{"id":860,"date":"2024-07-16T14:24:50","date_gmt":"2024-07-16T06:24:50","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=860"},"modified":"2024-07-25T15:16:11","modified_gmt":"2024-07-25T07:16:11","slug":"202407-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-renderdiffusion-image-diffusion-for-3d-reconstruction-inpainting-and-generation","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202407-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-renderdiffusion-image-diffusion-for-3d-reconstruction-inpainting-and-generation\/","title":{"rendered":"202407 \u8bba\u6587\u7814\u8bfb-RenderDiffusion: Image Diffusion for 3D Reconstruction, Inpainting and Generation"},"content":{"rendered":"\n

\u4f5c\u8005\uff1aTitas Anciukevi\u010dius\u3001Zexiang Xu\u3001Matthew Fisher
\u5355\u4f4d\uff1aUniversity of Edinburgh\u3001Adobe Research
\u6765\u6e90\uff1aCVPR
\u65f6\u95f4\uff1a2023<\/p>\n\n\n\n

\u4e00. \u80cc\u666f<\/h1>\n\n\n\n

\u76ee\u524d\u4e3a\u6b62\uff0c\u56fe\u50cf\u6269\u6563\u6a21\u578b\u4e0d\u652f\u6301 3D \u7406\u89e3\u6240\u9700\u7684\u4efb\u52a1\uff0c\u4f8b\u5982\u89c6\u56fe\u4e00\u81f4\u7684 3D \u751f\u6210\u6216\u5355\u89c6\u56fe\u5bf9\u8c61\u91cd\u5efa\u3002\u672c\u6587\u63d0\u51fa\u4e86 RenderDiffusion\uff0c\u8fd9\u662f\u7b2c\u4e00\u4e2a\u7528\u4e8e3D \u751f\u6210\u548c\u63a8\u7406\u7684\u6269\u6563\u6a21\u578b\uff0c\u4ec5\u4f7f\u7528\u5355\u76ee 2D \u76d1\u7763\u8fdb\u884c\u8bad\u7ec3\u3002\u672c\u6587\u65b9\u6cd5\u7684\u6838\u5fc3\u662f\u4e00\u79cd\u65b0\u9896\u7684\u56fe\u50cf\u53bb\u566a\u67b6\u6784\uff0c\u5b83\u5728\u6bcf\u4e2a\u53bb\u566a\u6b65\u9aa4\u4e2d\u751f\u6210\u5e76\u6e32\u67d3\u573a\u666f\u7684\u4e2d\u95f4\u4e09\u7ef4\u8868\u793a\u3002\u8fd9\u5728\u6269\u6563\u8fc7\u7a0b\u4e2d\u5f3a\u5316\u4e86\u5f3a\u5927\u7684\u5f52\u7eb3\u7ed3\u6784\uff0c\u63d0\u4f9b\u4e86 3D\u4e00\u81f4\u7684\u8868\u793a\uff0c\u540c\u65f6\u53ea\u9700\u8981 2D \u76d1\u7763\u3002\u751f\u6210\u7684 3D \u8868\u793a\u53ef\u4ee5\u4ece\u4efb\u4f55\u89c6\u56fe\u5448\u73b0\u3002\u6211\u4eec\u5728 FFHQ\u3001AFHQ\u3001ShapeNet \u548c CLEVR \u6570\u636e\u96c6\u4e0a\u8bc4\u4f30RenderDiffusion\uff0c\u663e\u793a\u4e86\u751f\u6210 3D \u573a\u666f\u4ee5\u53ca\u4ece 2D \u56fe\u50cf\u63a8\u65ad 3D \u573a\u666f\u7684\u7ade\u4e89\u6027\u80fd\u3002\u6b64\u5916\uff0c\u6211\u4eec\u57fa\u4e8e\u6269\u6563\u7684\u65b9\u6cd5\u5141\u8bb8\u6211\u4eec\u4f7f\u7528 2D \u4fee\u590d\u6765\u7f16\u8f91 3D \u573a\u666f\u3002<\/p>\n\n\n\n

\u4e8c. \u521b\u65b0<\/h1>\n\n\n\n

\u2460\u5355\u76ee\u4e09\u7ef4\u91cd\u5efa\u3002
\u2461\u4e09\u7ef4\u611f\u77e5\u4fee\u590d–\u5141\u8bb8\u4e8c\u7ef4\u4fee\u590d\u6765\u7f16\u8f91\u4e09\u7ef4\u573a\u666f\u3002
\u2462\u65e0\u6761\u4ef6\u751f\u6210\u3002<\/p>\n\n\n\n

\u4e09. \u7b97\u6cd5\u4e0e\u5b9e\u73b0<\/h1>\n\n\n\n

\u663e\u5f0f\u6f5c\u57283D\u8868\u793a\u7684\u53bb\u566a\u67b6\u6784<\/strong><\/h5>\n\n\n\n

