[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-53":104},{"enabled":4,"title":5,"statusText":5,"buttonText":5,"campSlug":5,"stage":6,"classroomId":7,"hasClassroom":4},false,"",0,null,{"total":9,"list":10,"pageNum":44,"pageSize":45,"size":9,"startRow":44,"endRow":9,"pages":44,"prePage":6,"nextPage":6,"isFirstPage":46,"isLastPage":46,"hasPreviousPage":4,"hasNextPage":4,"navigatePages":47,"navigatepageNums":48,"navigateFirstPage":44,"navigateLastPage":44},3,[11,25,35],{"slug":12,"title":13,"logo":7,"liveUrl":14,"prize":15,"sponsor":16,"description":17,"organizer":18,"startTime":19,"endTime":20,"registrationDeadline":21,"status":22,"dashboardVisibility":23,"createdAt":24},"spring2026","2026春季启元人工智能大赛","https:\u002F\u002Fwww.bilibili.com\u002Fblackboard\u002Flive\u002Flive-activity-player.html?cid=1822703816&quality=0","200,000","麒麟软件有限公司","“2026春季启元人工智能大赛”是由启元实验室主办的面向全国人工智能领域开发者、科研人员及高校学生的权威赛事。大赛旨在促进与培养具备工程实践能力和创新能力的优秀人工智能领域开发者、推广国产软硬件技术与生态、助力国家人工智能产业高质量发展。\n各参赛选手需在规定的比赛周期内，依据赛事规则选定赛题并完成相应挑战，优胜者将共同角逐总额20万元的赛事奖金池，并且还将有机会获得实习和知名企业优质岗位内推资格。","启元实验室","2026-04-19 16:00:00","2026-07-13 16:00:00","2026-06-16 16:00:00","ongoing","private","2026-04-16 08:00:00",{"slug":26,"title":27,"logo":7,"liveUrl":14,"prize":28,"sponsor":7,"description":29,"organizer":18,"startTime":30,"endTime":31,"registrationDeadline":32,"status":33,"dashboardVisibility":23,"createdAt":34},"autumn2025","2025秋季启元人工智能大赛","150,000","“2025秋季启元人工智能大赛”是由启元实验室主办的面向全国人工智能领域开发者、科研人员及高校学生的权威赛事。大赛旨在促进与培养具备工程实践能力和创新能力的优秀人工智能领域开发者、推广国产软硬件技术与生态、助力国家人工智能产业高质量发展。","2025-11-08 08:00:00","2026-01-13 08:00:00","2025-12-30 08:00:00","ended","2025-11-03 08:00:00",{"slug":36,"title":37,"logo":7,"liveUrl":14,"prize":38,"sponsor":7,"description":39,"organizer":18,"startTime":40,"endTime":41,"registrationDeadline":42,"status":33,"dashboardVisibility":23,"createdAt":43},"summer2025","启元人工智能大赛(系统方向)","80,000","启元人工智能大赛由启元实验室主办，其中系统赛道基于启元实验室九源软件栈的相关开源项目，通过拉动开源社区力量，一方面促进国产智能计算生态的发展，另一方面也为人工智能系统方向的学生及从业者提供一个实践平台。","2025-07-08 08:00:00","2025-08-25 08:00:00","2025-07-22 08:00:00","2025-06-22 08:00:00",1,5,true,8,[44],{"total":45,"list":50,"pageNum":44,"pageSize":45,"size":45,"startRow":44,"endRow":45,"pages":44,"prePage":6,"nextPage":6,"isFirstPage":46,"isLastPage":46,"hasPreviousPage":4,"hasNextPage":4,"navigatePages":47,"navigatepageNums":103,"navigateFirstPage":44,"navigateLastPage":44},[51,62,73,84,94],{"id":52,"slug":53,"title":54,"logo":55,"publicLiveEmbedUrl":14,"description":56,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":16,"startTime":58,"endTime":59,"status":22,"dashboardVisibility":60,"createdAt":58,"updatedAt":61},"019eeeff-6b00-7d55-a629-efafded3e232","summer2026","2026年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2026-06\u002F019eef7f-7816-7828-9309-79c2c61dd568.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=DdQTMKVo1nful94EsDeKYRm5DII%3D","由启元实验室发起的InfiniTensor大模型与人工智能系统训练营，面向开源社区、高校及产业界伙伴开放。