{"id":2117,"date":"2025-01-14T00:00:00","date_gmt":"2025-01-14T00:00:00","guid":{"rendered":"urn:uuid:8b3e41f8-8bfc-4e24-b8af-4991d67a8d8d"},"modified":"2025-01-14T00:00:00","modified_gmt":"2025-01-14T00:00:00","slug":"gong-kai-sareteirusheng-cheng-aihurokuramu-stable-diffusion","status":"publish","type":"post","link":"https:\/\/www.sekaiken.com\/?p=2117","title":{"rendered":"\u516c\u958b\u3055\u308c\u3066\u3044\u308b\u751f\u6210AI\u30d7\u30ed\u30b0\u30e9\u30e0 Stable Diffusion"},"content":{"rendered":"<p>Stability AI\u793e\uff08\u3082\u3068\u306f\u30df\u30e5\u30f3\u30d8\u30f3\u5927\u5b66\uff09\u306e\u30d7\u30ed\u30b0\u30e9\u30e0 Stable Diffusion\u3092\u898b\u3066\u307f\u307e\u3057\u3087\u3046\u3002\u30bd\u30fc\u30b9\u30b3\u30fc\u30c9\u3068\u57fa\u790e\u8a13\u7df4\u30c7\u30fc\u30bf\u3092\u516c\u958b\u3057\u3066\u3044\u308b\u306e\u304c\u5927\u5b66\u3089\u3057\u3044\u3068\u8a00\u3048\u307e\u3059\u3002\u6211\u3005\u304c\u52c9\u5f37\u3059\u308b\u306e\u306b\u52a9\u3051\u306b\u306a\u308a\u307e\u3059\u3002<br \/>\nhttps:\/\/github.com\/Stability-AI\/generative-models<br \/>\n\u5c55\u958b\u3057\u305f\u30d5\u30a1\u30a4\u30eb\u306e\u5927\u304d\u3055\u306f2022\u5e74\u306e\u6bb5\u968e\u3067\u306f\u6700\u521d\u306f5-6GB\u3067\u3001\u8a13\u7df4\u3067\u3060\u3093\u3060\u3093\u5927\u304d\u304f\u306a\u308b\u305d\u3046\u3067\u3059\u3002\u30d1\u30e9\u30e1\u30fc\u30bf\u306f\u6700\u521d\u306f10\u5104\u500b\u3067\u300120\u5104\u306e\u753b\u50cf\u3068\u30c6\u30ad\u30b9\u30c8\u306e\u30da\u30a2\u3092\u5b66\u7fd2\u3057\u3066\u5f97\u3089\u308c\u3066\u3044\u308b\u305d\u3046\u3067\u3059\u3002<br \/>\n\u3082\u3068\u3082\u3068\u306f\u3001\u30ce\u30a4\u30ba\u304c\u5165\u3063\u305f\u308a\u30d4\u30f3\u30dc\u30b1\u3057\u305f\u308a\u3057\u305f\u753b\u50cf\u3092\u4fee\u5fa9\u3059\u308b\u305f\u3081\u306e\u30bd\u30d5\u30c8\u3092\u958b\u767a\u3057\u3066\u3044\u307e\u3057\u305f\u3002\u3053\u3046\u3044\u3046\u306e\u306f\u62e1\u6563\u65b9\u7a0b\u5f0f\uff08diffusion equation\uff09\u306e\u9006\u554f\u984c\u306b\u4f3c\u3066\u3044\u3066\u3001\u6570\u5b66\u7684\u306b\u306f\u8a08\u7b97\u304c\u96e3\u3057\u3044\u3053\u3068\u304c\u77e5\u3089\u308c\u3066\u3044\u307e\u3057\u305f\u3002\u305d\u3053\u306b\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3092\u4f7f\u3063\u305f\u5b66\u7fd2\u3067\u3001\u300c\u3053\u3046\u3044\u3046\u307c\u3051\u305f\u753b\u50cf\u306a\u3089\u5143\u306f\u3053\u3046\u3060\u308d\u3046\u300d\u3068\u3044\u3046\u63a8\u6e2c\u3092\u3055\u305b\u308b\u65b9\u6cd5\u3092\u7814\u7a76\u3057\u3066\u3044\u3066\u3001\u30d2\u30f3\u30c8\u3068\u3057\u3066\u8a00\u8449\u3092\u4e0e\u3048\u308b\u65b9\u5411\u306b\u9032\u3093\u3067\u3001\u5b9f\u306f\u8a00\u8449\u3092\u597d\u304d\u306a\u3088\u3046\u306b\u7d44\u307f\u5408\u308f\u305b\u308c\u3070\u7d75\u3092\u7d44\u307f\u5408\u308f\u305b\u3089\u308c\u308b\u3068\u3044\u3046\u3053\u3068\u304c\u308f\u304b\u3063\u305f\u3001\u3068\u3044\u3046\u306e\u304c\u3054\u304f\u7c21\u5358\u306a\u8aac\u660e\u3067\u3059\u3002<br \/>\n\u4f7f\u3063\u3066\u307f\u308b\u3068\u3001\u6614\u306e\u30c6\u30ec\u30d3\u3067\u96fb\u6ce2\u304c\u5f31\u3044\u6642\u306b\u51fa\u305f\u300c\u7802\u5d50\u300d\u306e\u3088\u3046\u306a\u753b\u50cf\u304b\u3089\u3060\u3093\u3060\u3093\u306f\u3063\u304d\u308a\u3057\u3066\u304f\u308b\u69d8\u5b50\u304c\u898b\u3048\u307e\u3059\u3002\u4f8b\u3048\u3070\u4e0b\u8a18\u753b\u50cf\u3067\u3059\u3002<br \/>\nhttps:\/\/en.wikipedia.org\/wiki\/Stable_Diffusion#\/media\/File:X-Y_plot_of_algorithmically-generated_AI_art_of_European-style_castle_in_Japan_demonstrating_DDIM_diffusion_steps.png<br \/>\n\u539f\u7406\u306b\u3064\u3044\u3066\u306f\u4e0b\u8a18\u304c\u308f\u304b\u308a\u3084\u3059\u3044\u3067\u3059\u3002<br \/>\nhttps:\/\/qiita.com\/omiita\/items\/ecf8d60466c50ae8295b<br \/>\n\u5b66\u7fd2\uff08AI\u306e\u8a13\u7df4\uff1d10\u5104\u500b\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u6c7a\u5b9a\uff09\u3092\u3069\u3046\u3059\u308b\u304b\u3068\u3044\u3046\u3068\u3001\u8a00\u8449\u3068\u753b\u50cf\u3092\u4e0e\u3048\u3066\u3001\u753b\u50cf\u306b1000\u6bb5\u968e\u3067\u5c11\u3057\u305a\u3064\u30ce\u30a4\u30ba\u3092\u52a0\u3048\u3066\u300c\u7802\u5d50\u300d\u306b\u3057\u306a\u304c\u3089\u305d\u306e\u9664\u53bb\u306e\u4ed5\u65b9\u3092AI\u306b\u899a\u3048\u3055\u305b\u307e\u3059\u3002\u5177\u4f53\u7684\u306b\u306f\u3001\u52a0\u3048\u305f\u30ce\u30a4\u30ba\u306f\u308f\u304b\u3063\u3066\u3044\u308b\u306e\u3067\u30011000\u6bb5\u968e\u3067\u305d\u308c\u3092\u6c42\u3081\u3055\u305b\u3001\u6b63\u89e3\u3068\u6bd4\u8f03\u3057\u3066\u6b63\u89e3\u304c\u5f97\u3089\u308c\u308b\u3088\u3046\u306b\u8abf\u6574\u3057\u307e\u3059\u3002\u305d\u306e\u904e\u7a0b\u306710\u5104\u500b\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u4e2d\u306b\u753b\u50cf\u3068\u8a00\u8449\u306e\u60c5\u5831\u304c\u57cb\u3081\u8fbc\u307e\u308c\u308b\uff08\u300c\u6f5c\u5728\u7a7a\u9593\u300d\u3068\u540d\u4ed8\u3051\u3089\u308c\u3066\u3044\u308b\uff09\u305f\u3081\u3001\u5fdc\u7528\u304c\u52b9\u304f\u3088\u3046\u306b\u306a\u308b\u305d\u3046\u3067\u3059\u3002\u3053\u308c\u306f\u4eba\u9593\u306e\u8133\u304c\u4eba\u306e\u9854\u306a\u3069\u306e\u753b\u50cf\u3092\u899a\u3048\u308b\u3068\u304d\u3082\u540c\u3058\u3088\u3046\u306a\u3053\u3068\u3092\u3084\u3063\u3066\u3044\u308b\u306e\u3067\u3057\u3087\u3046\u306d\u3002\u4eba\u306e\u540d\u524d\u3068\u7279\u5fb4\u3092\u6f5c\u5728\u7684\u306b\u899a\u3048\u3066\u3044\u308b\u308f\u3051\u3067\u3059\u3002\u305f\u3057\u304b\u306b\u3001\u5927\u767a\u660e\u3060\u3068\u601d\u3044\u307e\u3059\u3002<br \/>\n\u8b1b\u7fa9\u306e\u52d5\u753b\u3082\u7d39\u4ecb\u3055\u308c\u3066\u3044\u307e\u3057\u305f\u3002<br \/>\n<iframe loading=\"lazy\" title=\"\u3010Deep Learning\u7814\u4fee\uff08\u767a\u5c55\uff09\u3011\u30c7\u30fc\u30bf\u751f\u6210\u30fb\u5909\u63db\u306e\u305f\u3081\u306e\u6a5f\u68b0\u5b66\u7fd2\u3000\u7b2c\uff17\u56de\u524d\u7de8\u300cDiffusion