Author name cluster

Zejian Li

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

8 papers

2 author rows

AAAI Conference 2026 Conference Paper

Diffusion Distillation with Direct Preference Optimization for Efficient 3D LiDAR Scene Completion

An Zhao
Shengyuan Zhang
Zejian Li
Ling Yang
Pei Chen
Jiale Wu
Haoran Xu
AnYang Wei

The slow sampling speed of diffusion models hinders their application in 3D LiDAR scene completion. To address this, we propose Distillation-DPO, a novel framework that accelerates sampling through score distillation while simultaneously enhancing generation quality via preference alignment. Distillation-DPO follows a three-step procedure. First, the student model generates paired completion scenes with different initial noises. Second, using LiDAR scene evaluation metrics as preference, we construct winning and losing sample pairs. Third, as our core innovation, Distillation-DPO optimizes the student model by exploiting the difference in score functions between the teacher and student models on the paired completion scenes. This operation performs variational score distillation of the student model but simultaneously encourages the distilled student to prefer the winning samples over the losing ones. Extensive experiments demonstrate that Distillation-DPO achieves higher-quality scene completion than state-of-the-art diffusion models, while accelerating sampling by over 5-fold. To our knowledge, our work is the first to integrate the preference learning principle of DPO into the distillation of diffusion models, offering a new framework of preference-aligned distillation.

PDF Details DOI

ICLR Conference 2025 Conference Paper

Distribution Backtracking Builds A Faster Convergence Trajectory for Diffusion Distillation

Shengyuan Zhang
Ling Yang 0006
Zejian Li
An Zhao
Chenye Meng
Changyuan Yang
Guang Yang 0022
Zhiyuan Yang

Accelerating the sampling speed of diffusion models remains a significant challenge. Recent score distillation methods distill a heavy teacher model into a student generator to achieve one-step generation, which is optimized by calculating the difference between two score functions on the samples generated by the student model. However, there is a score mismatch issue in the early stage of the score distillation process, since existing methods mainly focus on using the endpoint of pre-trained diffusion models as teacher models, overlooking the importance of the convergence trajectory between the student generator and the teacher model. To address this issue, we extend the score distillation process by introducing the entire convergence trajectory of the teacher model and propose $\textbf{Dis}$tribution $\textbf{Back}$tracking Distillation ($\textbf{DisBack}$). DisBask is composed of two stages: $\textit{Degradation Recording}$ and $\textit{Distribution Backtracking}$. $\textit{Degradation Recording}$ is designed to obtain the convergence trajectory by recording the degradation path from the pre-trained teacher model to the untrained student generator. The degradation path implicitly represents the intermediate distributions between the teacher and the student, and its reverse can be viewed as the convergence trajectory from the student generator to the teacher model. Then $\textit{Distribution Backtracking}$ trains the student generator to backtrack the intermediate distributions along the path to approximate the convergence trajectory of the teacher model. Extensive experiments show that DisBack achieves faster and better convergence than the existing distillation method and achieves comparable or better generation performance, with an FID score of 1.38 on the ImageNet 64$\times$64 dataset. DisBack is easy to implement and can be generalized to existing distillation methods to boost performance.

Details

AAAI Conference 2024 Conference Paper

Reducing Spatial Fitting Error in Distillation of Denoising Diffusion Models

Shengzhe Zhou
Zejian Li
Shengyuan Zhang
Lefan Hou
Changyuan Yang
Guang Yang
Zhiyuan Yang
Lingyun Sun

Denoising Diffusion models have exhibited remarkable capabilities in image generation. However, generating high-quality samples requires a large number of iterations. Knowledge distillation for diffusion models is an effective method to address this limitation with a shortened sampling process but causes degraded generative quality. Based on our analysis with bias-variance decomposition and experimental observations, we attribute the degradation to the spatial fitting error occurring in the training of both the teacher and student model in the distillation. Accordingly, we propose Spatial Fitting-Error Reduction Distillation model (SFERD). SFERD utilizes attention guidance from the teacher model and a designed semantic gradient predictor to reduce the student's fitting error. Empirically, our proposed model facilitates high-quality sample generation in a few function evaluations. We achieve an FID of 5.31 on CIFAR-10 and 9.39 on ImageNet 64x64 with only one step, outperforming existing diffusion methods. Our study provides a new perspective on diffusion distillation by highlighting the intrinsic denoising ability of models.