活动以「普及AI全栈知识技术、培养工程实践与创新型人才、推广国产软硬件技术生态」为核心目标，助力国家人工智能产业高质量发展。","...","2026-06-22 08:00:00","2026-10-19 08:00:00","public","2026-10-12 08:00:00",{"id":63,"slug":64,"title":65,"logo":66,"publicLiveEmbedUrl":7,"description":67,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":68,"startTime":69,"endTime":70,"status":33,"dashboardVisibility":23,"createdAt":71,"updatedAt":72},"019af3f3-e06a-7245-aa8d-37dc99a69c73","winter2025","2025年冬季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2025-12\u002F019afc69-b7ca-7805-91fb-605429671c01.png?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HWdVGs1NgNzaJvqpNa66tZqqOTo%3D","由启元实验室发起的InfiniTensor大模型与人工智能系统训练营，面向开源社区、高校及产业界伙伴开放。活动以「普及AI全栈知识技术、培养工程实践与创新型人才、推广国产软硬件技术生态」为核心目标，助力国家人工智能产业高质量发展。 \n","新华三技术有限公司","2025-12-25 08:00:00","2026-03-27 08:00:00","2025-12-06 08:00:00","2026-10-22 20:24:30",{"id":74,"slug":36,"title":75,"logo":76,"publicLiveEmbedUrl":7,"description":77,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":78,"startTime":79,"endTime":80,"status":81,"dashboardVisibility":23,"createdAt":82,"updatedAt":83},"01980342-208c-7803-b3f3-19e1d31960ba","2025年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2025-07\u002F0198045e-28fe-78f0-95fe-23835ce58395.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5IITjeecr8z%2B54QsS2SaoytOVRM%3D","InfiniTensor大模型与人工智能系统训练营由启元实验室及清华大学 AI 系统方向团队发起，面向开源社区、高校和产业界，普及人工智能各方面知识和技术，培养相关人才。训练营联合开源项目开发者、学术界大牛、工业界资深工程师参与授课与项目指导。课程采用「线上直播授课 + 在线答疑 + 项目实践辅导 」的方式进行，确保每位同学都能获得充分成长。训练营全程免费，通过视频、直播课、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。","北京积算科技有限公司","2025-07-14 08:00:00","2025-09-30 08:00:00","upcoming","2025-07-13 08:00:00","2026-10-22 20:24:58",{"id":85,"slug":86,"title":87,"logo":88,"publicLiveEmbedUrl":7,"description":89,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":5,"startTime":90,"endTime":91,"status":81,"dashboardVisibility":23,"createdAt":92,"updatedAt":93},"01980cf9-a7d8-7796-ae86-0bbe05334add","winter2024","2024年冬季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Ftesting-cdn.infinitensor.com\u002Fuploads\u002F2025-07\u002F01981269-ce23-7624-a287-9f07dd0d85f6.jpg","InfiniTensor大模型与人工智能系统训练营由清华大学发起，面向开源社区、高校和产业界，普及人工智能各方面知识和技术，培养相关人才。训练营全程免费，通过视频、直播课、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。报名学员请阅读“导学”课程页的训练营介绍。","2025-01-05 08:00:00","2025-03-22 08:00:00","2025-07-15 08:00:00","2026-10-22 20:24:39",{"id":95,"slug":96,"title":97,"logo":98,"publicLiveEmbedUrl":7,"description":99,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":5,"startTime":100,"endTime":101,"status":81,"dashboardVisibility":23,"createdAt":92,"updatedAt":102},"01980c20-363a-7dc3-8f5a-58a68e08dcc4","summer2024","2024年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Ftesting-cdn.infinitensor.com\u002Fuploads\u002F2025-07\u002F01981269-fbfa-788a-afcd-17fa9be74ff9.jpg","InfiniTensor大模型与人工智能系统训练营由清华大学发起，面向开源社区和高校，普及人工智能各方面知识和技术，培养相关人才。