https://1zeryu.github.io/Recent content on Yukun ZhouHugozh-cnFri, 24 Apr 2026 10:00:00 +0800https://1zeryu.github.io/p/deepseek-v4/Fri, 24 Apr 2026 10:00:00 +0800https://1zeryu.github.io/p/deepseek-v4/<p>4 月 24 日，DeepSeek 发布其最新版本 DeepSeek-V4<sup id="fnref:1"><a href="#fn:1" class="footnote-ref" role="doc-noteref">1</a></sup> 技术报告，并同步开源代码与模型权重。DeepSeek-V4 在多项评测中达到 SOTA 水平；在效率方面，相较于 DeepSeek-V3.2，DeepSeek-V4-Pro 的单 token 推理 FLOPs 降至 27%，KV 缓存缩减至 10%，从而稳定支持百万级 token 的长上下文处理。</p>https://1zeryu.github.io/p/hunyuan-mirror2/Thu, 23 Apr 2026 10:00:00 +0800https://1zeryu.github.io/p/hunyuan-mirror2/<h2 id="什么是3d重建">什么是3D重建？</h2> <p><a id="world-models-figure"></a></p> <figure> <img src="2Dvs3D.png" alt="3D重建示意图" style="width:90%;" /> </figure> <p>在数字世界里，2D 信息就是一张张图像：它们由像素排成的平面画布组成，每个像素只携带 (x, y) 坐标和颜色、亮度等属性，像一幅挂在墙上的画。而 3D 信息则像把画撕下来、折成立体纸雕——点不再只是平面上的像素，而是拥有了 (x, y, z) 三维坐标的“点云”；它们可以描述凹凸、距离与体积。</p>https://1zeryu.github.io/p/jepa/Fri, 17 Apr 2026 10:15:00 +0800https://1zeryu.github.io/p/jepa/<h2 id="引言">引言</h2> <p>JEPA<sup id="fnref:1"><a href="#fn:1" class="footnote-ref" role="doc-noteref">1</a></sup>（Joint Embedding Predictive Architecture）是由 <a href="http://yann.lecun.com/">LeCun</a> 等人提出的一类预测学习框架。本篇笔记分四个部分：先介绍 JEPA 的核心直觉（无数学推导），再解释 JEPA 的技术架构，然后以开源实现为例说明其推理与训练细节，最后补充一些问题与个人理解。</p>https://1zeryu.github.io/p/qwen3-omni/Mon, 13 Apr 2026 10:00:00 +0800https://1zeryu.github.io/p/qwen3-omni/<h2 id="引言大一统多模态大模型">引言：大一统多模态大模型</h2> <!-- 背景：单模态大模型的局限 --> <p>过去几年，尽管 GPT 系列在问答、编程、创作上不断刷新纪录，但它们都只能处理文本输入，无法直接感知视觉和听觉信息。与仅聚焦文本、对声画“关门”的单模态 GPT 大模型不同，人类天然是多模态智能体，能同步感知视觉与听觉并借文字、语音或动作完成信息交流。</p>https://1zeryu.github.io/about/Mon, 01 Jan 0001 00:00:00 +0000https://1zeryu.github.io/about/<h2 id="news">News</h2> <ul class="entry-list news-list"> <li> <span class="entry-date">2026.06</span> <span class="entry-body entry-text">Successfully concluded the GigaWorld Track of the GigaBrain Challenge (CVPR 2026 Workshop). <a href="https://gigaai-research.github.io/GigaBrain-Challenge-2026/index.html#about">link</a></span> </li> <li> <span class="entry-date">2025.07</span> <span class="entry-body entry-text">Two papers accepted to CVPR 2025 (HumanDreamer, SpeeD).</span> </li> <li> <span class="entry-date">2024.07</span> <span class="entry-body entry-text">One paper accepted to ECCV 2024 (InfoGrowth).</span> </li> <li> <span class="entry-date">2024.05</span> <span class="entry-body entry-text">Open-Sora has been released.</span> </li> </ul> <h2 id="selected-work">Selected Work</h2> <ul class="pub-list"> <li class="pub-item"> <div class="pub-thumb"> <img src="https://1zeryu.github.io/img/pubs/Speed_demo_hu_cdaacc5852eefe08.png" width="480" height="231" alt="A Closer Look at Time Steps is Worthy of Triple Speed-Up for Diffusion Model Training" loading="lazy"> </div> <div class="pub-body"> <div class="pub-venue">CVPR 2025</div> <div class="pub-title">A Closer Look at Time Steps is Worthy of Triple Speed-Up for Diffusion Model Training</div> <div class="pub-authors">Kai Wang, Mingjia Shi, <strong>Yukun Zhou</strong>, et al.</div> <div class="pub-desc">A plug-and-play, architecture-agnostic approach that achieves 3× acceleration across diffusion architectures, datasets, and tasks.