Zeyu Chen

AAAI Conference 2026 Conference Paper

Sortblock: Similarity-Aware Feature Reuse for Diffusion Model

• Hanqi Chen
• Xu Zhang
• Xiaoliu Guan
• Lielin Jiang
• Guanzhong Wang
• Zeyu Chen
• Yi Liu

Diffusion Transformers (DiTs) have demonstrated remarkable generative capabilities, particularly benefiting from Transformer architectures that enhance visual and artistic fidelity. However, their inherently sequential denoising process results in high inference latency, limiting their deployment in real-time scenarios. Existing training-free acceleration approaches typically reuse intermediate features at fixed timesteps or layers, overlooking the evolving semantic focus across denoising stages and Transformer blocks.To address this, we propose Sortblock, a training-free inference acceleration framework that dynamically caches block-wise features based on their similarity across adjacent timesteps. By ranking the evolution of residuals, Sortblock adaptively determines a recomputation ratio, selectively skipping redundant computations while preserving generation quality. Furthermore, we incorporate a lightweight linear prediction mechanism to reduce accumulated errors in skipped blocks.Extensive experiments across various tasks and DiT architectures demonstrate that Sortblock achieves over 2 times inference speedup with minimal degradation in output quality, offering an effective and generalizable solution for accelerating diffusion-based generative models.

IROS Conference 2025 Conference Paper

3D Gaussian Splatting for Fine-Detailed Surface Reconstruction in Large-Scale Scene

• Shihan Chen
• Zhaojin Li
• Zeyu Chen
• Qingsong Yan
• Gaoyang Shen
• Ran Duan

Recent developments in 3D Gaussian Splatting have made significant advances in surface reconstruction. However, scaling these methods to large-scale scenes remains challenging due to high computational demands and the complex dynamic appearances typical of outdoor environments. These challenges hinder the application in aerial surveying and autonomous driving. This paper proposes a novel solution to reconstruct large-scale surfaces with fine details, supervised by full-sized images. Firstly, we introduce a coarse-to-fine strategy to reconstruct a coarse model efficiently, followed by adaptive scene partitioning and sub-scene refining from image segments. Additionally, we integrate a decoupling appearance model to capture global appearance variations and a transient mask model to mitigate interference from moving objects. Finally, we expand the multi-view constraint and introduce a single-view regularization for texture-less areas. Our experiments were conducted on the publicly available dataset GauU-Scene V2, which was captured using unmanned aerial vehicles. To the best of our knowledge, our method outperforms existing NeRF-based and Gaussian-based methods, achieving high-fidelity visual results and accurate surface from full-size image optimization. Open-source code will be available on GitHub.

ICLR Conference 2025 Conference Paper

FlashMask: Efficient and Rich Mask Extension of FlashAttention

• Guoxia Wang
• Jinle Zeng
• Xiyuan Xiao
• Siming Wu
• Jiabin Yang
• Lujing Zheng
• Zeyu Chen
• Jiang Bian

The computational and memory demands of vanilla attention scale quadratically with the sequence length $N$, posing significant challenges for processing long sequences in Transformer models. FlashAttention alleviates these challenges by eliminating the $\mathcal{O}(N^2)$ memory dependency and reducing attention latency through IO-aware memory optimizations. However, its native support for certain attention mask types is limited, and it does not inherently accommodate more complex masking requirements. Previous approaches resort to using dense masks with $\mathcal{O}(N^2)$ memory complexity, leading to inefficiencies. In this paper, we propose \ours{}, an extension of FlashAttention that introduces a column-wise sparse representation of attention masks. This approach efficiently represents a wide range of mask types and facilitates the development of optimized kernel implementations. By adopting this novel representation, \ours{} achieves linear memory complexity $\mathcal{O}(N)$, making it suitable for modeling long-context sequences. Moreover, this representation enables kernel optimizations that eliminate unnecessary computations by leveraging sparsity in the attention mask, without sacrificing computational accuracy, resulting in higher computational efficiency. We evaluate \ours{}'s performance in fine-tuning and alignment training of LLMs such as SFT, LoRA, DPO, and RM. \ours{} achieves significant throughput improvements, with end-to-end speedups ranging from 1.65x to 3.22x compared to existing FlashAttention dense method. Additionally, our kernel-level comparisons demonstrate that \ours{} surpasses the latest counterpart, FlexAttention, by 12.1\% to 60.7\% in terms of kernel TFLOPs/s, achieving 37.8\% to 62.3\% of the theoretical maximum FLOPs/s on the A100 GPU. The code is open-sourced on PaddlePaddle\footnote{\url{https://github.com/PaddlePaddle/Paddle}} and integrated into PaddleNLP\footnote{\url{https://github.com/PaddlePaddle/PaddleNLP}}, supporting models with over 100 billion parameters for contexts extending up to 128K tokens.