\u5728\u6bcf\u4e2a\u53bb\u566a\u6b65\u9aa4\u4e2d\u751f\u6210\u5e76\u6e32\u67d3\u4e00\u4e2a\u573a\u666f\u7684\u4e2d\u95f43D\u8868\u793a\uff0c\u53ef\u4ee5\u5728\u6269\u6563\u8fc7\u7a0b\u4e2d\u6709\u4e00\u4e2a\u5f3a\u5f52\u7eb3\u7ed3\u6784\u3002
\u2460\u53ea\u4f7f\u75282D\u56fe\u50cf\u8bad\u7ec3\u4e14\u65e0\u660e\u786e\u76843D\u76d1\u7763\u3002
\u2461\u6f5c\u5728\u76843D\u8868\u793a\u5408\u5e76\u5728denoiser\u4e2d\uff0c\u4ea7\u751f\u4e00\u79cd\u5f52\u7eb3\u504f\u5dee\u3002
Latent 3D structure\uff1a
\uff081\uff09decoder(\u7531noise image\u521b\u5efa)
\uff082\uff09volumetric renderer(\u5c063d\u8868\u793a \u6e32\u67d3\u56de \u21922d denoised image)
\u4f18\u52bf\uff1a
\u907f\u514d\u4e86\u4f53\u79ef\u6570\u636e\u7684\u7acb\u65b9\u5185\u5b58\u589e\u957f\uff08\u76f4\u63a5\u5904\u74062d,\u800c\u65e0\u97003d\u76d1\u7763\uff09<\/p>\n\n\n\n

Model<\/strong><\/h5>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u4e0a\u56fe\u4e3a\u6a21\u578b\u603b\u4f53\u7684\u4f53\u7cfb\u7ed3\u6784\uff1a
\u5bf9\u6709\u566a\u58f0\u7684\u8f93\u5165\u56fe\u50cf\u5e94\u7528\u5230g_\u03b8\uff0c\u9010\u6b65\u53bb\u566a\u751f\u6210\u56fe\u50cf\u3002
\u53bb\u566a\u5668g_\u03b8\u4ee5\u4e09\u5e73\u9762\u8868\u793a\uff08P\uff09<\/p>\n\n\n\n

Image diffusion model<\/strong><\/h5>\n\n\n\n

\u57fa\u672c\u6d41\u7a0b\u548cDDPM\u4e00\u81f4\uff0c\u6709\u533a\u522b\u7684\u662f\u53bb\u566a\u5668g_\u03b8\u4e0d\u662f\u6765\u9884\u6d4b\u566a\u58f0(x_t\u2212x_0)\u800c\u662f\u6765\u9884\u6d4b\uff0c\u56e0\u6b64Train Loss\u4e3a\u5982\u4e0b\uff1a<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u5f53\u8bad\u7ec3\u5b8c\u6210\uff0c\u5728\u751f\u6210\u9636\u6bb5\u65f6\uff0c\u6a21\u578bg_\u03b8\u3002\u5c31\u53ef\u5c06u_t\u8fd1\u4f3c\u4e0e\u5982\u4e0b\uff1a
<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n
3D-Aware Denoiser<\/strong><\/h5>\n\n\n\n

Novel decoder structure:\u5f15\u5165\u6f5c\u5728\u76843D\u8868\u793a\uff0c\u66f4\u6539decoder\u7684\u4f53\u7cfb\u7ed3\u6784\uff0c\u4ee5\u4e0b\u662f\u52a0\u5165\u7684\u989d\u5916\u4e24\u4e2a\u7ec4\u4ef6:
triplane encoder e_\u03b8\u3002 \u4e09\u9762\u7f16\u7801\u5668:\u4f7f\u7528\u6444\u50cf\u673a\u89c6\u56fev\u6784\u6210\u7684\u8f93\u5165\u56fe\u50cfx_t\u21923d\u4e09\u9762\u8868\u793a\u3002
triplane renderer r_\u03a8 \u4e09\u9762\u6e32\u67d3\u5668:\u5c063d\u4e09\u9762\u8868\u793a\u2192\u6e32\u67d3\u56de\u53bb\u566a\u56fe\u50cfx_0
<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n
Triplane representation<\/strong><\/h5>\n\n\n\n

