Bin Lin林彬

Ph.D. Candidate in Computer Science计算机科学博士研究生

Peking University北京大学  ·  Shenzhen, China中国 · 深圳

linbin.ece@stu.pku.edu.cn  ·  linbin203279@gmail.com

I work on multimodal large models — building open, scalable systems that understand and generate across vision and language, and increasingly unify the two. Advised by Prof. Li Yuan (袁粒) at Peking University.

我的研究方向是多模态大模型——构建开放、可扩展的系统，使其能够跨视觉与语言进行理解与生成，并逐步走向二者的统一。师从北京大学袁粒教授。

I'm a strong believer in open research: I lead and contribute to widely-used projects such as Open-Sora Plan, UniWorld, and Video-LLaVA, releasing the code, models, and data behind every paper.

我笃信开放研究：我主导并参与了多个被广泛使用的项目，如 Open-Sora Plan、UniWorld 和 Video-LLaVA，并为每篇论文开源代码、模型与数据。

Multimodal Understanding多模态理解 Multimodal Generation多模态生成 Unified Models统一模型 Video Generation视频生成

Tencent Project UP Scholarship (腾讯青云奖学金)

One of only 15 awardees nationwide · ~400 applicants from 70+ universities · 2026全国仅 15 人入选 · 来自 70+ 所高校的约 400 名申请者 · 2026

4,700+

Citations总引用量

1 paper 1,700+ · 7 papers 100+1 篇 1,700+ · 7 篇 100+

20K+

GitHub StarsGitHub Stars

14M+

Model & Data Downloads模型与数据下载

Top 100

Hugging Face InfluencerHugging Face 影响力

2024.02–2025.01

• 2026.06Released OSP-Next and GEAR — our latest work on efficient video and autoregressive visual generation.发布 OSP-Next 与 GEAR —— 高效视频生成与自回归视觉生成方向的最新工作。
• 2026.01Selected for the Tencent Project UP Scholarship (腾讯青云奖学金) — one of only 15 awardees nationwide in AI.入选腾讯青云奖学金 —— 全国 AI 领域仅 15 人获此殊荣。
• 2025.07MoE-LLaVA accepted to IEEE Transactions on Multimedia (TMM).MoE-LLaVA 被 IEEE Transactions on Multimedia (TMM) 接收。
• 2025.06Released Open-Sora Plan v1.5 (12K+ ⭐) and launched UniWorld.发布 Open-Sora Plan v1.5（12K+ ⭐），并推出 UniWorld。
• 2025.03Video-LLaVA ranked #1 among the Most Influential EMNLP 2025 Papers.Video-LLaVA 位列 EMNLP 2025 最具影响力论文第 1 名。
• 2025.02WF-VAE accepted to CVPR 2025.WF-VAE 被 CVPR 2025 接收。
• 2024.09Video-LLaVA accepted to EMNLP 2025.Video-LLaVA 被 EMNLP 2025 接收。
• 2024.01LanguageBind accepted to ICLR 2024.LanguageBind 被 ICLR 2024 接收。

PreprintFirst Author第一作者

GEAR: Guided End-to-End AutoRegression for Image Synthesis

Bin Lin, Zheyuan Liu, et al., Li Yuan

Jointly trains a VQ tokenizer and an autoregressive generator end-to-end, guided by representation alignment — up to 10× faster AR training than LlamaGen-REPA, with better spatial features that generalize across quantizers and to text-to-image.将 VQ 分词器与自回归生成器端到端联合训练，并由表征对齐引导——自回归训练较 LlamaGen-REPA 最高加速 10×，特征更具空间一致性，且可泛化到多种量化器与文生图。

arXiv Code 0 🤗 Models

EMNLP 2025First Author第一作者 #6 GitHub TrendingGitHub 趋势榜第 6

Video-LLaVA: Learning United Visual Representation by Alignment Before Projection

Bin Lin, Yang Ye, Bin Zhu, et al., Li Yuan

#1 Most Influential EMNLP 2025 Paper with 1,000+ citations — reached video-QA SOTA in just two days on a single node.EMNLP 2025 最具影响力论文第 1 名，引用量 1,000+——在单机上仅用两天即达到视频问答 SOTA。

arXiv Code 3.5k 🤗 Models·Dataset Cited by –

For the complete list, see my Google Scholar. ^* denotes equal contribution.完整列表见我的 Google Scholar。^* 表示共同贡献。

• IEEE TMMFirst Author第一作者 #6 GitHub TrendingGitHub 趋势榜第 6 MoE-LLaVA: Mixture of Experts for Large Vision-Language Models. The first sparse multimodal LLM trainable on a single node.首个可在单机上训练的稀疏多模态大模型。
  arXiv Code 2.3k 🤗 Models·Dataset Cited by –
• CVPR 2025Co-First Author共同一作 WF-VAE: Enhancing Video VAE by Wavelet-Driven Energy Flow for Latent Video Diffusion. A wavelet-driven video VAE with higher compression efficiency and reconstruction quality.小波驱动的视频 VAE，兼具更高的压缩效率与重建质量。
  arXiv Code 203 🤗 Model Cited by –
• ICLR 2024Co-First Author共同一作 LanguageBind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment. Binds N modalities through language, enabling unified cross-modal pretraining and retrieval.以语言为桥连接 N 种模态，实现统一的跨模态预训练与检索。
  arXiv Code 881 🤗 Models Cited by –
• NeurIPS 2025Co-First Author共同一作 ImgEdit: A Unified Image Editing Dataset and Benchmark. A 1.2M-pair image-editing dataset and benchmark, with the ImgEdit-E1 editing model.包含 120 万对样本的图像编辑数据集与基准，并提供 ImgEdit-E1 编辑模型。
  arXiv Code 316 🤗 Dataset Cited by –
• ICML 2026Co-Author合作作者 WISE: A World Knowledge-Informed Semantic Evaluation for Text-to-Image Generation. A world-knowledge-informed benchmark for evaluating text-to-image generation.引入世界知识的文本到图像生成评测基准。
  arXiv Code 205 🤗 Dataset Cited by –
• NeurIPS 2024Co-Author合作作者 ShareGPT4Video: Improving Video Understanding and Generation with Better Captions. Boosts video understanding and generation with dense, high-quality captions.借助密集且高质量的视频描述，提升视频理解与生成能力。
  arXiv Code 1.1k 🤗 Dataset Cited by –

Open source is at the core of my research — collectively 20K+ GitHub stars and 14M+ model & data downloads.开源是我研究的核心 —— 累计获得 20K+ GitHub Stars，相关模型与数据下载量超过 14M+。

• 2023.06Outstanding Graduate of Sichuan Province, China.四川省优秀毕业生。
• 2022.11Outstanding Student, Sichuan Agricultural University (top 10 university-wide).四川农业大学优秀学生（全校前 10）。
• 2022.10National Scholarship — highest scholarship from the Ministry of Education, China.国家奖学金 —— 教育部最高奖学金。
• 2021.11National First Prize, National Undergraduate Mathematical Modeling Contest.全国一等奖，全国大学生数学建模竞赛。
• 2021.10National Scholarship — highest scholarship from the Ministry of Education, China.国家奖学金 —— 教育部最高奖学金。