models\u300d\" width=\"500\" height=\"281\" src=\"https:\/\/www.youtube.com\/embed\/10ki2IS55Q4?feature=oembed\" frameborder=\"0\" allow=\"accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share\" referrerpolicy=\"strict-origin-when-cross-origin\" allowfullscreen><\/iframe><br \/>\n\u5f53\u7136\u3001\u601d\u3063\u3066\u3044\u308b\u306e\u3068\u9055\u3046\u3082\u306e\u3092\u300c\u5fa9\u5143\u300d\u3059\u308b\u3053\u3068\u3082\u3042\u308b\u306f\u305a\u306a\u306e\u3067\u3001\u30a4\u30e9\u30b9\u30c8\u306e\u5834\u5408\u306a\u3089\u3070\u3001\u3044\u304f\u3064\u3082\u5019\u88dc\u3092\u4f5c\u3089\u305b\u3066\u4eba\u9593\u304c\u3044\u3044\u306e\u3092\u9078\u3076\u3068\u3044\u3046\u65b9\u6cd5\u3092\u3068\u308a\u307e\u3059\u3002<\/p>\n<p>\u82f1\u8a9e\u306f\u3001\u3000https:\/\/en.wikipedia.org\/wiki\/Stable_Diffusion\u3000\u304b\u3089\u3002<br \/>\n\u201dIntroduced in 2015, diffusion models are trained with the objective of removing successive applications of Gaussian noise on training images, which can be thought of as a sequence of denoising autoencoders.&rdquo;<br \/>\nobjective of &hellip; ~\u3092\u76ee\u7684\u3068\u3057\u3066<br \/>\nsuccessive applications of ~ \u3092\u9806\u7e70\u308a\u306b\u9069\u7528\u3059\u308b<br \/>\nGauusian noise \u30ac\u30a6\u30b7\u30a2\u30f3\u30ce\u30a4\u30ba\u3000\u6b63\u898f\u5206\u5e03\u306b\u5f93\u3046\u30ce\u30a4\u30ba<br \/>\ndenoising\u3000\u30ce\u30a4\u30ba\u9664\u53bb<br \/>\nautoencoder \u30aa\u30fc\u30c8\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\u3000\u3053\u308c\u306f\u6a5f\u68b0\u5b66\u7fd2\u3092\u7fd2\u3046\u3068\u6700\u521d\u306b\u51fa\u3066\u304d\u3066\u4f55\u306e\u305f\u3081\u306b\u3084\u308b\u306e\u304b\u6700\u521d\u306f\u5168\u304f\u308f\u304b\u308a\u307e\u305b\u3093\u304c\u3001\u7c21\u5358\u306b\u8a00\u3046\u3068\u60c5\u5831\u91cf\u3092\u6b63\u3057\u304f\u5727\u7e2e\u3059\u308b\u305f\u3081\u306e\u4ed5\u7d44\u307f\u3067\u3059\u3002<br \/>\n&ldquo;Stable Diffusion consists of 3 parts: the variational autoencoder (VAE), U-Net, and an optional text encoder.<br \/>\nThe VAE encoder compresses the image from pixel space to a smaller dimensional latent space, capturing a more fundamental semantic meaning of the image.&rdquo;<br \/>\nvariational \u5909\u5206\uff08\u3078\u3093\u3076\u3093\uff09\u3000\u6570\u5b66\u7528\u8a9e\u3067\u3059\u3002<br \/>\nlatent space \u6f5c\u5728\u7a7a\u9593\u3000\u3000cf. latent heat \u6f5c\u71b1<br \/>\nsemantic \u610f\u5473\u8ad6\u7684\u306a<br \/>\n&ldquo;Gaussian noise is iteratively applied to the compressed latent representation during forward diffusion.&rdquo;<br \/>\niteratively \u300c\u30a4\u300d\u30c6\u30e9\u30c6\u30a3\u30f4\u308a\u3000\u7e70\u308a\u8fd4\u3057\u3000\u3000iterate \u300c\u30a4\u300d\u30c6\u30ec\u30a4\u30c8\u3000\u7e70\u308a\u8fd4\u3059<br \/>\n&ldquo;The U-Net block, composed of a ResNet backbone, denoises the output from forward diffusion backwards to obtain a latent representation.