experiments conducted on the ETH-XGaze dataset, we demonstrate that our framework is capable of generating diverse novel gaze images, achieving superior image quality and gaze estimation accuracy compared to previous state-of-the-art methods.

Zheng Cheng, Wenri Wang, Guangyong Chen et al.

Underwater image enhancement (UIE) techniques aim to improve visual quality of images captured in aquatic environments by addressing degradation issues caused by light absorption and scattering effects, including color distortion, blurring, and low contrast. Current mainstream solutions predominantly employ multi-scale feature extraction (MSFE) mechanisms to enhance reconstruction quality through multi-resolution feature fusion. However, our extensive experiments demonstrate that high-quality image reconstruction does not necessarily rely on multi-scale feature fusion. Contrary to popular belief, our experiments show that single-scale feature extraction alone can match or surpass the performance of multi-scale methods, significantly reducing complexity. To comprehensively explore single-scale feature potential in underwater enhancement, we propose an innovative Single-Scale Decomposition Network (SSD-Net). This architecture introduces an asymmetrical decomposition mechanism that disentangles input image into clean layer along with degradation layer. The former contains scene-intrinsic information and the latter encodes medium-induced interference. It uniquely combines CNN's local feature extraction capabilities with Transformer's global modeling strengths through two core modules: 1) Parallel Feature Decomposition Block (PFDB), implementing dual-branch feature space decoupling via efficient attention operations and adaptive sparse transformer; 2) Bidirectional Feature Communication Block (BFCB), enabling cross-layer residual interactions for complementary feature mining and fusion. This synergistic design preserves feature decomposition independence while establishing dynamic cross-layer information pathways, effectively enhancing degradation decoupling capacity.

Yizhou Huang, Yihua Cheng, Kezhi Wang

Understanding human motion is crucial for accurate pedestrian trajectory prediction. Conventional methods typically rely on supervised learning, where ground-truth labels are directly optimized against predicted trajectories. This amplifies the limitations caused by long-tailed data distributions, making it difficult for the model to capture abnormal behaviors. In this work, we propose a self-supervised pedestrian trajectory prediction framework that explicitly models position, velocity, and acceleration. We leverage velocity and acceleration information to enhance position prediction through feature injection and a self-supervised motion consistency mechanism. Our model hierarchically injects velocity features into the position stream. Acceleration features are injected into the velocity stream. This enables the model to predict position, velocity, and acceleration jointly. From the predicted position, we compute corresponding pseudo velocity and acceleration, allowing the model to learn from data-generated pseudo labels and thus achieve self-supervised learning. We further design a motion consistency evaluation strategy grounded in physical principles; it selects the most reasonable predicted motion trend by comparing it with historical dynamics and uses this trend to guide and constrain trajectory generation. We conduct experiments on the ETH-UCY and Stanford Drone datasets, demonstrating that our method achieves state-of-the-art performance on both datasets.

Boeun Kim, Hea In Jeong, JungHoon Sung et al.

This paper introduces Motion Personalization, a new task that generates personalized motions aligned with text descriptions using several basic motions containing Persona. To support this novel task, we introduce a new large-scale motion dataset called PerMo (PersonaMotion), which captures the unique personas of multiple actors. We also propose a multi-modal finetuning method of a pretrained motion diffusion model called PersonaBooth. PersonaBooth addresses two main challenges: i) A significant distribution gap between the persona-focused PerMo dataset and the pretraining datasets, which lack persona-specific data, and ii) the difficulty of capturing a consistent persona from the motions vary in content (action type). To tackle the dataset distribution gap, we introduce a persona token to accept new persona features and perform multi-modal adaptation for both text and visuals during finetuning. To capture a consistent persona, we incorporate a contrastive learning technique to enhance intra-cohesion among samples with the same persona. Furthermore, we introduce a context-aware fusion mechanism to maximize the integration of persona cues from multiple input motions. PersonaBooth outperforms state-of-the-art motion style transfer methods, establishing a new benchmark for motion personalization.

Yihua Cheng, Ziyi Zhang, Hanchen Li et al.