</div> <div class="pub-links"> <a href="https://bdemo.github.io/SpeeD/">Project</a><a href="https://github.com/NUS-HPC-AI-Lab/SpeeD">Code</a><a href="https://openaccess.thecvf.com/content/CVPR2025/papers/Wang_A_Closer_Look_at_Time_Steps_is_Worthy_of_Triple_CVPR_2025_paper.pdf">Paper</a> </div> </div> </li> <li class="pub-item"> <div class="pub-thumb"> <img src="https://1zeryu.github.io/img/pubs/dataset_growth_hu_351a1cd3b09565a1.png" width="480" height="304" alt="Dataset Growth" loading="lazy"> </div> <div class="pub-body"> <div class="pub-venue">ECCV 2024</div> <div class="pub-title">Dataset Growth</div> <div class="pub-authors">Ziheng Qin, Zhaopan Xu, <strong>Yukun Zhou</strong>, et al.</div> <div class="pub-desc">Improves data quality and efficiency on both single-modal and multi-modal tasks with an efficient, scalable design.</div> <div class="pub-links"> <a href="https://github.com/NUS-HPC-AI-Lab/InfoGrowth?tab=readme-ov-file">Code</a><a href="https://arxiv.org/pdf/2405.18347">Paper</a> </div> </div> </li> <li class="pub-item"> <div class="pub-thumb"> <img src="https://1zeryu.github.io/img/pubs/gigaworld-0_hu_a61819688c966e70.png" width="480" height="353" alt="GigaWorld-0: World Models as Data Engine to Empower Embodied AI" loading="lazy"> </div> <div class="pub-body"> <div class="pub-venue">Technical Report</div> <div class="pub-title">GigaWorld-0: World Models as Data Engine to Empower Embodied AI</div> <div class="pub-authors">GigaAI</div> <div class="pub-desc">A unified world model framework designed explicitly as a data engine for Vision-Language-Action learning. Responsible for training and evaluating the video foundation model, downstream fine-tuning, and inference acceleration.</div> <div class="pub-links"> <a href="https://giga-world-0.github.io/">Project</a><a href="https://github.com/open-gigaai/giga-world-0">Code</a><a href="https://arxiv.org/pdf/2511.19861">Paper</a><a href="https://huggingface.co/open-gigaai">Model</a> </div> </div> </li> <li class="pub-item"> <div class="pub-thumb"> <img src="https://1zeryu.github.io/img/pubs/opensora_demo_hu_9973b817b09e4f9d.png" width="480" height="324" alt="Open-Sora: Democratizing Efficient Video Production for All" loading="lazy"> </div> <div class="pub-body"> <div class="pub-venue">Open-Source Project</div> <div class="pub-title">Open-Sora: Democratizing Efficient Video Production for All</div> <div class="pub-authors">HPC-AI Tech</div> <div class="pub-desc">An open-source video generation model with full training, inference, data preparation code and model weights. Project participant as an intern at Luchen Technology.</div> <div class="pub-links"> <a href="https://github.com/hpcaitech/Open-Sora">Code</a><a href="https://arxiv.org/pdf/2412.20404">Paper</a><a href="https://huggingface.co/collections/hpcai-tech/open-sora-66373960c8950692a4394e2d">Model</a> </div> </div> </li> </ul> <h2 id="education">Education</h2> <ul> <li>B.Eng., Xidian University, 2025</li> </ul>https://1zeryu.github.io/search/Mon, 01 Jan 0001 00:00:00 +0000https://1zeryu.github.io/search/https://1zeryu.github.io/writing/Mon, 01 Jan 0001 00:00:00 +0000https://1zeryu.github.io/writing/