{"id":64782,"date":"2024-05-31T14:06:56","date_gmt":"2024-05-31T05:06:56","guid":{"rendered":"https:\/\/smilegate.ai\/?p=64782"},"modified":"2024-08-12T10:57:15","modified_gmt":"2024-08-12T01:57:15","slug":"%eb%94%94%ed%93%a8%ec%a0%84-%eb%a7%90%ea%b3%a0-cfm-%eb%93%a4%ec%96%b4%eb%b4%a4%eb%8b%88","status":"publish","type":"post","link":"https:\/\/smilegate.ai\/cn\/2024\/05\/31\/%eb%94%94%ed%93%a8%ec%a0%84-%eb%a7%90%ea%b3%a0-cfm-%eb%93%a4%ec%96%b4%eb%b4%a4%eb%8b%88\/","title":{"rendered":"\ub514\ud4e8\uc804 \ub9d0\uace0 CFM \ub4e4\uc5b4\ubd24\ub2c8?"},"content":{"rendered":"

\uc778\uacf5\uc9c0\ub2a5\uc758 \ubc1c\uc804\uc740 \uc6b0\ub9ac\uc758 \uc0b6\uc5d0 \ud070 \ubcc0\ud654\ub97c \uac00\uc838\uc654\uc2b5\ub2c8\ub2e4. \ud2b9\ud788 generative \ubaa8\ub378\uc740 \uc774\ub7ec\ud55c \ubcc0\ud654\uc758 \uc911\uc2ec\uc5d0 \uc11c \uc788\uc2b5\ub2c8\ub2e4. \ub514\ud4e8\uc804\uc744 \uae30\ubc18\uc73c\ub85c \ud55c \uc0dd\uc131 \ubaa8\ub378\uc774 \uc774\ubbf8\uc9c0, \uc624\ub514\uc624 \uc0dd\uc131 \ubd84\uc57c\uc5d0 \ud65c\uc6a9\ub418\uace0 \uc788\uc73c\uba70, \ub2e4\uc591\ud55c \uc131\uacfc\ub97c \ubcf4\uc5ec \uc8fc\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \ub514\ud4e8\uc804 \uae30\ubc18 \uc0dd\uc131 \ubaa8\ub378\uc740 \ucd94\ub860\uc744 \uc704\ud55c \uc5f0\uc0b0 \ubcf5\uc7a1\ub3c4\uac00 \ub192\uae30 \ub54c\ubb38\uc5d0 \ucd94\ub860 \uc18d\ub3c4\uc5d0 \ub300\ud55c \uace0\ubbfc\uc774 \uc788\uc2b5\ub2c8\ub2e4. \ucd5c\uadfc \uc774\ub7ec\ud55c \uace0\ubbfc\uc774 \ubc1c\uc804 \ud558\uba74\uc11c One-step \uc0dd\uc131\uc774\ub098 Conditional Flow Matching \uacfc \uac19\uc740 \uae30\uc220\uc774 \uc8fc\ubaa9 \ubc1b\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uae30\uc220\ub4e4\uc740 \uae30\uc874 \ub514\ud4e8\uc804\uacfc \uc720\uc0ac\ud55c \uc0dd\uc131 \ubc29\ubc95\uc744 \ub530\ub974\uc9c0\ub9cc \uc791\uc740 Number of Function Evaluation (NFE)\ub85c \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \ubcf4\uc5ec\uc8fc\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \ud3ec\uc2a4\ud305\uc5d0\uc11c\ub294 2023\ub144 ICLR\uc5d0\uc11c Lipman \uc120\uc0dd\ub2d8\uc774 \ubc1c\ud45c\ud558\uc2e0 \ub17c\ubb38 Flow Matching for Generative Modeling\uc758 \ub0b4\uc6a9\uc778 Conditional Flow Matching\uc5d0 \ub300\ud55c \uc18c\uac1c\uc640 \ud65c\uc6a9\ub418\uace0 \uc788\ub294 \uc5f0\uad6c \ubd84\uc57c\uc5d0 \ub300\ud574\uc11c \uc774\uc57c\uae30 \ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n

