{"id":479505,"date":"2023-08-09T10:41:18","date_gmt":"2023-08-09T10:41:18","guid":{"rendered":""},"modified":"2023-09-05T11:18:58","modified_gmt":"2023-09-05T11:18:58","slug":"vector-quantized-generative-adversarial-network-vqgan","status":"publish","type":"wiki","link":"https:\/\/oneproxy.pro\/jp\/wiki\/vector-quantized-generative-adversarial-network-vqgan\/","title":{"rendered":"\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u751f\u6210\u6575\u5bfe\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN)"},"content":{"rendered":"<p>\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u751f\u6210\u6575\u5bfe\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN) \u306f\u3001\u751f\u6210\u6575\u5bfe\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (GAN) \u3068\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316 (VQ) \u3068\u3044\u3046 2 \u3064\u306e\u4e00\u822c\u7684\u306a\u6a5f\u68b0\u5b66\u7fd2\u624b\u6cd5\u306e\u8981\u7d20\u3092\u7d44\u307f\u5408\u308f\u305b\u305f\u3001\u9769\u65b0\u7684\u3067\u5f37\u529b\u306a\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0 \u30e2\u30c7\u30eb\u3067\u3059\u3002VQGAN \u306f\u3001\u9ad8\u54c1\u8cea\u3067\u4e00\u8cab\u6027\u306e\u3042\u308b\u753b\u50cf\u3092\u751f\u6210\u3067\u304d\u308b\u305f\u3081\u3001\u4eba\u5de5\u77e5\u80fd\u7814\u7a76\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3067\u5927\u304d\u306a\u6ce8\u76ee\u3092\u96c6\u3081\u3066\u304a\u308a\u3001\u753b\u50cf\u5408\u6210\u3001\u30b9\u30bf\u30a4\u30eb\u8ee2\u9001\u3001\u30af\u30ea\u30a8\u30a4\u30c6\u30a3\u30d6 \u30b3\u30f3\u30c6\u30f3\u30c4\u751f\u6210\u306a\u3069\u3001\u3055\u307e\u3056\u307e\u306a\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u306b\u6709\u671b\u306a\u30c4\u30fc\u30eb\u3068\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<h2>\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN) \u306e\u8d77\u6e90\u3068\u305d\u306e\u6700\u521d\u306e\u8a00\u53ca\u306e\u6b74\u53f2\u3002<\/h2>\n<p>GAN \u306e\u6982\u5ff5\u306f\u30012014 \u5e74\u306b Ian Goodfellow \u6c0f\u3068\u305d\u306e\u540c\u50da\u306b\u3088\u3063\u3066\u521d\u3081\u3066\u5c0e\u5165\u3055\u308c\u307e\u3057\u305f\u3002GAN \u306f\u3001\u30b8\u30a7\u30cd\u30ec\u30fc\u30bf\u30fc\u3068\u30c7\u30a3\u30b9\u30af\u30ea\u30df\u30cd\u30fc\u30bf\u30fc\u306e 2 \u3064\u306e\u30cb\u30e5\u30fc\u30e9\u30eb \u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3067\u69cb\u6210\u3055\u308c\u308b\u751f\u6210\u30e2\u30c7\u30eb\u3067\u3042\u308a\u3001\u30df\u30cb\u30de\u30c3\u30af\u30b9 \u30b2\u30fc\u30e0\u3092\u30d7\u30ec\u30a4\u3057\u3066\u73fe\u5b9f\u7684\u306a\u5408\u6210\u30c7\u30fc\u30bf\u3092\u751f\u6210\u3057\u307e\u3059\u3002GAN \u306f\u753b\u50cf\u751f\u6210\u306b\u304a\u3044\u3066\u512a\u308c\u305f\u7d50\u679c\u3092\u793a\u3057\u3066\u3044\u307e\u3059\u304c\u3001\u30e2\u30fc\u30c9\u5d29\u58ca\u3084\u751f\u6210\u3055\u308c\u305f\u51fa\u529b\u306e\u5236\u5fa1\u306e\u6b20\u5982\u306a\u3069\u306e\u554f\u984c\u306b\u60a9\u307e\u3055\u308c\u308b\u3053\u3068\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n<p>2020 \u5e74\u3001DeepMind \u306e\u7814\u7a76\u8005\u306f\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u5909\u5206\u30aa\u30fc\u30c8\u30a8\u30f3\u30b3\u30fc\u30c0 (VQ-VAE) \u30e2\u30c7\u30eb\u3092\u767a\u8868\u3057\u307e\u3057\u305f\u3002VQ-VAE \u306f\u5909\u5206\u30aa\u30fc\u30c8\u30a8\u30f3\u30b3\u30fc\u30c0 (VAE) \u30e2\u30c7\u30eb\u306e\u30d0\u30ea\u30a8\u30fc\u30b7\u30e7\u30f3\u3067\u3042\u308a\u3001\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u3092\u7d44\u307f\u8fbc\u3093\u3067\u5165\u529b\u30c7\u30fc\u30bf\u306e\u96e2\u6563\u7684\u304b\u3064\u30b3\u30f3\u30d1\u30af\u30c8\u306a\u8868\u73fe\u3092\u751f\u6210\u3057\u307e\u3059\u3002\u3053\u308c\u306f\u3001VQGAN \u306e\u958b\u767a\u306b\u5411\u3051\u305f\u91cd\u8981\u306a\u30b9\u30c6\u30c3\u30d7\u3067\u3057\u305f\u3002<\/p>\n<p>\u305d\u306e\u5f8c\u3001\u540c\u5e74\u3001Ali Razavi \u7387\u3044\u308b\u7814\u7a76\u8005\u30b0\u30eb\u30fc\u30d7\u304c VQGAN \u3092\u767a\u8868\u3057\u307e\u3057\u305f\u3002\u3053\u306e\u30e2\u30c7\u30eb\u306f\u3001GAN \u306e\u30d1\u30ef\u30fc\u3068 VQ-VAE \u306e\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6280\u8853\u3092\u7d44\u307f\u5408\u308f\u305b\u3066\u3001\u54c1\u8cea\u3001\u5b89\u5b9a\u6027\u3001\u5236\u5fa1\u6027\u304c\u5411\u4e0a\u3057\u305f\u753b\u50cf\u3092\u751f\u6210\u3057\u307e\u3059\u3002VQGAN \u306f\u3001\u751f\u6210\u30e2\u30c7\u30eb\u306e\u5206\u91ce\u306b\u304a\u3051\u308b\u753b\u671f\u7684\u306a\u9032\u6b69\u3068\u306a\u308a\u307e\u3057\u305f\u3002<\/p>\n<h2>\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN) \u306b\u95a2\u3059\u308b\u8a73\u7d30\u60c5\u5831\u3002\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN) \u306e\u30c8\u30d4\u30c3\u30af\u3092\u62e1\u5f35\u3057\u307e\u3059\u3002<\/h2>\n<h3>\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\uff08VQGAN\uff09\u306e\u4ed5\u7d44\u307f<\/h3>\n<p>VQGAN \u306f\u3001\u5f93\u6765\u306e GAN \u3068\u540c\u69d8\u306b\u3001\u30b8\u30a7\u30cd\u30ec\u30fc\u30bf\u30fc\u3068\u30c7\u30a3\u30b9\u30af\u30ea\u30df\u30cd\u30fc\u30bf\u30fc\u3067\u69cb\u6210\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u30b8\u30a7\u30cd\u30ec\u30fc\u30bf\u30fc\u306f\u30e9\u30f3\u30c0\u30e0\u306a\u30ce\u30a4\u30ba\u3092\u5165\u529b\u3068\u3057\u3066\u53d7\u3051\u53d6\u308a\u3001\u30ea\u30a2\u30eb\u306a\u753b\u50cf\u3092\u751f\u6210\u3057\u3088\u3046\u3068\u3057\u307e\u3059\u3002\u4e00\u65b9\u3001\u30c7\u30a3\u30b9\u30af\u30ea\u30df\u30cd\u30fc\u30bf\u30fc\u306f\u3001\u5b9f\u969b\u306e\u753b\u50cf\u3068\u751f\u6210\u3055\u308c\u305f\u753b\u50cf\u3092\u533a\u5225\u3059\u308b\u3053\u3068\u3092\u76ee\u7684\u3068\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>VQGAN \u306e\u91cd\u8981\u306a\u9769\u65b0\u306f\u3001\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc \u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306b\u3042\u308a\u307e\u3059\u3002\u9023\u7d9a\u8868\u73fe\u3092\u4f7f\u7528\u3059\u308b\u4ee3\u308f\u308a\u306b\u3001\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\u306f\u5165\u529b\u753b\u50cf\u3092\u96e2\u6563\u6f5c\u5728\u30b3\u30fc\u30c9\u306b\u30de\u30c3\u30d4\u30f3\u30b0\u3057\u3001\u753b\u50cf\u306e\u3055\u307e\u3056\u307e\u306a\u8981\u7d20\u3092\u8868\u3057\u307e\u3059\u3002\u3053\u308c\u3089\u306e\u96e2\u6563\u30b3\u30fc\u30c9\u306f\u3001\u5b9a\u7fa9\u6e08\u307f\u306e\u57cb\u3081\u8fbc\u307f\u307e\u305f\u306f\u30d9\u30af\u30c8\u30eb\u306e\u30bb\u30c3\u30c8\u3092\u542b\u3080\u30b3\u30fc\u30c9\u30d6\u30c3\u30af\u306b\u6e21\u3055\u308c\u307e\u3059\u3002\u30b3\u30fc\u30c9\u30d6\u30c3\u30af\u5185\u306e\u6700\u3082\u8fd1\u3044\u57cb\u3081\u8fbc\u307f\u304c\u5143\u306e\u30b3\u30fc\u30c9\u306b\u7f6e\u304d\u63db\u3048\u3089\u308c\u3001\u91cf\u5b50\u5316\u3055\u308c\u305f\u8868\u73fe\u306b\u306a\u308a\u307e\u3059\u3002\u3053