训练营全程免费，通过课程、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。","2024-07-20 08:00:00","2024-09-29 08:00:00","2026-10-22 20:24:49",[44],{"currentItem":105,"processedData":125},{"img":106,"keywords":107,"treeType":44,"description":108,"showStatus":44,"updateTime":109,"type":110,"title":111,"content":112,"nextArticle":113,"createTime":109,"sortOrder":117,"canonicalPath":118,"id":119,"contentType":120,"prevArticle":121},"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_fm.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TVY1JmiKi4PxhsToCMwcYVmt9ik%3D","InfiniTensor,冬季训练营,多模态智能,跨模态理解,应用实践","讲解多模态智能的核心方法与落地场景，涵盖图文融合、跨模态理解与应用实践。","2026-10-22 20:38:11",2,"InfiniTensor 冬季训练营前沿模型第五课《多模态智能及其应用》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw01.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BeJHdYop0xVtrCUoChLTYA8BmYk%3D)\r\n\r\n\r\n\r\n2026 年 2 月 3 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营前沿模型系列第五课《多模态智能及其应用》开讲。\r\n\r\n本节课介绍当前热门的理解生成一体化技术，特别是针对 **Nano Banana Pro** 模型，探讨其核心技术、应用场景及未来发展趋势。\r\n\r\n\r\n### Nano Banana Pro 技术\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw02.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HzQaxPhXQx3vSudg2HUq9%2BLFvNM%3D)\r\n\r\n\r\n### 原生多模态模型与长上下文思考\r\n- **原生多模态模型**：统一生成与理解的能力\r\n- **长上下文窗口**：能够接收多种图片、多轮对话及复杂长文描述，实现对话式出图。\r\n\r\n### 数据生产与评估反馈\r\n- 高质量数据的重要性：强调高质量数据对模型训练的重要性，特别是图文交织的训练方式。\r\n- 人类评估体系：人类美学评估、面部实用性评估等评估体系，以及文字渲染能力的量化评估方法。\r\n- 数据闭环与反馈：通过人类反馈形成数据制造闭环，提升模型生成准确性。\r\n\r\n### 开源多模态生成模型的挑战\r\n- 物理规律违背问题：开源多模态生成模型在生成内容时违背用户指令或物理规律的问题，如跑步机上反着跑的人、水杯破碎顺序错误等。\r\n- 理解能力的缺失：开源生成模型在理解人类意图方面不足。\r\n\r\n### 理解生成一体化技术脉络\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw03.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CE40bTVeQrGXfAQ66zaoYRYRbak%3D)\r\n\r\n\r\n- **图文对比学习：** CLIP 模型利用对比学习实现图像与文本对齐的技术。\r\n- **GPT4 与 Gemini 2.0：** GPT4 与 Gemini 2.0 等原生多模态模型在图文理解方面的强大能力。\r\n- **多模态生成技术：** Stable Diffusion、DALL·E 等模型在文生图方面的技术突破，以及 Flux 系列模型在指令遵循能力方面提升。\r\n\r\n### 理解生成一体化框架设计\r\n#### 1. 理想的一体化框架\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw04.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=b53PL8kj0jExsyWkVzSu8IhubCQ%3D)\r\n\r\n#### 2. 现有理解生成技术\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw05.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WrLZJvoYb339HmR0dt3I3R4fFYs%3D)\r\n\r\n- 编解码结构设计\r\n- 视觉特征方式\r\n- 视觉表示特性\r\n\r\n#### 3. 现有理解生成模型框架\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw06.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cj7MyVImeiDzy0J7u1KjnCSgyfY%3D)\r\n\r\n##### （1）生成外挂：AD-QC-LS\r\n- 优点：可快速使模型同时具有不错的理解和生成能力。\r\n- 缺点：参数量大；特征割裂，无充分融合。\r\n\r\n##### （2）极致统一\r\n- **A-Q-L**\r\n  - 优点：可快速嵌入 AR 框架。\r\n  - 缺点：难以做高维度语义理解任务；表示空间少；词表固定。\r\n- **A-C-L**\r\n  - 优点：表示空间更丰富，细节生成更好。\r\n  - 缺点：难以做高维度语义理解任务；可能出现上下文遗忘等问题。\r\n- **A-Q-S**\r\n  - 优点：一个视觉编码器完成两种视觉表示建模，同时提升理解和生成性能；离散后便于直接嵌入 AR 架构。\r\n  - 缺点：表示空间少，细节和语义难以完全保持；词表固定，视觉编码器固定。\r\n- **D-Q-L**\r\n  - 优点：鉴于 LLaDA 的 Diffusion 解码框架，符合图像直观的解码方式，同时解码效率相对高。\r\n  - 缺点：1. 基础模型仍在探索中，基础能力不足; 2. 仍局限于 VQ，受其本身能力局限。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw07.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=svuMDxN9rWt44n6k4UjIlqSLO%2FQ%3D)\r\n\r\n\r\n##### （3）融合建模\r\n- **A-Q-LS**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw08.