Conditional Flow Matching (CFM) \uc774\ub780?<\/strong><\/p>\n\n\n\n

CFM\uc740 \ucd08\uae30 \uc0c1\ud0dc X0<\/sub>\ub97c \ubaa9\ud45c \uc0c1\ud0dc X1<\/sub> \uc73c\ub85c \uc774\ub3d9 \uc2dc\ud0a4\ub294 \uac83\uc744 \ubaa9\ud45c\ub85c \ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uc811\uadfc\uc740 \ub514\ud4e8\uc804 \ubaa8\ub378\uacfc \ub3d9\uc77c\ud55c Score \uae30\ubc18 \uc0dd\uc131 \ubaa8\ub378\uc744 \ub530\ub974\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\ub9cc \ub514\ud4e8\uc804 \uae30\ubc18 \ubaa8\ub378\uacfc\uc758 \ucc28\uc774\uc810\uc740 X0 <\/sub>-> X1<\/sub>\uc73c\ub85c \ubcc0\ud658 \ud558\uae30 \uc704\ud55c \ub370\uc774\ud130\uc758 \uc774\ub3d9 \uacbd\ub85c\ub97c \uc815\uc758 \ud558\ub294\ub370 \uc788\uc2b5\ub2c8\ub2e4. \ub514\ud4e8\uc804 \uae30\ubc18 \uc0dd\uc131 \ubaa8\ub378\uc758 \uacbd\uc6b0 Forward\uc640 Backward process\ub97c \uc815\uc758 \ud558\ub294\ub370 \uc788\uc5b4\uc11c Stochastic Differential Equation (SDE) path\ub97c \ub530\ub974\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ubc18\uba74 CFM \uae30\ubc18 \uc0dd\uc131 \ubaa8\ub378\uc740 Ordinary Differential Equation (ODE) path\ub97c \ub530\ub974\uace0 \uc788\uae30 \ub54c\ubb38\uc5d0 \uc2dc\uac04\ubcc0\uc218 t \uc5d0 \ub300\ud55c \ubbf8\ubd84\uc744 \ud1b5\ud574 \uc774\ub3d9\uacbd\ub85c\ub97c \uc815\uc758 \ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n

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

<\/p>\n\n\n\n

\ub0b4\uc6a9\uc774 \uc5b4\ub835\ub098\uc694? \uadf8\ub7fc \uc880 \ub354 \uad6c\uccb4\uc801\uc73c\ub85c \uc124\uba85 \ub4dc\ub9ac\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n

\ubb3c\uccb4\uac00 \uc2dc\uac04 \ubcc0\ud654\uc5d0 \ub530\ub978 \ud568\uc218 F(t)\ub85c \uc6c0\uc9c1\uc778\ub2e4\uace0 \uac00\uc815\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \ubb3c\uccb4\uac00 \ud604\uc7ac \uc2dc\uac04 t \uc5d0\uc11c t+1\ub85c \uc6c0\uc9c1\uc77c \ub54c\ub294 dF(t)\/dt \ub77c\ub294 \ubbf8\ubd84 \uc815\ubcf4\ub97c \ud1b5\ud574 \ubb3c\uc81c\uac00 \uc5b4\ub5bb\uac8c \uc6c0\uc9c1\uc77c \uc218 \uc788\ub294\uc9c0 \uc54c \uc218 \uc788\uac8c \ub429\ub2c8\ub2e4.<\/p>\n\n\n\n

<\/p>\n\n\n\n

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

\uc774\ub7ec\ud55c \uc6d0\ub9ac\ub97c \uc0dd\uc131 \ubaa8\ub378\uc5d0 \uc801\uc6a9\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. X0<\/sub> \uc640 X1<\/sub> \uc774 \uc2dc\uac04\uc5d0 \ub530\ub978 \ubbf8\ubd84 \uac00\ub2a5\ud55c \ud568\uc218 F(t)\ub85c \uc815\uc758\uac00 \ub420 \uc218 \uc788\ub2e4\uba74, \uadf8\ub9ac\uace0 F(0) = X0<\/sub> , F(1) = X1<\/sub>\uc744 \ub9cc\uc871\ud55c\ub2e4\uba74 F(t)\ub294 CFM\uc758 path \ud65c\uc6a9 \ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uadf8\ub9ac\uace0 Neural Network\ub97c \ud1b5\ud574 Xn<\/sub>=F(tn<\/sub>) \uc744 \uc785\ub825\uc73c\ub85c \ubc1b\uc544\uc11c F(tn<\/sub>)’ \ub97c \ucd94\ub860 \ud558\ub3c4\ub85d Supervised learning\uc744 \ud1b5\ud55c \uc9c1\uad00\uc801\uc73c\ub85c \ubaa8\ub378 \ud559\uc2b5\uc774 \uac00\ub2a5\ud558\uba70, ODE sampling\uc744 \ud1b5\ud55c X0<\/sub> -> X1<\/sub> \ubcc0\ud658\uc774 \uac00\ub2a5\ud574\uc9d1\ub2c8\ub2e4.<\/p>\n\n\n\n

\uc5ec\uae30\uc11c X0<\/sub>\ub97c \uac00\uc6b0\uc2dc\uac04 \ubd84\ud3ec X1<\/sub>\uc744 \ubaa9\ud45c \uc774\ubbf8\uc9c0\ub85c \uc0dd\uac01\ud574 \ubcf4\uba74, Noise\uc5d0\uc11c Sampling\uc744 \ud1b5\ud574 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294 \ub514\ud4e8\uc804\uacfc \uc720\uc0ac\ud558\uac8c \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131 \ud560 \uc218 \uc788\uac8c \ub429\ub2c8\ub2e4.<\/p>\n\n\n\n

