arXiv subjects

Wentao Xiong

Publications and source records attributed to Wentao Xiong.

4 recordsLinked to original sources

arXiv2025

SynthVLM: Towards High-Quality and Efficient Synthesis of Image-Caption Datasets for Vision-Language Models

Vision-Language Models (VLMs) have recently emerged, demonstrating remarkable vision-understanding capabilities. However, training these models requires large-scale datasets, which brings challenges related to efficiency, effectiveness, and quality of web data. In this paper, we introduce SynthVLM, a new data synthesis and curation method for generating image-caption pairs. Unlike traditional methods, where captions are generated from images, SynthVLM utilizes advanced diffusion models and high-quality captions to synthesize and select images from text captions, thereby creating precisely aligned image-text pairs. We further introduce SynthVLM-100K, a high-quality dataset consisting of 100K curated and synthesized image-caption pairs. In both model and human evaluations, SynthVLM-100K outperforms traditional real-world datasets. Leveraging this dataset, we develop a new family of multimodal large language models (MLLMs), SynthVLM-7B and SynthVLM-13B, which achieve state-of-the-art (SOTA) performance on various vision question-answering (VQA) tasks. Notably, our models outperform LLaVA across most metrics with only 18\% pretrain data. Furthermore, SynthVLM-7B and SynthVLM-13B attain SOTA performance on the MMLU benchmark, demonstrating that the high-quality SynthVLM-100K dataset preserves language abilities.

cs.CV↗

arXiv2024

Harnessing Vision-Language Pretrained Models with Temporal-Aware Adaptation for Referring Video Object Segmentation

The crux of Referring Video Object Segmentation (RVOS) lies in modeling dense text-video relations to associate abstract linguistic concepts with dynamic visual contents at pixel-level. Current RVOS methods typically use vision and language models pretrained independently as backbones. As images and texts are mapped to uncoupled feature spaces, they face the arduous task of learning Vision-Language (VL) relation modeling from scratch. Witnessing the success of Vision-Language Pretrained (VLP) models, we propose to learn relation modeling for RVOS based on their aligned VL feature space. Nevertheless, transferring VLP models to RVOS is a deceptively challenging task due to the substantial gap between the pretraining task (static image/region-level prediction) and the RVOS task (dynamic pixel-level prediction). To address this transfer challenge, we introduce a framework named VLP-RVOS which harnesses VLP models for RVOS through temporal-aware adaptation. We first propose a temporal-aware prompt-tuning method, which not only adapts pretrained representations for pixel-level prediction but also empowers the vision encoder to model temporal contexts. We further customize a cube-frame attention mechanism for robust spatial-temporal reasoning. Besides, we propose to perform multi-stage VL relation modeling while and after feature extraction for comprehensive VL understanding. Extensive experiments demonstrate that our method performs favorably against state-of-the-art algorithms and exhibits strong generalization abilities.

cs.CV↗

arXiv2017

A Novel Thiophene-fused Ending Group Enabling an Excellent Small Molecule Acceptor for High-performance Fullerene-free Polymer Solar Cells with 11.8% Efficiency

In this article, we designed and synthesized a novel small molecule acceptor of ITCPTC with thiophene-fused ending group by employing a new active methylene precursor of CPTCN. The ITCPTC based polymer solar cells with PBT1-EH as donor achieved very high PCEs of up to 11.8% with a remarkably enhanced fill factor (FF) of 0.751, a near 20% boost in PCE with respect to the ITIC based control device. These values are among the highest PCEs and FFs for PSCs. In the whole study, we made contrasts with ITIC to understand the reasons of excellent performance of ITCPTC-based PSCs through various measurements, such as GIWAXS and RSoXS. We revealed that the simple modification of ITIC into ITCPTC not only change the material electronic structure, but also mediate the material interactions and crystallization, which contribute together to the excellent performance of ITCPTC based PSCs.

physics.chem-ph↗

arXiv2011

The Limiting Spectral Measure for Ensembles of Symmetric Block Circulant Matrices

Given an ensemble of NxN random matrices, a natural question to ask is whether or not the empirical spectral measures of typical matrices converge to a limiting spectral measure as N --> oo. While this has been proved for many thin patterned ensembles sitting inside all real symmetric matrices, frequently there is no nice closed form expression for the limiting measure. Further, current theorems provide few pictures of transitions between ensembles. We consider the ensemble of symmetric m-block circulant matrices with entries i.i.d.r.v. These matrices have toroidal diagonals periodic of period m. We view m as a "dial" we can "turn" from the thin ensemble of symmetric circulant matrices, whose limiting eigenvalue density is a Gaussian, to all real symmetric matrices, whose limiting eigenvalue density is a semi-circle. The limiting eigenvalue densities f_m show a visually stunning convergence to the semi-circle as m tends to infinity, which we prove. In contrast to most studies of patterned matrix ensembles, our paper gives explicit closed form expressions for the densities. We prove that f_m is the product of a Gaussian and a degree 2m-2 polynomial; the formula equals that of the m x m Gaussian Unitary Ensemble (GUE). The proof is by the moments. The new feature, which allows us to obtain closed form expressions, is converting the central combinatorial problem in the moment calculation into an equivalent counting problem in algebraic topology. We end with a generalization of the m-block circulant pattern, dropping the assumption that the m random variables be distinct. We prove that the limiting spectral distribution exists and is determined by the pattern of the independent elements within an m-period, depending on not only the frequency at which each element appears, but also the way the elements are arranged.

math.PR↗