IROS Conference 2025 Conference Paper

KDMOS: Knowledge Distillation for Motion Segmentation

• Chunyu Cao
• Jintao Cheng
• Zeyu Chen
• Linfan Zhan
• Rui Fan 0001
• Zhijian He
• Xiaoyu Tang

Motion Object Segmentation (MOS) is crucial for autonomous driving, as it enhances localization, path planning, map construction, scene flow estimation, and future state prediction. While existing methods achieve strong performance, balancing accuracy and real-time inference remains a challenge. To address this, we propose a logits-based knowledge distillation framework for MOS, aiming to improve accuracy while maintaining real-time efficiency. Specifically, we adopt a Bird’s Eye View (BEV) projection-based model as the student and a non-projection model as the teacher. To handle the severe imbalance between moving and non-moving classes, we decouple them and apply tailored distillation strategies, allowing the teacher model to better learn key motion-related features. This approach significantly reduces false positives and false negatives. Additionally, we introduce dynamic upsampling, optimize the network architecture, and achieve a 7. 69% reduction in parameter count, mitigating overfitting. Our method achieves a notable IoU of 78. 8% on the hidden test set of the SemanticKITTI-MOS dataset and delivers competitive results on the Apollo dataset. The KDMOS implementation is available at https://github.com/SCNU-RISLAB/KDMOS.

NeurIPS Conference 2025 Conference Paper

Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation

• Yifu Luo
• Xinhao Hu
• Keyu Fan
• Haoyuan Sun
• Zeyu Chen
• Bo Xia
• Tiantian Zhang
• Yongzhe Chang

Reinforcement learning (RL) has garnered increasing attention in text-to-image (T2I) generation. However, most existing RL approaches are tailored to either diffusion models or autoregressive models, overlooking an important alternative: masked generative models. In this work, we propose Mask-GRPO, the first method to incorporate Group Relative Policy Optimization (GRPO)-based RL into this overlooked paradigm. Our core insight is to redefine the transition probability, which is different from current approaches, and formulate the unmasking process as a multi-step decision-making problem. To further enhance our method, we explore several useful strategies, including removing the Kullback–Leibler constraint, applying the reduction strategy, and filtering out low-quality samples. Using Mask-GRPO, we improve a base model, Show-o, with substantial improvements on standard T2I benchmarks and preference alignment, outperforming existing state-of-the-art approaches.

IROS Conference 2024 Conference Paper

A Point-Line Features Fusion Method for Fast and Robust Monocular Visual-Inertial Initialization