PDF Details DOI

IJCAI Conference 2023 Conference Paper

Learning Object Consistency and Interaction in Image Generation from Scene Graphs

Yangkang Zhang
Chenye Meng
Zejian Li
Pei Chen
Guang Yang
Changyuan Yang
Lingyun Sun

This paper is concerned with synthesizing images conditioned on a scene graph (SG), a set of object nodes and their edges of interactive relations. We divide existing works into image-oriented and code-oriented methods. In our analysis, the image-oriented methods do not consider object interaction in spatial hidden feature. On the other hand, in empirical study, the code-oriented methods lose object consistency as their generated images miss certain objects in the input scene graph. To alleviate these two issues, we propose Learning Object Consistency and Interaction (LOCI). To preserve object consistency, we design a consistency module with a weighted augmentation strategy for objects easy to be ignored and a matching loss between scene graphs and image codes. To learn object interaction, we design an interaction module consisting of three kinds of message propagation between the input scene graph and the learned image code. Experiments on COCO-stuff and Visual Genome datasets show our proposed method alleviates the ignorance of objects and outperforms the state-of-the-art on visual fidelity of generated images and objects.

PDF Details DOI

AAAI Conference 2023 Conference Paper

Preserving Structural Consistency in Arbitrary Artist and Artwork Style Transfer

Jingyu Wu
Lefan Hou
Zejian Li
Jun Liao
Li Liu
Lingyun Sun

Deep generative models are effective in style transfer. Previous methods learn one or several specific artist-style from a collection of artworks. These methods not only homogenize the artist-style of different artworks of the same artist but also lack generalization for the unseen artists. To solve these challenges, we propose a double-style transferring module (DSTM). It extracts different artist-style and artwork-style from different artworks (even untrained) and preserves the intrinsic diversity between different artworks of the same artist. DSTM swaps the two styles in the adversarial training and encourages realistic image generation given arbitrary style combinations. However, learning style from single artwork can often cause over-adaption to it, resulting in the introduction of structural features of style image. We further propose an edge enhancing module (EEM) which derives edge information from multi-scale and multi-level features to enhance structural consistency. We broadly evaluate our method across six large-scale benchmark datasets. Empirical results show that our method achieves arbitrary artist-style and artwork-style extraction from a single artwork, and effectively avoids introducing the style image’s structural features. Our method improves the state-of-the-art deception rate from 58.9% to 67.2% and the average FID from 48.74 to 42.83.

PDF Details DOI

AAAI Conference 2020 Conference Paper

FET-GAN: Font and Effect Transfer via K-shot Adaptive Instance Normalization

Wei Li
Yongxing He
Yanwei Qi
Zejian Li
Yongchuan Tang

Text effect transfer aims at learning the mapping between text visual effects while maintaining the text content. While remarkably successful, existing methods have limited robustness in font transfer and weak generalization ability to unseen effects. To address these problems, we propose FET-GAN, a novel end-to-end framework to implement visual effects transfer with font variation among multiple text effects domains. Our model achieves remarkable results both on arbitrary effect transfer between texts and effect translation from text to graphic objects. By a few-shot ﬁne-tuning strategy, FET-GAN can generalize the transfer of the pre-trained model to the new effect. Through extensive experimental validation and comparison, our model advances the state-of-the-art in the text effect transfer task. Besides, we have collected a font dataset including 100 fonts of more than 800 Chinese and English characters. Based on this dataset, we demonstrated the generalization ability of our model by the application that complements the font library automatically by few-shot samples. This application is signiﬁcant in reducing the labor cost for the font designer.

PDF Details

AAAI Conference 2019 Conference Paper

Learning Disentangled Representation with Pairwise Independence

Zejian Li
Yongchuan Tang
Wei Li
Yongxing He

Unsupervised disentangled representation learning is one of the foundational methods to learn interpretable factors in the data. Existing learning methods are based on the assumption that disentangled factors are mutually independent and incorporate this assumption with the evidence lower bound. However, our experiment reveals that factors in real-world data tend to be pairwise independent. Accordingly, we propose a new method based on a pairwise independence assumption to learn the disentangled representation. The evidence lower bound implicitly encourages mutual independence of latent codes so it is too strong for our assumption. Therefore, we introduce another lower bound in our method. Extensive experiments show that our proposed method gives competitive performances as compared with other state-of-the-art methods.

PDF Details

IJCAI Conference 2018 Conference Paper

Unsupervised Disentangled Representation Learning with Analogical Relations

Zejian Li
Yongchuan Tang
Yongxing He

Learning the disentangled representation of interpretable generative factors of data is one of the foundations to allow artificial intelligence to think like people. In this paper, we propose the analogical training strategy for the unsupervised disentangled representation learning in generative models. The analogy is one of the typical cognitive processes, and our proposed strategy is based on the observation that sample pairs in which one is different from the other in one specific generative factor show the same analogical relation. Thus, the generator is trained to generate sample pairs from which a designed classifier can identify the underlying analogical relation. In addition, we propose a disentanglement metric called the subspace score, which is inspired by subspace learning methods and does not require supervised information. Experiments show that our proposed training strategy allows the generative models to find the disentangled factors, and that our methods can give competitive performances as compared with the state-of-the-art methods.

PDF Details