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OCfdcvRVx0QQAFE7i%2BsP30Z0DLU%3D)\r\n\r\n- **A-QC-SL**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw09.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CQ4ZG3S%2Bf7J6DAK7g6bQb8wQj38%3D)\r\n\r\n- **AD-C-L**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw10.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9biUB4m4y4zrudvOy4XyDbKOoh8%3D)\r\n\r\n- **AD-Q-SL**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw11.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Uitx0eONWWCPiMT2xIIKGtJAIf0%3D)\r\n\r\n- **AD-QC-SL**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw12.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QYzrGNwyR26dxHIITpfT2l9yicI%3D)\r\n\r\n- **AD-C-SL**\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw13.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=UN3dXJqfcZy32t3GZ8uCbskvjHw%3D)\r\n\r\n### DeepMind 工作启发\r\n从 DeepMind 相关的研究基础来看，有两点技术坚持：\r\n- 1.坚持利用连续特征 + Diffusion Loss 用于图像生成。\r\n- 2.是利用 AR 还是 AR+Diffusion 仍在探索，但基本上是在 LLM 内部完成一体化建模。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw14.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2NHYTwRsyNQPZh0ZvxnBChTSzGM%3D)\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw15.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=jTnv8OQ436y%2Fwu1Bjsi2%2Fh3UYcE%3D)\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw16.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5QYcN0gfmgmzbS5%2FzjX%2FAdsZ5E4%3D)\r\n\r\n\r\n\r\n![2025冬季训练营课程表](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_kcb.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=tmWtTrH27PV8kQApQCBd1QaHaNU%3D)\r\n\r\n### 💡 学习贴士\r\n更多细节及完整课程内容，请观看直播或查看课程回放。\r\n\r\n链接：https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\r\n\r\n### 📚 训练营报名持续进行中\r\n扫码报名与加入群聊，或点击 **「阅读原文」** 直达官网报名\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_ewm.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=O1NL5bahW5nvf%2FlqFbFObNY5fbM%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-52","study-52","InfiniTensor 冬季训练营推理引擎第七课《InfiniLM》",56,"\u002Fstudy-material\u002Fstudy-53","study-53","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-54","study-54","2025 冬季 InfiniTensor 训练营｜C++ 实践课（一）",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw01.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BeJHdYop0xVtrCUoChLTYA8BmYk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 2 月 3 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营前沿模型系列第五课《多模态智能及其应用》开讲。\u003C\u002Fp>\n\u003Cp>本节课介绍当前热门的理解生成一体化技术，特别是针对 \u003Cstrong>Nano Banana Pro\u003C\u002Fstrong> 模型，探讨其核心技术、应用场景及未来发展趋势。\u003C\u002Fp>\n\u003Ch3 id=\"nano-banana-pro-技术\">\u003Ca class=\"anchor\" href=\"#nano-banana-pro-技术\" aria-label=\"跳转到标题：Nano Banana Pro 技术\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Nano Banana Pro 技术\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw02.