• Guoqiang Xie
• Jie Chen
• Tianhang Tang
• Zeyu Chen
• Ling Lei
• Yiguang Liu

Fast and robust initialization is essential for highly accurate monocular visual-inertial odometer (VIO), but at present majority of initialization methods rely only on point features, unstable in low texture and blurring situations. Therefore, we propose a novel point-line features fusion method for monocular visual-inertial initialization, as line features are more stable and provide richer geometric information than point features: 1) a closed-form line features initialization method is presented, and combined with point features to obtain a more integrated and robust linear system; 2) a monocular depth network is adopted to provide learned affine-invariant depth map, requiring only one prior depth map for the first frame, which can improve performance under low-parallax scenarios; 3) we can easily use RANSAC to reject outliers in solving linear system based on our formulation. Moreover, line feature re-projection residual is added to visual-inertial bundle adjustment (VI-BA) to obtain more accurate initial parameters. The proposed method is more accurate and robust than state-of-the-art methods due to the line features, especially under extreme low-parallax scenarios, and extensive experiments on popular datasets have confirmed, 0. 5s initialization window on EuRoC MAV, 0. 3s initialization window on TUM-VI, while the standard method normally waits for a window of 2s.

AAAI Conference 2024 Conference Paper

G–LIME: Statistical Learning for Local Interpretations of Deep Neural Networks Using Global Priors (Abstract Reprint)

• Xuhong Li
• Haoyi Xiong
• Xingjian Li
• Xiao Zhang
• Ji Liu
• Haiyan Jiang
• Zeyu Chen
• Dejing Dou

To explain the prediction result of a Deep Neural Network (DNN) model based on a given sample, LIME [1] and its derivatives have been proposed to approximate the local behavior of the DNN model around the data point via linear surrogates. Though these algorithms interpret the DNN by finding the key features used for classification, the random interpolations used by LIME would perturb the explanation result and cause the instability and inconsistency between repetitions of LIME computations. To tackle this issue, we propose G-LIME that extends the vanilla LIME through high-dimensional Bayesian linear regression using the sparsity and informative global priors. Specifically, with a dataset representing the population of samples (e.g., the training set), G-LIME first pursues the global explanation of the DNN model using the whole dataset. Then, with a new data point, -LIME incorporates an modified estimator of ElasticNet-alike to refine the local explanation result through balancing the distance to the global explanation and the sparsity/feature selection in the explanation. Finally, G-LIME uses Least Angle Regression (LARS) and retrieves the solution path of a modified ElasticNet under varying -regularization, to screen and rank the importance of features [2] as the explanation result. Through extensive experiments on real world tasks, we show that the proposed method yields more stable, consistent, and accurate results compared to LIME.

IROS Conference 2024 Conference Paper

UW-SDF: Exploiting Hybrid Geometric Priors for Neural SDF Reconstruction from Underwater Multi-view Monocular Images

• Zeyu Chen
• Jingyi Tang
• Gu Wang 0001
• Shengquan Li
• Xinghui Li
• Xiangyang Ji
• Xiu Li 0001

Due to the unique characteristics of underwater environments, accurate 3D reconstruction of underwater objects poses a challenging problem in tasks such as underwater exploration and mapping. Traditional methods that rely on multiple sensor data for 3D reconstruction are time-consuming and face challenges in data acquisition in underwater scenarios. We propose UW-SDF, a framework for reconstructing target objects from multi-view underwater images based on neural SDF. We introduce hybrid geometric priors to optimize the reconstruction process, markedly enhancing the quality and efficiency of neural SDF reconstruction. Additionally, to address the challenge of segmentation consistency in multi-view images, we propose a novel few-shot multi-view target segmentation strategy using the general-purpose segmentation model (SAM), enabling rapid automatic segmentation of unseen objects. Through extensive qualitative and quantitative experiments on diverse datasets, we demonstrate that our proposed method outperforms the traditional underwater 3D reconstruction method and other neural rendering approaches in the field of underwater 3D reconstruction.

