Upgrade to Pro — share decks privately, control downloads, hide ads and more …

카카오브레인의 텍스트 기반 이미지 생성 기술

kakao
December 09, 2022
Programming
0
410

카카오브레인의 텍스트 기반 이미지 생성 기술

#GenerativeModel

카카오브레인에서 개발한 텍스트 기반 이미지 생성 기술을 소개합니다

발표자 : sam.ksh
카카오브레인에서 생성 모델을 개발하고 있는 연구원입니다.

kakao

December 09, 2022
Tweet

More Decks by kakao

See All by kakao
나의 컨텍스트를 아는 친구, Context-Aware AI Mate
kakao
0
50
카카오i를 통한 AIOT 서비스의 사례와 미래
kakao
0
60
데이터 분석과 머신러닝을 통한 유저 방문 맛집 발굴하기
kakao
0
55
최애 작품 이용권 선물해주는 ‘Helix 푸시’ 개발기
kakao
0
30
지연 시간 순삭! LLM 추론 구조와 효율적 애플리케이션 설계
kakao
0
52
문서 검토는 이제 Document AI로 한방에!
kakao
0
75
이미지까지 이해하는 Multimodal LLM의 학습 방법 밝혀내기
kakao
0
110
나만의 프로필 이미지를 만드는 Personlized T2I 모델 개발기
kakao
0
40
생성형AI를 활용한 개체명인식(NER)
kakao
0
61

Other Decks in Programming

See All in Programming
責務を分離するための例外設計 - PHPカンファレンス 2024
kajitack
9
2.3k
快速入門可觀測性
blueswen
0
480
LLM Supervised Fine-tuningの理論と実践
datanalyticslabo
8
1.8k
PHPで作るWebSocketサーバー ~リアクティブなアプリケーションを知るために~ / WebSocket Server in PHP - To know reactive applications
seike460
PRO
2
720
KubeCon NA 2024の全DB関連セッションを紹介
nnaka2992
0
110
103 Early Hints
sugi_0000
1
330
Scalaから始めるOpenFeature入門 / Scalaわいわい勉強会 #4
arthur1
1
380
fs2-io を試してたらバグを見つけて直した話
chencmd
0
280
Go の GC の不得意な 部分を克服したい
taiyow
3
980
オニオンアーキテクチャを使って、 Unityと.NETでコードを共有する
soi013
0
350
Запуск 1С:УХ в крупном энтерпрайзе: мечта и реальность ПМа
lamodatech
0
870
テストコードのガイドライン 〜作成から運用まで〜
riku929hr
6
1.3k

Featured

See All Featured
Building Flexible Design Systems
yeseniaperezcruz
328
38k
The Psychology of Web Performance [Beyond Tellerrand 2023]
tammyeverts
45
2.3k
The MySQL Ecosystem @ GitHub 2015
samlambert
250
12k
[RailsConf 2023 Opening Keynote] The Magic of Rails
eileencodes
28
9.2k
Why You Should Never Use an ORM
jnunemaker
PRO
54
9.1k
Art, The Web, and Tiny UX
lynnandtonic
298
20k
"I'm Feeling Lucky" - Building Great Search Experiences for Today's Users (#IAC19)
danielanewman
226
22k
Gamification - CAS2011
davidbonilla
80
5.1k
The Cult of Friendly URLs
andyhume
78
6.1k
Principles of Awesome APIs and How to Build Them.
keavy
126
17k
Visualizing Your Data: Incorporating Mongo into Loggly Infrastructure
mongodb
44
9.3k
Rails Girls Zürich Keynote
gr2m
94
13k

Transcript

  1. $PQZSJHIU ,BLBP $PSQ"MMSJHIUTSFTFSWFE3FEJTUSJCVUJPOPSQVCMJDEJTQMBZJTOPUQFSNJUUFEXJUIPVUXSJUUFOQFSNJTTJPOGSPN,BLBP ஠஠য়࠳ۨੋ੄ ఫझ౟ ӝ߈ ੉޷૑ ࢤࢿ ӝࣿ JG