\u5206\u89e33D\u7279\u5f81\u7f51\u683c:\u4e09\u5e73\u9762\u8868\u793aP\u5c06\u4e00\u4e2a\u5b8c\u6574\u76843D\u7279\u5f81\u7f51\u683c\u5206\u89e3\u4e3a\u4e09\u4e2a(\u6cbf\u4e09\u4e2a\u5750\u6807\u5e73\u9762\u653e\u7f6e\u76842D feature map
\u5bf9P\u53cc\u7ebf\u6027\u62bd\u6837:\u5c06\u4efb\u610f\u7ed9\u5b9a\u76843D\u70b9p\u6295\u5f71\u5230\u6bcf\u4e2a\u5750\u6807\u5e73\u9762\uff0c\u5bf9\u6bcf\u4e2a\u7279\u5f81\u56fe\u4f5c\u53cc\u7ebf\u6027\u63d2\u503c\u5e76\u5c06\u4e09\u4e2a\u7ed3\u679c\u76f8\u52a0\uff0c\u6700\u540e\u5f97\u5230\u4e00\u4e2asize\u4e3an_f\u7684\u5355\u7279\u5f81\u5411\u91cf\u3002
Triplane encoder\uff1a
e_\u03b8\u5c06MxMx3\u7684\u8f93\u5165\u56fe\u50cfx_t\u2192NxNx3n_f \u7684\u4e09\u5e73\u9762\u8868\u793a(N\u2265M)
U-Net\u4f53\u7cfb\u7ed3\u6784\u4f5c\u4e3abasis,\u4f46\u5728\u4e09\u5e73\u9762\u5927\u5c0f\u7684\u8f93\u51fa feature maps\u4e2d\u6dfb\u52a0\u989d \u5916\u7684\u5c42\u3002
Triplane renderer:
r_\u03a8\u4f7f\u7528\u4e0a\u8ff0\u7684\u4e09\u5e73\u9762\u7279\u5f81\u8fdb\u884c\u4f53\u6e32\u67d3\uff0c\u8f93\u51faMxMx3\u7684\u56fe\u50cfxt-1
\u5bf9\u6bcf\u4e2a3D\u91c7\u6837\u70b9\uff0c\u6cbf\u7740\u4ece\u56fe\u50cf\u6295\u5c04\u7684\u5c04\u7ebf\u5f97\u5230density y<\/strong>and color c\uff1a
<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u50cf\u7d20\u7684\u6700\u7ec8\u989c\u8272\u662f\u901a\u8fc7\u6cbf\u7740\u5c04\u7ebf\u96c6\u6210density yand color c\u4ea7\u751f\u7684(\u4f7f\u7528MipNeRF\u76f8\u540c\u7684\u663e\u5f0f\u4f53\u7ed8\u5236\u65b9\u6cd5)\uff0c\u4f7f\u7528NERF\u7684\u4e24\u6b21\u91cd\u8981\u6027\u91c7\u6837:
\u7b2c\u4e00\u6b21\u4f20\u9012\u4f7f\u7528\u5206\u5c42\u62bd\u6837\uff0c\u6cbf\u7740\u6bcf\u6761\u5c04\u7ebf\u653e\u7f6e\u6837\u672c
\u5bf9\u7b2c\u4e00\u6b21\u4f20\u9012\u7684\u7ed3\u679c\u8fdb\u884c\u62bd\u6837<\/p>\n\n\n\n

\u7ed3\u679c<\/strong><\/h5>\n\n\n\n
\"\"<\/figure>\n\n\n\n

3D \u611f\u77e5\u4fee\u590d\u3002\u672c\u6587\u4ece\u4e24\u4e2a\u4e0d\u540c\u7684\u566a\u58f0\u6837\u672c\u5f00\u59cb\u4fee\u590d\u5de6\u4fa7\u663e\u793a\u7684\u566a\u58f0 2D \u533a\u57df\uff0c\u4ece\u800c\u4ea7\u751f\u4e24\u4e2a\u4e0d\u540c\u7684\u4fee\u590d 3D \u573a\u666f\u3002\u672c\u6587\u8fd8\u5c55\u793a\u4e86\u7b2c\u4e8c\u4e2a\u573a\u666f\u7684\u65b0\u9896\u89c6\u89d2\u3002<\/p>\n\n\n\n

\u56db. \u603b\u7ed3<\/h1>\n\n\n\n

\u2460\u7b2c\u4e00\u4e2a\u53ef\u4ee5\u5229\u7528posed 2D-images\u800c\u975e3D\u76d1\u7763\u76843D\u6269\u6563\u6a21\u578b\u3002
\u2461\u65b0\u7684\u53bb\u566a\u67b6\u6784:triplane\u6e32\u67d3\u6a21\u578b\uff0c\u52a0\u5f3a\u4e86\u5f52\u7eb3\u504f\u5dee\uff0c\u4ea7\u751f3D\u573a\u666f\u4e00\u81f4\u6027\u3002
\u4f7f\u7528\u60c5\u666f:
2d image \u21923d \u573a\u666f
2d inpainting \u21923d\u7f16\u8f91
3d\u573a\u666f\u751f\u6210
\u5c40\u9650\u6027:
\u8981\u6c42\u8bad\u7ec3\u56fe\u50cf\u6709\u5df2\u77e5\u7684\u76f8\u673a\u53c2\u6570\u3002
\u4e09\u5e73\u9762\u9700\u8981\u88ab\u653e\u7f6e\u5230\u4e00\u4e2a\u5168\u5c40\u5750\u6807\u7cfb\u4e2d\uff0c\u9650\u5236\u4e86\u8de8\u5bf9\u8c61\u653e\u7f6e\u7684\u6cdb\u5316\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u4f5c\u8005\uff1aTitas Anciukevi\u010dius\u3001Zexiang Xu\u3001Matthew Fisher\u5355\u4f4d\uff1aUniv […]<\/p>\n","protected":false},"author":5,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-860","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/860","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=860"}],"version-history":[{"count":3,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/860\/revisions"}],"predecessor-version":[{"id":1023,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/860\/revisions\/1023"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=860"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=860"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=860"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}