AIJ Journal 2023 Journal Article

G -LIME: Statistical learning for local interpretations of deep neural networks using global priors

• Xuhong Li
• Haoyi Xiong
• Xingjian Li
• Xiao Zhang
• Ji Liu
• Haiyan Jiang
• Zeyu Chen
• Dejing Dou

JMLR Journal 2022 Journal Article

InterpretDL: Explaining Deep Models in PaddlePaddle

• Xuhong Li
• Haoyi Xiong
• Xingjian Li
• Xuanyu Wu
• Zeyu Chen
• Dejing Dou

Techniques to explain the predictions of deep neural networks (DNNs) have been largely required for gaining insights into the black boxes. We introduce InterpretDL, a toolkit of explanation algorithms based on PaddlePaddle, with uniformed programming interfaces and "plug-and-play" designs. A few lines of codes are needed to obtain the explanation results without modifying the structure of the model. InterpretDL currently contains 16 algorithms, explaining training phases, datasets, global and local behaviors of post-trained deep models. InterpretDL also provides a number of tutorial examples and showcases to demonstrate the capability of InterpretDL working on a wide range of deep learning models, e.g., Convolutional Neural Networks (CNNs), Multi-Layer Preceptors (MLPs), Transformers, etc., for various tasks in both Computer Vision (CV) and Natural Language Processing (NLP). Furthermore, InterpretDL modularizes the implementations, making efforts to support the compatibility across frameworks. The project is available at https://github.com/PaddlePaddle/InterpretDL. [abs] [ pdf ][ bib ] [ code ] &copy JMLR 2022. ( edit, beta )

IJCAI Conference 2022 Conference Paper

Simple and Effective Relation-based Embedding Propagation for Knowledge Representation Learning

• Huijuan Wang
• Siming Dai
• Weiyue Su
• Hui Zhong
• Zeyang Fang
• Zhengjie Huang
• Shikun Feng
• Zeyu Chen

Relational graph neural networks have garnered particular attention to encode graph context in knowledge graphs (KGs). Although they achieved competitive performance on small KGs, how to efficiently and effectively utilize graph context for large KGs remains an open problem. To this end, we propose the Relation-based Embedding Propagation (REP) method. It is a post-processing technique to adapt pre-trained KG embeddings with graph context. As relations in KGs are directional, we model the incoming head context and the outgoing tail context separately. Accordingly, we design relational context functions with no external parameters. Besides, we use averaging to aggregate context information, making REP more computation-efficient. We theoretically prove that such designs can avoid information distortion during propagation. Extensive experiments also demonstrate that REP has significant scalability while improving or maintaining prediction quality. Particularly, it averagely brings about 10% relative improvement to triplet-based embedding methods on OGBL-WikiKG2 and takes 5%-83% time to achieve comparable results as the state-of-the-art GC-OTE.

TIST Journal 2021 Journal Article

“In-Network Ensemble”: Deep Ensemble Learning with Diversified Knowledge Distillation

• Xingjian Li
• Haoyi Xiong
• Zeyu Chen
• Jun Huan
• Cheng-Zhong Xu
• Dejing Dou

Ensemble learning is a widely used technique to train deep convolutional neural networks (CNNs) for improved robustness and accuracy. While existing algorithms usually first train multiple diversified networks and then assemble these networks as an aggregated classifier, we propose a novel learning paradigm, namely, “In-Network Ensemble” ( INE ) that incorporates the diversity of multiple models through training a SINGLE deep neural network. Specifically, INE segments the outputs of the CNN into multiple independent classifiers, where each classifier is further fine-tuned with better accuracy through a so-called diversified knowledge distillation process. We then aggregate the fine-tuned independent classifiers using an Averaging-and-Softmax operator to obtain the final ensemble classifier. Note that, in the supervised learning settings, INE starts the CNN training from random, while, under the transfer learning settings, it also could start with a pre-trained model to incorporate the knowledge learned from additional datasets. Extensive experiments have been done using eight large-scale real-world datasets, including CIFAR, ImageNet, and Stanford Cars, among others, as well as common deep network architectures such as VGG, ResNet, and Wide ResNet. We have evaluated the method under two tasks: supervised learning and transfer learning. The results show that INE outperforms the state-of-the-art algorithms for deep ensemble learning with improved accuracy.