[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"product:pixal3d":3},{"id":4,"name":5,"slug":6,"slogan":7,"description":8,"organization":9,"images":10,"avatar_url":11,"image_url":12,"website_url":13,"access_links":14,"status":19,"build_by":20,"categories":21,"stats":25,"upload_relationship":30,"create_at":31,"update_at":32,"owner_id":27,"upload_user_id":33,"reviewer_id":34,"video_info":35,"prerelease_cover_url":10,"scheduled_publish_at":35,"user_scheduled":36,"emoji_reacts":37,"tag":35,"pass_review":35,"user_star":36},1993,"Pixal3D","pixal3d","多视角图像转 3D 资产生成平台。","{\"type\": \"doc\", \"content\": [{\"type\": \"paragraph\", \"content\": [{\"text\": \"【产品介绍】\", \"type\": \"text\"}]}, {\"type\": \"bulletList\", \"content\": [{\"type\": \"listItem\", \"content\": [{\"type\": \"paragraph\", \"content\": [{\"text\": \"Pixal3D 是由清华大学、腾讯 ARC 实验室及惠灵顿维多利亚大学联合研发的下一代图像转 3D 生成算法模型。\", \"type\": \"text\"}]}]}, {\"type\": \"listItem\", \"content\": [{\"type\": \"paragraph\", \"content\": [{\"text\": \"该项目针对现有 3D 原生生成器在规范空间中生成时导致 2D-3D 对应关系模糊、无法达到像素级精确契合的行业难题进行了攻关。\", \"type\": \"text\"}]}]}, {\"type\": \"listItem\", \"content\": [{\"type\": \"paragraph\", \"content\": [{\"text\": \"平台引入创新的“像素对齐”生成范式，使生成的 3D 资产在视点上与原始输入图像保持强一致性，将 3D 生成的保真度提升至接近 3D 重建的级别。\", \"type\": \"text\"}]}]}, {\"type\": \"listItem\", \"content\": [{\"type\": \"paragraph\", \"content\": [{\"text\": \"该研究成果已成功入选 SIGGRAPH 2026，为高精度数字孪生、游戏资产创建及多视角 3D 场景合成提供了全新的底层技术支撑。\", \"type\": \"text\"}]}]}]}, {\"type\": \"paragraph\"}, {\"type\": \"paragraph\", \"content\": [{\"text\": \"【产品功能】\", \"type\": \"text\"}]}, {\"type\": \"bulletList\", \"content\": [{\"type\": \"listItem\", \"content\": [{\"type\": \"paragraph\", \"content\": [{\"text\": \"像素对齐表征学习：利用定制的 VAE 压缩技术，将像素对齐的稀疏有向距离函数（SDF）转化为高效的稀疏潜变量，奠定高精度几何解码基础。\", \"type\": \"text\"}]}]}, {\"type\": \"listItem\", \"content\": [{\"type\": \"paragraph\", \"content\": [{\"text\": \"图像反向投影调理：引入反向投影机制，明确将多尺度 2D 图像特征提升至 3D 特征体中，从根本上消除了像素到 3D 空间映射的歧义。\", \"type\": \"text\"}]}]}, {\"type\": \"listItem\", \"content\": [{\"type\": \"paragraph\", \"content\": [{\"text\": \"双阶段高保真生成：通过双阶段生成式流水线，先期预测物体的粗糙结构，随后精确填充细节潜变量，最终解码出高保真的 3D 纹理网格。\", \"type\": \"text\"}]}]}, {\"type\": \"listItem\", \"content\": [{\"type\": \"paragraph\", \"content\": [{\"text\": \"跨视角特征聚合：原生支持多视角扩展，能够将来自不同视角的反向投影特征体进行深度融合，生成更为完整的 3D 资产。\", \"type\": \"text\"}]}]}, {\"type\": \"listItem\", \"content\": [{\"type\": \"paragraph\", \"content\": [{\"text\": \"物体解耦场景合成：提供模块化流水线，支持直接从场景图像中全自动合成出高保真、且各个物体相互独立分离的复杂 3D 场景。\", \"type\": \"text\"}]}]}]}, {\"type\": \"paragraph\"}]}","清华大学 & 腾讯 ARC 实验室","","https:\u002F\u002Fwatcha.tos-cn-beijing.volces.com\u002Fprod\u002Fuser\u002Fuploads\u002F10009357_1778948269_a6f048b3-ed1d-4499-a340-d765d36fbed2.png","https:\u002F\u002Fwatcha.tos-cn-beijing.volces.com\u002Fprod\u002Fuser\u002Fuploads\u002F10009357_1778948222_db5a3c6f-c069-455b-8e51-56b2185e9438.png","https:\u002F\u002Fldyang694.github.io\u002Fprojects\u002Fpixal3d\u002F",{"items":15},[16],{"platform":17,"url":13,"is_primary":18},"WEB",true,"PUBLISHED",[],[22],{"id":23,"name":24},12,"其他类型",{"upvotes":26,"stars":26,"review_count":26,"reply_count":27,"score":28,"update_at":29},1,0,2.0654329147389294,"2026-09-08T08:05:15.674Z","THIRD_PARTY","2026-05-16T16:17:56.371Z","2026-05-16T16:36:14.903Z",10009357,2,null,false,{"reacts":38},[]]