In real-time video communication, retransmitting lost packets over high-latency networks is not viable due to strict latency requirements. To counter packet losses without retransmission, two primary strategies are employed -- encoder-based forward error correction (FEC) and decoder-based error concealment. The former encodes data with redundancy before transmission, yet determining the optimal redundancy level in advance proves challenging. The latter reconstructs video from partially received frames, but dividing a frame into independently coded partitions inherently compromises compression efficiency, and the lost information cannot be effectively recovered by the decoder without adapting the encoder. We present a loss-resilient real-time video system called GRACE, which preserves the user's quality of experience (QoE) across a wide range of packet losses through a new neural video codec. Central to GRACE's enhanced loss resilience is its joint training of the neural encoder and decoder under a spectrum of simulated packet losses. In lossless scenarios, GRACE achieves video quality on par with conventional codecs (e.g., H.265). As the loss rate escalates, GRACE exhibits a more graceful, less pronounced decline in quality, consistently outperforming other loss-resilient schemes. Through extensive evaluation on various videos and real network traces, we demonstrate that GRACE reduces undecodable frames by 95% and stall duration by 90% compared with FEC, while markedly boosting video quality over error concealment methods. In a user study with 240 crowdsourced participants and 960 subjective ratings, GRACE registers a 38% higher mean opinion score (MOS) than other baselines.

Yihua Cheng, Haofei Wang, Yiwei Bao et al.

Human gaze provides valuable information on human focus and intentions, making it a crucial area of research. Recently, deep learning has revolutionized appearance-based gaze estimation. However, due to the unique features of gaze estimation research, such as the unfair comparison between 2D gaze positions and 3D gaze vectors and the different pre-processing and post-processing methods, there is a lack of a definitive guideline for developing deep learning-based gaze estimation algorithms. In this paper, we present a systematic review of the appearance-based gaze estimation methods using deep learning. Firstly, we survey the existing gaze estimation algorithms along the typical gaze estimation pipeline: deep feature extraction, deep learning model design, personal calibration and platforms. Secondly, to fairly compare the performance of different approaches, we summarize the data pre-processing and post-processing methods, including face/eye detection, data rectification, 2D/3D gaze conversion and gaze origin conversion. Finally, we set up a comprehensive benchmark for deep learning-based gaze estimation. We characterize all the public datasets and provide the source code of typical gaze estimation algorithms. This paper serves not only as a reference to develop deep learning-based gaze estimation methods, but also a guideline for future gaze estimation research. The project web page can be found at https://phi-ai.buaa.edu.cn/Gazehub.

Yiwei Bao, Yihua Cheng, Yunfei Liu et al.

Recently, many multi-stream gaze estimation methods have been proposed. They estimate gaze from eye and face appearances and achieve reasonable accuracy. However, most of the methods simply concatenate the features extracted from eye and face appearance. The feature fusion process has been ignored. In this paper, we propose a novel Adaptive Feature Fusion Network (AFF-Net), which performs gaze tracking task in mobile tablets. We stack two-eye feature maps and utilize Squeeze-and-Excitation layers to adaptively fuse two-eye features according to their similarity on appearance. Meanwhile, we also propose Adaptive Group Normalization to recalibrate eye features with the guidance of facial feature. Extensive experiments on both GazeCapture and MPIIFaceGaze datasets demonstrate consistently superior performance of the proposed method.

Yihua Cheng, Shiyao Huang, Fei Wang et al.

Human gaze is essential for various appealing applications. Aiming at more accurate gaze estimation, a series of recent works propose to utilize face and eye images simultaneously. Nevertheless, face and eye images only serve as independent or parallel feature sources in those works, the intrinsic correlation between their features is overlooked. In this paper we make the following contributions: 1) We propose a coarse-to-fine strategy which estimates a basic gaze direction from face image and refines it with corresponding residual predicted from eye images. 2) Guided by the proposed strategy, we design a framework which introduces a bi-gram model to bridge gaze residual and basic gaze direction, and an attention component to adaptively acquire suitable fine-grained feature. 3) Integrating the above innovations, we construct a coarse-to-fine adaptive network named CA-Net and achieve state-of-the-art performances on MPIIGaze and EyeDiap.