\u306e\u30d7\u30ed\u30bb\u30b9\u306f\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u3068\u547c\u3070\u308c\u307e\u3059\u3002<\/p>\n<p>\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u4e2d\u3001\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\u3001\u30b8\u30a7\u30cd\u30ec\u30fc\u30bf\u30fc\u3001\u304a\u3088\u3073\u30c7\u30a3\u30b9\u30af\u30ea\u30df\u30cd\u30fc\u30bf\u30fc\u304c\u9023\u643a\u3057\u3066\u3001\u518d\u69cb\u7bc9\u640d\u5931\u3068\u6575\u5bfe\u7684\u640d\u5931\u3092\u6700\u5c0f\u9650\u306b\u6291\u3048\u3001\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0 \u30c7\u30fc\u30bf\u306b\u4f3c\u305f\u9ad8\u54c1\u8cea\u306e\u753b\u50cf\u3092\u751f\u6210\u3057\u307e\u3059\u3002VQGAN \u306f\u96e2\u6563\u6f5c\u5728\u30b3\u30fc\u30c9\u3092\u4f7f\u7528\u3059\u308b\u3053\u3068\u3067\u3001\u610f\u5473\u306e\u3042\u308b\u69cb\u9020\u3092\u30ad\u30e3\u30d7\u30c1\u30e3\u3059\u308b\u80fd\u529b\u3092\u9ad8\u3081\u3001\u3088\u308a\u5236\u5fa1\u3055\u308c\u305f\u753b\u50cf\u751f\u6210\u3092\u53ef\u80fd\u306b\u3057\u307e\u3059\u3002<\/p>\n<h3>\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN) \u306e\u4e3b\u306a\u7279\u5fb4<\/h3>\n<ol>\n<li>\n<p><strong>\u96e2\u6563\u6f5c\u5728\u30b3\u30fc\u30c9<\/strong>VQGAN \u306f\u96e2\u6563\u6f5c\u5728\u30b3\u30fc\u30c9\u3092\u63a1\u7528\u3057\u3066\u304a\u308a\u3001\u591a\u69d8\u3067\u5236\u5fa1\u3055\u308c\u305f\u753b\u50cf\u51fa\u529b\u3092\u751f\u6210\u3067\u304d\u307e\u3059\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u968e\u5c64\u69cb\u9020<\/strong>: \u30e2\u30c7\u30eb\u306e\u30b3\u30fc\u30c9\u30d6\u30c3\u30af\u306f\u3001\u8868\u73fe\u5b66\u7fd2\u30d7\u30ed\u30bb\u30b9\u3092\u5f37\u5316\u3059\u308b\u968e\u5c64\u69cb\u9020\u3092\u5c0e\u5165\u3057\u307e\u3059\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u5b89\u5b9a\u6027<\/strong>VQGAN \u306f\u3001\u5f93\u6765\u306e GAN \u3067\u898b\u3089\u308c\u308b\u4e0d\u5b89\u5b9a\u6027\u306e\u554f\u984c\u306e\u4e00\u90e8\u306b\u5bfe\u51e6\u3057\u3001\u3088\u308a\u30b9\u30e0\u30fc\u30ba\u3067\u4e00\u8cab\u6027\u306e\u3042\u308b\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u3092\u5b9f\u73fe\u3057\u307e\u3059\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u9ad8\u54c1\u8cea\u753b\u50cf\u751f\u6210<\/strong>: VQGAN \u306f\u3001\u5370\u8c61\u7684\u306a\u30c7\u30a3\u30c6\u30fc\u30eb\u3068\u4e00\u8cab\u6027\u3092\u5099\u3048\u305f\u3001\u9ad8\u89e3\u50cf\u5ea6\u3067\u8996\u899a\u7684\u306b\u9b45\u529b\u7684\u306a\u753b\u50cf\u3092\u751f\u6210\u3067\u304d\u307e\u3059\u3002<\/p>\n<\/li>\n<\/ol>\n<h2>\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN) \u306e\u7a2e\u985e<\/h2>\n<p>VQGAN \u306f\u8a95\u751f\u4ee5\u6765\u9032\u5316\u3092\u7d9a\u3051\u3066\u304a\u308a\u3001\u3044\u304f\u3064\u304b\u306e\u30d0\u30ea\u30a8\u30fc\u30b7\u30e7\u30f3\u3084\u6539\u826f\u304c\u63d0\u6848\u3055\u308c\u3066\u304d\u307e\u3057\u305f\u3002\u6ce8\u76ee\u3059\u3079\u304d VQGAN \u306e\u7a2e\u985e\u306b\u306f\u6b21\u306e\u3088\u3046\u306a\u3082\u306e\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u30bf\u30a4\u30d7<\/th>\n<th>\u8aac\u660e<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>VQ-VAE-2<\/td>\n<td>\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u3092\u6539\u826f\u3057\u305f VQ-VAE \u306e\u62e1\u5f35\u3002<\/td>\n<\/tr>\n<tr>\n<td>VQGAN+\u30af\u30ea\u30c3\u30d7<\/td>\n<td>VQGAN \u3068 CLIP \u30e2\u30c7\u30eb\u3092\u7d44\u307f\u5408\u308f\u305b\u3066\u3001\u3088\u308a\u512a\u308c\u305f\u753b\u50cf\u5236\u5fa1\u3092\u5b9f\u73fe\u3057\u307e\u3059\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u62e1\u6563\u30e2\u30c7\u30eb<\/td>\n<td>\u9ad8\u54c1\u8cea\u306a\u753b\u50cf\u5408\u6210\u306e\u305f\u3081\u306e\u62e1\u6563\u30e2\u30c7\u30eb\u3092\u7d71\u5408\u3057\u307e\u3059\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN) \u306e\u4f7f\u7528\u65b9\u6cd5\u3001\u4f7f\u7528\u306b\u95a2\u9023\u3059\u308b\u554f\u984c\u3068\u305d\u306e\u89e3\u6c7a\u7b56\u3002<\/h2>\n<h3>\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN) \u306e\u7528\u9014<\/h3>\n<ol>\n<li>\n<p><strong>\u753b\u50cf\u5408\u6210<\/strong>: VQGAN \u306f\u30ea\u30a2\u30eb\u3067\u591a\u69d8\u306a\u753b\u50cf\u3092\u751f\u6210\u3067\u304d\u308b\u305f\u3081\u3001\u30af\u30ea\u30a8\u30a4\u30c6\u30a3\u30d6\u306a\u30b3\u30f3\u30c6\u30f3\u30c4\u306e\u751f\u6210\u3001\u30a2\u30fc\u30c8\u3001\u30c7\u30b6\u30a4\u30f3\u306b\u5f79\u7acb\u3061\u307e\u3059\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u30b9\u30bf\u30a4\u30eb\u8ee2\u9001<\/strong>: \u6f5c\u5728\u30b3\u30fc\u30c9\u3092\u64cd\u4f5c\u3059\u308b\u3053\u3068\u3067\u3001VQGAN \u306f\u30b9\u30bf\u30a4\u30eb\u8ee2\u9001\u3092\u5b9f\u884c\u3057\u3001\u753b\u50cf\u306e\u69cb\u9020\u3092\u7dad\u6301\u3057\u306a\u304c\u3089\u753b\u50cf\u306e\u5916\u89b3\u3092\u5909\u66f4\u3067\u304d\u307e\u3059\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u30c7\u30fc\u30bf\u62e1\u5f35<\/strong>VQGAN \u306f\u3001\u4ed6\u306e\u30b3\u30f3\u30d4\u30e5\u30fc\u30bf\u30fc \u30d3\u30b8\u30e7\u30f3 \u30bf\u30b9\u30af\u306e\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0 \u30c7\u30fc\u30bf\u3092\u62e1\u5f35\u3057\u3001\u6a5f\u68b0\u5b66\u7fd2\u30e2\u30c7\u30eb\u306e\u4e00\u822c\u5316\u3092\u5411\u4e0a\u3055\u305b\u308b\u305f\u3081\u306b\u4f7f\u7528\u3067\u304d\u307e\u3059\u3002<\/p>\n<\/li>\n<\/ol>\n<h3>\u554f\u984c\u3068\u89e3\u6c7a\u7b56<\/h3>\n<ol>\n<li>\n<p><strong>\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u306e\u4e0d\u5b89\u5b9a\u3055<\/strong>\u591a\u304f\u306e\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0 \u30e2\u30c7\u30eb\u3068\u540c\u69d8\u306b\u3001VQGAN \u306f\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u306e\u4e0d\u5b89\u5b9a\u6027\u306b\u60a9\u307e\u3055\u308c\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u3001\u305d\u306e\u7d50\u679c\u3001\u30e2\u30fc\u30c9\u306e\u5d29\u58ca\u3084\u53ce\u675f\u4e0d\u826f\u304c\u767a\u751f\u3057\u307e\u3059\u3002\u7814\u7a76\u8005\u306f\u3001\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u8abf\u6574\u3001\u6b63\u898f\u5316\u6280\u8853\u306e\u4f7f\u7528\u3001\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306e\u6539\u5584\u306e\u5c0e\u5165\u306b\u3088\u3063\u3066\u3001\u3053\u306e\u554f\u984c\u306b\u5bfe\u51e6\u3057\u3066\u304d\u307e\u3057\u305f\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u30b3\u30fc\u30c9\u30d6\u30c3\u30af\u306e\u30b5\u30a4\u30ba<\/strong>\u30b3\u30fc\u30c9\u30d6\u30c3\u30af\u306e\u30b5\u30a4\u30ba\u306f\u3001\u30e2\u30c7\u30eb\u306e\u30e1\u30e2\u30ea\u8981\u4ef6\u3068\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u6642\u9593\u306b\u5927\u304d\u306a\u5f71\u97ff\u3092\u4e0e\u3048\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002\u7814\u7a76\u8005\u306f\u3001\u753b\u50cf\u306e\u54c1\u8cea\u3092\u72a0\u7272\u306b\u3059\u308b\u3053\u3068\u306a\u304f\u30b3\u30fc\u30c9\u30d6\u30c3\u30af\u306e\u30b5\u30a4\u30ba\u3092\u6700\u9069\u5316\u3059\u308b\u65b9\u6cd5\u3092\u7814\u7a76\u3057\u3066\u304d\u307e\u3057\u305f\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u5236\u5fa1\u6027<\/strong>VQGAN \u3067\u306f\u753b\u50cf\u751f\u6210\u3092\u3042\u308b\u7a0b\u5ea6\u5236\u5fa1\u3067\u304d\u307e\u3059\u304c\u3001\u6b63\u78ba\u306a\u5236\u5fa1\u3092\u5b9f\u73fe\u3059\u308b\u306e\u306f\u4f9d\u7136\u3068\u3057\u3066\u56f0\u96e3\u3067\u3059\u3002\u7814\u7a76\u8005\u306f\u3001\u30e2\u30c7\u30eb\u306e\u5236\u5fa1\u6027\u3092\u5411\u4e0a\u3055\u305b\u308b\u65b9\u6cd5\u3092\u7a4d\u6975\u7684\u306b\u7814\u7a76\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<\/li>\n<\/ol>\n<h2>\u4e3b\u306a\u7279\u5fb4\u3084\u305d\u306e\u4ed6\u306e\u985e\u4f3c\u7528\u8a9e\u3068\u306e\u6bd4\u8f03\u3092\u8868\u3084\u30ea\u30b9\u30c8\u306e\u5f62\u5f0f\u3067\u793a\u3057\u307e\u3059\u3002<\/h2>\n<h3>\u5f93\u6765\u306eGAN\u304a\u3088\u3073VAE\u3068\u306e\u6bd4\u8f03<\/h3>\n<table>\n<thead>\n<tr>\n<th>\u7279\u6027<\/th>\n<th>VQGAN<\/th>\n<th>\u5f93\u6765\u306eGAN<\/th>\n<th>VAEs<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6f5c\u5728\u7a7a\u9593\u8868\u73fe<\/td>\n<td>\u96e2\u6563\u30b3\u30fc\u30c9<\/td>\n<td>\u9023\u7d9a\u5024<\/td>\n<td>\u9023\u7d9a\u5024<\/td>\n<\/tr>\n<tr>\n<td>\u753b\u8cea<\/td>\n<td>\u9ad8\u54c1\u8cea<\/td>\n<td>\u591a\u69d8\u306a\u54c1\u8cea<\/td>\n<td>\u4e2d\u7a0b\u5ea6\u306e\u54c1\u8cea<\/td>\n<\/tr>\n<tr>\n<td>\u30e2\u30fc\u30c9\u306e\u6298\u308a\u305f\u305f\u307f<\/td>\n<td>\u524a\u6e1b<\/td>\n<td>\u5d29\u58ca\u3057\u3084\u3059\u3044<\/td>\n<td>\u9069\u7528\u3067\u304d\u306a\u3044<\/td>\n<\/tr>\n<tr>\n<td>\u5236\u5fa1\u6027<\/td>\n<td>\u5236\u5fa1\u306e\u6539\u5584<\/td>\n<td>\u5236\u9650\u3055\u308c\u305f\u5236\u5fa1<\/td>\n<td>\u512a\u308c\u305f\u30b3\u30f3\u30c8\u30ed\u30fc\u30eb<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>\u4ed6\u306e\u751f\u6210\u30e2\u30c7\u30eb\u3068\u306e\u6bd4\u8f03<\/h3>\n<table>\n<thead>\n<tr>\n<th>\u30e2\u30c7\u30eb<\/th>\n<th>\u7279\u5fb4<\/th>\n<th>\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>VQ-VAE<\/td>\n<td>\u5909\u5206\u30aa\u30fc\u30c8\u30a8\u30f3\u30b3\u30fc\u30c0 \u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3067\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u3092\u4f7f\u7528\u3057\u307e\u3059\u3002<\/td>\n<td>\u753b\u50cf\u5727\u7e2e\u3001\u30c7\u30fc\u30bf\u8868\u73fe\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u30af\u30ea\u30c3\u30d7<\/td>\n<td>\u8996\u899a\u3068\u8a00\u8a9e\u306e\u4e8b\u524d\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0 \u30e2\u30c7\u30eb\u3002<\/td>\n<td>\u753b\u50cf\u30ad\u30e3\u30d7\u30b7\u30e7\u30f3\u3001\u30c6\u30ad\u30b9\u30c8\u304b\u3089\u753b\u50cf\u3078\u306e\u751f\u6210\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u62e1\u6563\u30e2\u30c7\u30eb<\/td>\n<td>\u753b\u50cf\u5408\u6210\u306e\u305f\u3081\u306e\u78ba\u7387\u30e2\u30c7\u30eb\u3002<\/td>\n<td>\u9ad8\u54c1\u8cea\u306a\u753b\u50cf\u751f\u6210\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN) \u306b\u95a2\u9023\u3059\u308b\u5c06\u6765\u306e\u5c55\u671b\u3068\u6280\u8853\u3002<\/h2>\n<p>VQGAN \u306f\u3059\u3067\u306b\u3055\u307e\u3056\u307e\u306a\u30af\u30ea\u30a8\u30a4\u30c6\u30a3\u30d6 \u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u3067\u9a5a\u304f\u3079\u304d\u53ef\u80fd\u6027\u3092\u793a\u3057\u3066\u304a\u308a\u3001\u5c06\u6765\u3082\u6709\u671b\u3067\u3059\u3002VQGAN \u306b\u95a2\u9023\u3059\u308b\u5c06\u6765\u306e\u958b\u767a\u3068\u6280\u8853\u306b\u306f\u6b21\u306e\u3088\u3046\u306a\u3082\u306e\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n<ol>\n<li>\n<p><strong>\u5236\u5fa1\u6027\u306e\u5411\u4e0a<\/strong>\u7814\u7a76\u306e\u9032\u6b69\u306b\u3088\u308a\u3001\u751f\u6210\u3055\u308c\u305f\u753b\u50cf\u3092\u3088\u308a\u6b63\u78ba\u304b\u3064\u76f4\u611f\u7684\u306b\u5236\u5fa1\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u3001\u82b8\u8853\u8868\u73fe\u306e\u65b0\u305f\u306a\u53ef\u80fd\u6027\u304c\u958b\u304b\u308c\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30eb\u751f\u6210<\/strong>\u7814\u7a76\u8005\u305f\u3061\u306f\u3001VQGAN \u304c\u8907\u6570\u306e\u30b9\u30bf\u30a4\u30eb\u3084\u69d8\u5f0f\u3067\u753b\u50cf\u3092\u751f\u6210\u3067\u304d\u308b\u3088\u3046\u306b\u3057\u3001\u3055\u3089\u306b\u591a\u69d8\u3067\u5275\u9020\u7684\u306a\u51fa\u529b\u3092\u53ef\u80fd\u306b\u3059\u308b\u65b9\u6cd5\u3092\u6a21\u7d22\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u30ea\u30a2\u30eb\u30bf\u30a4\u30e0\u751f\u6210<\/strong>\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u3068\u6700\u9069\u5316\u6280\u8853\u304c\u9032\u6b69\u3059\u308b\u306b\u3064\u308c\u3066\u3001VQGAN \u3092\u4f7f\u7528\u3057\u305f\u30ea\u30a2\u30eb\u30bf\u30a4\u30e0\u753b\u50cf\u751f\u6210\u304c\u3088\u308a\u5b9f\u73fe\u53ef\u80fd\u306b\u306a\u308a\u3001\u30a4\u30f3\u30bf\u30e9\u30af\u30c6\u30a3\u30d6\u306a\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u304c\u53ef\u80fd\u306b\u306a\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n<\/li>\n<\/ol>\n<h2>\u30d7\u30ed\u30ad\u30b7 \u30b5\u30fc\u30d0\u30fc\u3092 Vector Quantized Generative Adversarial Network (VQGAN) \u3067\u4f7f\u7528\u3059\u308b\u65b9\u6cd5\u3001\u307e\u305f\u306f\u95a2\u9023\u4ed8\u3051\u308b\u65b9\u6cd5\u3002<\/h2>\n<p>\u30d7\u30ed\u30ad\u30b7 \u30b5\u30fc\u30d0\u30fc\u306f\u3001\u7279\u306b\u5927\u898f\u6a21\u306a\u30c7\u30fc\u30bf\u51e6\u7406\u3084\u753b\u50cf\u751f\u6210\u304c\u95a2\u4fc2\u3059\u308b\u30b7\u30ca\u30ea\u30aa\u306b\u304a\u3044\u3066\u3001VQGAN \u306e\u4f7f\u7528\u3092\u30b5\u30dd\u30fc\u30c8\u3059\u308b\u4e0a\u3067\u91cd\u8981\u306a\u5f79\u5272\u3092\u679c\u305f\u3057\u307e\u3059\u3002\u30d7\u30ed\u30ad\u30b7 \u30b5\u30fc\u30d0\u30fc\u3092 VQGAN \u306b\u4f7f\u7528\u3057\u305f\u308a\u95a2\u9023\u4ed8\u3051\u305f\u308a\u3059\u308b\u65b9\u6cd5\u306f\u3001\u6b21\u306e\u3068\u304a\u308a\u3067\u3059\u3002<\/p>\n<ol>\n<li>\n<p><strong>\u30c7\u30fc\u30bf\u53ce\u96c6\u3068\u524d\u51e6\u7406<\/strong>: \u30d7\u30ed\u30ad\u30b7 \u30b5\u30fc\u30d0\u30fc\u306f\u3001\u3055\u307e\u3056\u307e\u306a\u30bd\u30fc\u30b9\u304b\u3089\u753b\u50cf\u30c7\u30fc\u30bf\u306e\u53ce\u96c6\u3068\u524d\u51e6\u7406\u306b\u5f79\u7acb\u3061\u3001VQGAN \u306e\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u306e\u305f\u3081\u306e\u591a\u69d8\u3067\u4ee3\u8868\u7684\u306a\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3092\u78ba\u4fdd\u3057\u307e\u3059\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u4e26\u5217\u51e6\u7406<\/strong>: \u5927\u898f\u6a21\u306a\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3067 VQGAN \u3092\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u3059\u308b\u3068\u3001\u8a08\u7b97\u8ca0\u8377\u304c\u5927\u304d\u304f\u306a\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002\u30d7\u30ed\u30ad\u30b7 \u30b5\u30fc\u30d0\u30fc\u306f\u3001\u30ef\u30fc\u30af\u30ed\u30fc\u30c9\u3092\u8907\u6570\u306e\u30de\u30b7\u30f3\u306b\u5206\u6563\u3057\u3001\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0 \u30d7\u30ed\u30bb\u30b9\u3092\u9ad8\u901f\u5316\u3067\u304d\u307e\u3059\u3002<\/p>\n<\/li>\n<li>\n<p><strong>API\u30a8\u30f3\u30c9\u30dd\u30a4\u30f3\u30c8<\/strong>: \u30d7\u30ed\u30ad\u30b7 \u30b5\u30fc\u30d0\u30fc\u306f\u3001VQGAN \u30e2\u30c7\u30eb\u3092\u5c55\u958b\u3059\u308b\u305f\u3081\u306e API \u30a8\u30f3\u30c9\u30dd\u30a4\u30f3\u30c8\u3068\u3057\u3066\u6a5f\u80fd\u3057\u3001\u30e6\u30fc\u30b6\u30fc\u304c\u30ea\u30e2\u30fc\u30c8\u3067\u30e2\u30c7\u30eb\u3092\u64cd\u4f5c\u3057\u3001\u30aa\u30f3\u30c7\u30de\u30f3\u30c9\u3067\u753b\u50cf\u3092\u751f\u6210\u3067\u304d\u308b\u3088\u3046\u306b\u3057\u307e\u3059\u3002<\/p>\n<\/li>\n<\/ol>\n<h2>\u95a2\u9023\u30ea\u30f3\u30af<\/h2>\n<p>\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN) \u3068\u95a2\u9023\u30c8\u30d4\u30c3\u30af\u306e\u8a73\u7d30\u306b\u3064\u3044\u3066\u306f\u3001\u6b21\u306e\u30ea\u30bd\u30fc\u30b9\u3092\u53c2\u7167\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n<ol>\n<li>\n<p><a href=\"https:\/\/deepmind.com\/blog\/article\/introducing-vq-vae-2\" target=\"_new\" rel=\"noopener nofollow\">DeepMind \u30d6\u30ed\u30b0 \u2013 VQ-VAE-2 \u306e\u7d39\u4ecb<\/a><\/p>\n<\/li>\n<li>\n<p><a href=\"https:\/\/arxiv.org\/abs\/2006.10905\" target=\"_new\" rel=\"noopener nofollow\">arXiv \u2013 VQ-VAE-2: GAN \u3068 VAE \u306e\u96e2\u6563\u6f5c\u5728\u5909\u6570\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u306e\u6539\u5584<\/a><\/p>\n<\/li>\n<li>\n<p><a href=\"https:\/\/github.com\/deepmind\/deepmind-research\/tree\/master\/vq_vae_2\" target=\"_new\" rel=\"noopener nofollow\">GitHub \u2013 VQ-VAE-2 \u5b9f\u88c5<\/a><\/p>\n<\/li>\n<li>\n<p><a href=\"https:\/\/openai.com\/research\/publications\/clip\" target=\"_new\" rel=\"noopener nofollow\">OpenAI \u2013 CLIP: \u30c6\u30ad\u30b9\u30c8\u3068\u753b\u50cf\u3092\u3064\u306a\u3050<\/a><\/p>\n<\/li>\n<li>\n<p><a href=\"https:\/\/arxiv.org\/abs\/2103.00020\" target=\"_new\" rel=\"noopener nofollow\">arXiv \u2013 CLIP: \u5927\u898f\u6a21\u306a\u30c6\u30ad\u30b9\u30c8\u3068\u753b\u50cf\u306e\u63a5\u7d9a<\/a><\/p>\n<\/li>\n<\/ol>\n<p>\u3053\u308c\u3089\u306e\u30ea\u30bd\u30fc\u30b9\u3092\u8abf\u3079\u308b\u3053\u3068\u3067\u3001\u30d9\u30af\u30c8\u30eb\u91cf\u5b50\u5316\u6575\u5bfe\u7684\u751f\u6210\u30cd\u30c3\u30c8\u30ef\u30fc\u30af (VQGAN) \u3068\u3001\u4eba\u5de5\u77e5\u80fd\u304a\u3088\u3073\u30af\u30ea\u30a8\u30a4\u30c6\u30a3\u30d6 \u30b3\u30f3\u30c6\u30f3\u30c4\u751f\u6210\u306e\u4e16\u754c\u306b\u304a\u3051\u308b\u305d\u306e\u5fdc\u7528\u306b\u3064\u3044\u3066\u3001\u3088\u308a\u6df1\u304f\u7406\u89e3\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3059\u3002<\/p>","protected":false},"featured_media":470817,"menu_order":0,"template":"","meta":{"_acf_changed":false,"content-type":"","inline_featured_image":false,"footnotes":""},"class_list":["post-479505","wiki","type-wiki","status-publish","has-post-thumbnail","hentry"],"acf":{"faq_title":"Frequently Asked Questions about <mark>Vector Quantized Generative Adversarial Network (VQGAN)<\/mark>","faq_items":[{"question":"What is Vector Quantized Generative Adversarial Network (VQGAN)?","answer":"<p>Vector Quantized Generative Adversarial Network (VQGAN) is an advanced deep learning model that combines Generative Adversarial Networks (GANs) and Vector Quantization (VQ) techniques. It excels in generating high-quality images and offers improved control over the creative content generation process.