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HzQaxPhXQx3vSudg2HUq9%2BLFvNM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"原生多模态模型与长上下文思考\">\u003Ca class=\"anchor\" href=\"#原生多模态模型与长上下文思考\" aria-label=\"跳转到标题：原生多模态模型与长上下文思考\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>原生多模态模型与长上下文思考\u003C\u002Fh3>\u003Cul>\n\u003Cli>\u003Cstrong>原生多模态模型\u003C\u002Fstrong>：统一生成与理解的能力\u003C\u002Fli>\n\u003Cli>\u003Cstrong>长上下文窗口\u003C\u002Fstrong>：能够接收多种图片、多轮对话及复杂长文描述，实现对话式出图。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"数据生产与评估反馈\">\u003Ca class=\"anchor\" href=\"#数据生产与评估反馈\" aria-label=\"跳转到标题：数据生产与评估反馈\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>数据生产与评估反馈\u003C\u002Fh3>\u003Cul>\n\u003Cli>高质量数据的重要性：强调高质量数据对模型训练的重要性，特别是图文交织的训练方式。\u003C\u002Fli>\n\u003Cli>人类评估体系：人类美学评估、面部实用性评估等评估体系，以及文字渲染能力的量化评估方法。\u003C\u002Fli>\n\u003Cli>数据闭环与反馈：通过人类反馈形成数据制造闭环，提升模型生成准确性。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"开源多模态生成模型的挑战\">\u003Ca class=\"anchor\" href=\"#开源多模态生成模型的挑战\" aria-label=\"跳转到标题：开源多模态生成模型的挑战\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>开源多模态生成模型的挑战\u003C\u002Fh3>\u003Cul>\n\u003Cli>物理规律违背问题：开源多模态生成模型在生成内容时违背用户指令或物理规律的问题，如跑步机上反着跑的人、水杯破碎顺序错误等。\u003C\u002Fli>\n\u003Cli>理解能力的缺失：开源生成模型在理解人类意图方面不足。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"理解生成一体化技术脉络\">\u003Ca class=\"anchor\" href=\"#理解生成一体化技术脉络\" aria-label=\"跳转到标题：理解生成一体化技术脉络\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>理解生成一体化技术脉络\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw03.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CE40bTVeQrGXfAQ66zaoYRYRbak%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>图文对比学习：\u003C\u002Fstrong> CLIP 模型利用对比学习实现图像与文本对齐的技术。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>GPT4 与 Gemini 2.0：\u003C\u002Fstrong> GPT4 与 Gemini 2.0 等原生多模态模型在图文理解方面的强大能力。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>多模态生成技术：\u003C\u002Fstrong> Stable Diffusion、DALL·E 等模型在文生图方面的技术突破，以及 Flux 系列模型在指令遵循能力方面提升。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"理解生成一体化框架设计\">\u003Ca class=\"anchor\" href=\"#理解生成一体化框架设计\" aria-label=\"跳转到标题：理解生成一体化框架设计\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>理解生成一体化框架设计\u003C\u002Fh3>\u003Ch4 id=\"1-理想的一体化框架\">\u003Ca class=\"anchor\" href=\"#1-理想的一体化框架\" aria-label=\"跳转到标题：1. 理想的一体化框架\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 理想的一体化框架\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw04.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=b53PL8kj0jExsyWkVzSu8IhubCQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-现有理解生成技术\">\u003Ca class=\"anchor\" href=\"#2-现有理解生成技术\" aria-label=\"跳转到标题：2. 现有理解生成技术\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 现有理解生成技术\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw05.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WrLZJvoYb339HmR0dt3I3R4fFYs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>编解码结构设计\u003C\u002Fli>\n\u003Cli>视觉特征方式\u003C\u002Fli>\n\u003Cli>视觉表示特性\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-现有理解生成模型框架\">\u003Ca class=\"anchor\" href=\"#3-现有理解生成模型框架\" aria-label=\"跳转到标题：3. 