&rdquo; &ldquo;Finally, the VAE decoder generates the final image by converting the representation back into pixel space.&rdquo;<br \/>\nrepresentation \u8868\u73fe<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Stability AI\u793e\uff08\u3082\u3068\u306f\u30df\u30e5\u30f3\u30d8\u30f3\u5927\u5b66\uff09\u306e\u30d7\u30ed\u30b0\u30e9\u30e0 Stable Diffusion\u3092\u898b\u3066\u307f\u307e\u3057\u3087\u3046\u3002\u30bd\u30fc\u30b9\u30b3\u30fc\u30c9\u3068\u57fa\u790e\u8a13\u7df4\u30c7\u30fc\u30bf\u3092\u516c\u958b\u3057\u3066\u3044\u308b\u306e\u304c\u5927\u5b66\u3089\u3057\u3044\u3068\u8a00\u3048\u307e\u3059\u3002\u6211\u3005\u304c\u52c9\u5f37\u3059\u308b\u306e\u306b\u52a9\u3051\u306b\u306a\u308a\u307e\u3059\u3002 https:\/\/github.com\/Stability-AI\/generative-models \u5c55\u958b\u3057\u305f\u30d5\u30a1\u30a4\u30eb\u306e\u5927\u304d\u3055\u306f2022\u5e74\u306e\u6bb5\u968e\u3067\u306f\u6700\u521d\u306f5-6GB\u3067\u3001\u8a13\u7df4\u3067\u3060\u3093\u3060\u3093\u5927\u304d\u304f\u306a\u308b\u305d\u3046\u3067\u3059\u3002\u30d1\u30e9\u30e1\u30fc\u30bf\u306f\u6700\u521d\u306f10\u5104\u500b\u3067\u300120\u5104\u306e\u753b\u50cf\u3068\u30c6\u30ad\u30b9\u30c8\u306e\u30da\u30a2\u3092\u5b66\u7fd2\u3057\u3066\u5f97\u3089\u308c\u3066\u3044\u308b\u305d\u3046\u3067\u3059\u3002 \u3082\u3068\u3082\u3068\u306f\u3001\u30ce\u30a4\u30ba\u304c\u5165\u3063\u305f\u308a\u30d4\u30f3\u30dc\u30b1\u3057\u305f\u308a\u3057\u305f\u753b\u50cf\u3092\u4fee\u5fa9\u3059\u308b\u305f\u3081\u306e\u30bd\u30d5\u30c8\u3092\u958b\u767a\u3057\u3066\u3044\u307e\u3057\u305f&hellip;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":""},"categories":[60,42],"tags":[23,5],"class_list":["post-2117","post","type-post","status-publish","format-standard","hentry","category-computer","category-tech","tag-computer","tag-tech"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/www.sekaiken.com\/index.php?rest_route=\/wp\/v2\/posts\/2117","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.sekaiken.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.sekaiken.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.sekaiken.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.sekaiken.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2117"}],"version-history":[{"count":0,"href":"https:\/\/www.sekaiken.com\/index.php?rest_route=\/wp\/v2\/posts\/2117\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.sekaiken.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2117"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.sekaiken.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2117"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.sekaiken.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2117"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}