<\/p>"},{"question":"How does VQGAN work?","answer":"<p>VQGAN consists of a generator and a discriminator, similar to traditional GANs. The key innovation lies in its encoder architecture, which maps input images to discrete latent codes. These codes are then quantized using a predefined set of embeddings in a codebook. The model is trained to minimize reconstruction and adversarial losses, resulting in realistic and visually appealing image synthesis.<\/p>"},{"question":"What are the key features of VQGAN?","answer":"<ul><li>Discrete Latent Codes: VQGAN uses discrete codes, enabling diverse and controlled image outputs.<\/li><li>Stability: VQGAN addresses stability issues common in traditional GANs, leading to smoother training.<\/li><li>High-Quality Image Generation: The model can generate high-resolution, detailed images.<\/li><\/ul>"},{"question":"What types of VQGAN exist?","answer":"<p>Some notable types of VQGAN include VQ-VAE-2, VQGAN+CLIP, and Diffusion Models. VQ-VAE-2 extends VQ-VAE with improved vector quantization, VQGAN+CLIP combines VQGAN with CLIP for better image control, and Diffusion Models integrate probabilistic models for high-quality image synthesis.<\/p>"},{"question":"How can VQGAN be used?","answer":"<p>VQGAN finds applications in various fields, including:<\/p><ul><li>Image Synthesis: Generating realistic and diverse images for creative content and art.<\/li><li>Style Transfer: Altering the appearance of images while preserving their structure.<\/li><li>Data Augmentation: Enhancing training data for better generalization in machine learning models.<\/li><\/ul>"},{"question":"What are the challenges and solutions related to using VQGAN?","answer":"<p>Challenges include training instability, codebook size, and achieving precise control over generated images. Researchers address these issues through hyperparameter adjustments, regularization techniques, and architectural improvements.<\/p>"},{"question":"What are the future perspectives of VQGAN?","answer":"<p>The future holds improved controllability, multi-modal generation, and real-time image synthesis using VQGAN. Advancements in research and hardware optimization will further enhance its capabilities.<\/p>"},{"question":"How are proxy servers associated with VQGAN?","answer":"<p>Proxy servers support VQGAN by assisting in data collection and preprocessing, enabling parallel processing for faster training, and serving as API endpoints for remote model deployment.<\/p>"}]},"_links":{"self":[{"href":"https:\/\/oneproxy.pro\/jp\/wp-json\/wp\/v2\/wiki\/479505","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/oneproxy.pro\/jp\/wp-json\/wp\/v2\/wiki"}],"about":[{"href":"https:\/\/oneproxy.pro\/jp\/wp-json\/wp\/v2\/types\/wiki"}],"version-history":[{"count":0,"href":"https:\/\/oneproxy.pro\/jp\/wp-json\/wp\/v2\/wiki\/479505\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/oneproxy.pro\/jp\/wp-json\/wp\/v2\/media\/470817"}],"wp:attachment":[{"href":"https:\/\/oneproxy.pro\/jp\/wp-json\/wp\/v2\/media?parent=479505"}],"curies":[{"name":"\u3046\u30fc\u3093","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}