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

<\/p>\n\n\n\n

CFM\uc758 \uc74c\uc131 \uc5f0\uad6c \ud65c\uc6a9<\/strong><\/p>\n\n\n\n

\uc774\ub7ec\ud55c CFM\uc740 \uace0\uc804\uc801\uc778 \uc774\ubbf8\uc9c0 \uc0dd\uc131\uc744 \ub118\uc5b4 \uc74c\uc131 \ud569\uc131 \ubd84\uc57c\uc5d0\ub3c4 \ub9ce\uc774 \ud65c\uc6a9 \ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. META\uc5d0\uc11c \ubc1c\ud45c\ub41c Voice Box \uc758 \uacbd\uc6b0 \ubaa9\ud45c \ud654\uc790\uac00 \uc815\uc758 \ub418\uc5b4 \uc788\ub294 \uc0c1\ud669\uc5d0\uc11c Text embedding\uc744 latent\ub85c \uc815\uc758 \ud558\uace0, Transformer \uae30\ubc18 Masked prediction\uc744 \ud1b5\ud574 CFM \ubaa8\ub378\uc744 \ud559\uc2b5 \ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c CFM \uae30\ubc18 \uc74c\uc131 \ud569\uc131 \uc5f0\uad6c\ub294 \ucd5c\uadfc \ud070 \uc8fc\ubaa9\uc744 \ubc1b\uace0 \uc788\uc73c\uba70, \ubcf4\ub2e4 \uc6b0\uc218\ud55c \ud569\uc131 \ud004\ub9ac\ud2f0\uc640 \uacbd\ub7c9\ud654 \uc5f0\uad6c\uac00 \uc9c4\ud589\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n

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

<\/p>\n\n\n\n

<\/p>\n\n\n\n

<\/p>\n\n\n\n

<\/p>\n

<\/span><\/div>","protected":false},"excerpt":{"rendered":"

\uc778\uacf5\uc9c0\ub2a5\uc758 \ubc1c\uc804\uc740 \uc6b0\ub9ac\uc758 \uc0b6\uc5d0 \ud070 \ubcc0\ud654\ub97c \uac00\uc838\uc654\uc2b5\ub2c8\ub2e4. \ud2b9\ud788 generative \ubaa8\ub378\uc740 \uc774\ub7ec\ud55c \ubcc0\ud654\uc758 \uc911\uc2ec\uc5d0 \uc11c \uc788\uc2b5\ub2c8\ub2e4. \ub514\ud4e8\uc804\uc744 \uae30\ubc18\uc73c\ub85c \ud55c \uc0dd\uc131 \ubaa8\ub378\uc774 \uc774\ubbf8\uc9c0, \uc624\ub514\uc624 \uc0dd\uc131 \ubd84\uc57c\uc5d0 \ud65c\uc6a9\ub418\uace0 \uc788\uc73c\uba70, \ub2e4\uc591\ud55c \uc131\uacfc\ub97c \ubcf4\uc5ec \uc8fc\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \ub514\ud4e8\uc804 \uae30\ubc18 \uc0dd\uc131 \ubaa8\ub378\uc740 \ucd94\ub860\uc744 \uc704\ud55c \uc5f0\uc0b0 \ubcf5\uc7a1\ub3c4\uac00 \ub192\uae30 \ub54c\ubb38\uc5d0 \ucd94\ub860 \uc18d\ub3c4\uc5d0 \ub300\ud55c \uace0\ubbfc\uc774 \uc788\uc2b5\ub2c8\ub2e4. \ucd5c\uadfc \uc774\ub7ec\ud55c \uace0\ubbfc\uc774 \ubc1c\uc804 \ud558\uba74\uc11c One-step \uc0dd\uc131\uc774\ub098 Conditional…<\/p>\n

<\/span><\/div>","protected":false},"author":1,"featured_media":64789,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"rank_math_lock_modified_date":false,"footnotes":""},"categories":[17,19],"tags":[188,296,710],"class_list":["post-64782","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tech02","category-tech04","tag-featured","tag-tts","tag-generative-ai","category-17","category-19","description-off"],"_links":{"self":[{"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/posts\/64782","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/comments?post=64782"}],"version-history":[{"count":2,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/posts\/64782\/revisions"}],"predecessor-version":[{"id":64913,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/posts\/64782\/revisions\/64913"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/media\/64789"}],"wp:attachment":[{"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/media?parent=64782"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/categories?post=64782"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/smilegate.ai\/cn\/wp-json\/wp\/v2\/tags?post=64782"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}