Spaces:
Running
Running
| <html lang="en"> | |
| <head> | |
| <meta http-equiv="X-UA-Compatible" content="IE=Edge" /> | |
| <meta charset="utf-8" /> | |
| </head> | |
| <body style="margin: 0;"> | |
| <div id="p9" style="overflow: hidden; position: relative; background-color: white; width: 2200px; height: 1237px;"> | |
| <!-- Begin shared CSS values --> | |
| <style class="shared-css" type="text/css" > | |
| .t { | |
| transform-origin: bottom left; | |
| z-index: 2; | |
| position: absolute; | |
| white-space: pre; | |
| overflow: visible; | |
| line-height: 1.5; | |
| } | |
| .text-container { | |
| white-space: pre; | |
| } | |
| @supports (-webkit-touch-callout: none) { | |
| .text-container { | |
| white-space: normal; | |
| } | |
| } | |
| </style> | |
| <!-- End shared CSS values --> | |
| <!-- Begin inline CSS --> | |
| <style type="text/css" > | |
| #t1_9{left:0px;bottom:1054px;letter-spacing:-0.23px;} | |
| #t2_9{left:0px;bottom:917px;letter-spacing:-0.24px;} | |
| #t3_9{left:0px;bottom:781px;letter-spacing:-0.24px;} | |
| #t4_9{left:0px;bottom:645px;letter-spacing:-0.21px;} | |
| #t5_9{left:565px;bottom:998px;letter-spacing:0.2px;} | |
| #t6_9{left:565px;bottom:934px;letter-spacing:0.2px;} | |
| #t7_9{left:565px;bottom:870px;letter-spacing:0.19px;} | |
| #t8_9{left:609px;bottom:1080px;letter-spacing:0.15px;} | |
| #t9_9{left:565px;bottom:731px;letter-spacing:0.21px;} | |
| #ta_9{left:861px;bottom:731px;letter-spacing:0.23px;} | |
| #tb_9{left:565px;bottom:667px;letter-spacing:0.2px;} | |
| #tc_9{left:823px;bottom:667px;letter-spacing:0.19px;} | |
| #td_9{left:565px;bottom:603px;letter-spacing:0.21px;} | |
| #te_9{left:208px;bottom:456px;letter-spacing:0.21px;} | |
| #tf_9{left:504px;bottom:456px;letter-spacing:0.2px;} | |
| #tg_9{left:208px;bottom:391px;letter-spacing:0.22px;} | |
| #th_9{left:208px;bottom:327px;letter-spacing:0.22px;} | |
| #ti_9{left:208px;bottom:263px;letter-spacing:0.21px;} | |
| #tj_9{left:208px;bottom:199px;letter-spacing:0.21px;} | |
| .s1_9{font-size:115px;font-family:IBMPlexSans_2d;color:#000;} | |
| .s2_9{font-size:115px;font-family:IBMPlexSans-Bold_2l;color:#000;} | |
| .s3_9{font-size:53px;font-family:IBMPlexSans-Italic_2o;color:#000;} | |
| .s4_9{font-size:99px;font-family:IBMPlexSans-Bold_2l;color:#000;} | |
| .s5_9{font-size:53px;font-family:IBMPlexSans_2d;color:#000;} | |
| </style> | |
| <!-- End inline CSS --> | |
| <!-- Begin embedded font definitions --> | |
| <style id="fonts9" type="text/css" > | |
| @font-face { | |
| font-family: IBMPlexSans-Bold_2l; | |
| src: url("fonts/IBMPlexSans-Bold_2l.woff") format("woff"); | |
| } | |
| @font-face { | |
| font-family: IBMPlexSans-Italic_2o; | |
| src: url("fonts/IBMPlexSans-Italic_2o.woff") format("woff"); | |
| } | |
| @font-face { | |
| font-family: IBMPlexSans_2d; | |
| src: url("fonts/IBMPlexSans_2d.woff") format("woff"); | |
| } | |
| </style> | |
| <!-- End embedded font definitions --> | |
| <!-- Begin page background --> | |
| <div id="pg9Overlay" style="width:100%; height:100%; position:absolute; z-index:1; background-color:rgba(0,0,0,0); -webkit-user-select: none;"></div> | |
| <div id="pg9" style="-webkit-user-select: none;"><object width="2200" height="1237" data="9/9.svg" type="image/svg+xml" id="pdf9" style="width:2200px; height:1237px; -moz-transform:scale(1); z-index: 0;"></object></div> | |
| <!-- End page background --> | |
| <!-- Begin text definitions (Positioned/styled in CSS) --> | |
| <div class="text-container"><span id="t1_9" class="t s1_9">Problems </span> | |
| <span id="t2_9" class="t s1_9">and </span> | |
| <span id="t3_9" class="t s2_9">Proposed </span> | |
| <span id="t4_9" class="t s2_9">Solutions </span> | |
| <span id="t5_9" class="t s3_9">Fine-tuning a diffusion model on a small set of subject images </span> | |
| <span id="t6_9" class="t s3_9">causes it to lose the ability to generate generic images of the same </span> | |
| <span id="t7_9" class="t s3_9">class and forget the class-specific prior. </span> | |
| <span id="t8_9" class="t s4_9">1.Language Drift </span> | |
| <span id="t9_9" class="t s5_9">Solution 1 </span><span id="ta_9" class="t s5_9">Dreambooth use the model's own generated samples </span> | |
| <span id="tb_9" class="t s5_9">by adding </span><span id="tc_9" class="t s5_9">a relative weight of the prior-preservation loss. </span> | |
| <span id="td_9" class="t s5_9">However the ratio of prior-preservation is not easy to determine. </span> | |
| <span id="te_9" class="t s5_9">Solution 2 </span><span id="tf_9" class="t s5_9">This is a method that requires a lot of GPU time - during the regular </span> | |
| <span id="tg_9" class="t s5_9">training process, we add auto-generated images from the current model with </span> | |
| <span id="th_9" class="t s5_9">prompt of a single word, with words chosen from a pre-estimated word frequency </span> | |
| <span id="ti_9" class="t s5_9">list randomly according to a certain ratio (we chose our word list from Danbooru </span> | |
| <span id="tj_9" class="t s5_9">Tags). To avoid overfitting, each auto-generated image is used only once. </span></div> | |
| <!-- End text definitions --> | |
| </div> | |
| </body> | |
| </html> | |