现有理解生成模型框架\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 现有理解生成模型框架\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw06.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cj7MyVImeiDzy0J7u1KjnCSgyfY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch5 id=\"1-生成外挂-ad-qc-ls\">\u003Ca class=\"anchor\" href=\"#1-生成外挂-ad-qc-ls\" aria-label=\"跳转到标题：（1）生成外挂：AD-QC-LS\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>（1）生成外挂：AD-QC-LS\u003C\u002Fh5>\u003Cul>\n\u003Cli>优点：可快速使模型同时具有不错的理解和生成能力。\u003C\u002Fli>\n\u003Cli>缺点：参数量大；特征割裂，无充分融合。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch5 id=\"2-极致统一\">\u003Ca class=\"anchor\" href=\"#2-极致统一\" aria-label=\"跳转到标题：（2）极致统一\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>（2）极致统一\u003C\u002Fh5>\u003Cul>\n\u003Cli>\u003Cstrong>A-Q-L\u003C\u002Fstrong>\u003Cul>\n\u003Cli>优点：可快速嵌入 AR 框架。\u003C\u002Fli>\n\u003Cli>缺点：难以做高维度语义理解任务；表示空间少；词表固定。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A-C-L\u003C\u002Fstrong>\u003Cul>\n\u003Cli>优点：表示空间更丰富，细节生成更好。\u003C\u002Fli>\n\u003Cli>缺点：难以做高维度语义理解任务；可能出现上下文遗忘等问题。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A-Q-S\u003C\u002Fstrong>\u003Cul>\n\u003Cli>优点：一个视觉编码器完成两种视觉表示建模，同时提升理解和生成性能；离散后便于直接嵌入 AR 架构。\u003C\u002Fli>\n\u003Cli>缺点：表示空间少，细节和语义难以完全保持；词表固定，视觉编码器固定。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Cstrong>D-Q-L\u003C\u002Fstrong>\u003Cul>\n\u003Cli>优点：鉴于 LLaDA 的 Diffusion 解码框架，符合图像直观的解码方式，同时解码效率相对高。\u003C\u002Fli>\n\u003Cli>缺点：1. 基础模型仍在探索中，基础能力不足; 2. 仍局限于 VQ，受其本身能力局限。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw07.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=svuMDxN9rWt44n6k4UjIlqSLO%2FQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch5 id=\"3-融合建模\">\u003Ca class=\"anchor\" href=\"#3-融合建模\" aria-label=\"跳转到标题：（3）融合建模\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>（3）融合建模\u003C\u002Fh5>\u003Cul>\n\u003Cli>\u003Cstrong>A-Q-LS\u003C\u002Fstrong>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw08.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OCfdcvRVx0QQAFE7i%2BsP30Z0DLU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>A-QC-SL\u003C\u002Fstrong>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw09.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CQ4ZG3S%2Bf7J6DAK7g6bQb8wQj38%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>AD-C-L\u003C\u002Fstrong>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw10.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9biUB4m4y4zrudvOy4XyDbKOoh8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>AD-Q-SL\u003C\u002Fstrong>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw11.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Uitx0eONWWCPiMT2xIIKGtJAIf0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>AD-QC-SL\u003C\u002Fstrong>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw12.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QYzrGNwyR26dxHIITpfT2l9yicI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>AD-C-SL\u003C\u002Fstrong>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw13.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=UN3dXJqfcZy32t3GZ8uCbskvjHw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"deepmind-工作启发\">\u003Ca class=\"anchor\" href=\"#deepmind-工作启发\" aria-label=\"跳转到标题：DeepMind 工作启发\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>DeepMind 工作启发\u003C\u002Fh3>\u003Cp>从 DeepMind 相关的研究基础来看，有两点技术坚持：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>1.坚持利用连续特征 + Diffusion Loss 用于图像生成。\u003C\u002Fli>\n\u003Cli>2.是利用 AR 还是 AR+Diffusion 仍在探索，但基本上是在 LLM 内部完成一体化建模。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw14.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2NHYTwRsyNQPZh0ZvxnBChTSzGM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw15.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=jTnv8OQ436y%2Fwu1Bjsi2%2Fh3UYcE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_53_zw16.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=5QYcN0gfmgmzbS5%2FzjX%2FAdsZ5E4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_kcb.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=tmWtTrH27PV8kQApQCBd1QaHaNU%3D\" alt=\"2025冬季训练营课程表\">\u003C\u002Fp>\n\u003Ch3 id=\"学习贴士\">\u003Ca class=\"anchor\" href=\"#学习贴士\" aria-label=\"跳转到标题：💡 学习贴士\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>💡 学习贴士\u003C\u002Fh3>\u003Cp>更多细节及完整课程内容，请观看直播或查看课程回放。\u003C\u002Fp>\n\u003Cp>链接：\u003Ca href=\"https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\">https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\u003C\u002Fa>\u003C\u002Fp>\n\u003Ch3 id=\"训练营报名持续进行中\">\u003Ca class=\"anchor\" href=\"#训练营报名持续进行中\" aria-label=\"跳转到标题：📚 训练营报名持续进行中\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>📚 训练营报名持续进行中\u003C\u002Fh3>\u003Cp>扫码报名与加入群聊，或点击 \u003Cstrong>「阅读原文」\u003C\u002Fstrong> 直达官网报名\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_ewm.jpg?Expires=1792752825&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=O1NL5bahW5nvf%2FlqFbFObNY5fbM%3D\" alt=\"\">\u003C\u002Fp>\n",[128,131,133,135,137,139,141,145,148,151,154,157,160,163,166],{"id":129,"text":130,"level":9},"nano-banana-pro-技术","Nano Banana Pro 技术",{"id":132,"text":132,"level":9},"原生多模态模型与长上下文思考",{"id":134,"text":134,"level":9},"数据生产与评估反馈",{"id":136,"text":136,"level":9},"开源多模态生成模型的挑战",{"id":138,"text":138,"level":9},"理解生成一体化技术脉络",{"id":140,"text":140,"level":9},"理解生成一体化框架设计",{"id":142,"text":143,"level":144},"1-理想的一体化框架","1. 理想的一体化框架",4,{"id":146,"text":147,"level":144},"2-现有理解生成技术","2. 现有理解生成技术",{"id":149,"text":150,"level":144},"3-现有理解生成模型框架","3. 现有理解生成模型框架",{"id":152,"text":153,"level":45},"1-生成外挂-ad-qc-ls","（1）生成外挂：AD-QC-LS",{"id":155,"text":156,"level":45},"2-极致统一","（2）极致统一",{"id":158,"text":159,"level":45},"3-融合建模","（3）融合建模",{"id":161,"text":162,"level":9},"deepmind-工作启发","DeepMind 工作启发",{"id":164,"text":165,"level":9},"学习贴士","💡 学习贴士",{"id":167,"text":168,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]