    LBLBP  ஠஠য়࠳ۨੋ ӣࣁള TBNLTI

  2. ஠஠য়࠳ۨੋ੄ 5*ࢤࢿ ݽ؛ ࣗѐ അ੤ ݽ؛ ѐࢶਸ ਤೠ োҳ ೐۽ં౟

  3. ஠஠য়࠳ۨੋ੄ 5*ࢤࢿ ݽ؛ ࣗѐ 5PXBSET&GGJDJFOU5FYU$POEJUJPOBM%JGGVTJPO.PEFMT CZ"EWFSTBSJBM5SBJOJOH

  4. %FOPJTJOH%JGGVTJPO1SPCBCJMJTUJD.PEFMT "QIPUPPGBUFEEZCFBSPOUIFSPBE "IJHIRVBMJUZ QJDUVSFPGBNFEJFWBMLOJHIUXJUIHPMEFOBSNPS  +)P "+BJO BOE1"CCFFMt%FOPJTJOH%JGGVTJPO1SPCBCJMJTUJD.PEFMTu /FVS*14s

  5. 0WFSBMM"SDIJUFDUVSF VO$-*1 BLB%"--&  ਬࢎೞѱ 1SJPSm %FDPEFSm 43 ݽٕ۽ ҳࢿغয

    ੓णפ׮  "3BNFTI 1%IBSJXBM "/JDIPM $$IV BOE.$IFOu)JFSBSDIJDBM5FYU$POEJUJPOBM*NBHF(FOFSBUJPOXJUI$-*1-BUFOUTuBS9JWW

  6. 0WFSBMM"SDIJUFDUVSF A high quality picture of a medieval knight with

    golden armor Prior Y/ Y Y CLIP (text encoder) Y/ CLIP (image feature)

  7. 0WFSBMM"SDIJUFDUVSF A high quality picture of a medieval knight with

    golden armor Y/ Y Y CLIP (text encoder) Y/ CLIP (image feature) Prior

  8. 0WFSBMM"SDIJUFDUVSF A high quality picture of a medieval knight with

    golden armor Y/ Y Y CLIP (text encoder) Y/ CLIP (image feature) Prior

  9. 0WFSBMM"SDIJUFDUVSF 1SJPS m %FDPEFSm 43

  10. 0WFSBMM"SDIJUFDUVSF 1SJPS m %FDPEFSm 43 $-*1UFYUGFBUVSF QPPMJOH੹ ੿ࠁب ನೣ

  11. 0WFSBMM"SDIJUFDUVSF 1SJPS m %FDPEFSm 43 $-*1UFYUGFBUVSF UJNF

  12. 0WFSBMM"SDIJUFDUVSF 1SJPSm %FDPEFS m 43 A high quality picture of

    a medieval knight with golden armor Causal Transformer Y/ Y Y/ CLIP (text encoder) CLIP (image feature)

  13. 0WFSBMM"SDIJUFDUVSF 1SJPSm %FDPEFSm 43 Y Y/

  14. 0WFSBMM"SDIJUFDUVSF 1SJPSm %FDPEFSm *NQSPWFE43 Y Y/ Y 1SFUSBJOFE 'SFF[F 5SBJOBCMF

    JOJUJBMJ[FEGSPNUIF QSFUSBJOFEOFU MSE loss Perceptual loss Adversarial loss

  15. &YQFSJNFOU4FUVQ 1SJPS %FDPEFS 43 ೟ण ؘ੉ఠ .JNBHFUFYUQBJST $0:0ؘ੉ఠࣇ ನೣ ೟ण

    ߑध झ௼ې஖ ೟ण $-*1ݽ؛ 7J5- 7J5-  QBSBNFUFST # . . . PQUJNJ[BUJPO TUFQT . . . . TBNQMJOHTUFQT TUFQT TUFQT TUFQT

  16. 2VBOUJUBUJWF3FTVMUT $$. $0$0 ,GSPNWBM $-*1T 7J5#   '*% 

    

  17. 2VBOUJUBUJWF3FTVMUT $$. $0$0 ,GSPNWBM $-*1T 7J5#   '*% 

      "3BNFTI 1%IBSJXBM "/JDIPM $$IV BOE.$IFOu)JFSBSDIJDBM5FYU$POEJUJPOBM*NBHF(FOFSBUJPOXJUI$-*1-BUFOUTuBS9JWW WT GSPNVO$-*1

  18. 2VBMJUBUJWF3FTVMUT ࠂ੟ೠ ஶࣆী ؀ೠ ؀਽ מ۱ ಣо

  19. 2VBMJUBUJWF3FTVMUT ࠂ੟ೠ ஶࣆী ؀ೠ ؀਽ מ۱ ಣо (PSZFP DFMBEPOJOUIFTIBQFPGEBSUI WBEFS

    IPMEJOHBTXPSE (PSZFP DFMBEPOJOUIFTIBQFPGEBSUI WBEFS

  20. 2VBMJUBUJWF3FTVMUT ࠂ੟ೠ ஶࣆী ؀ೠ ؀਽ מ۱ ಣо (PSZFP DFMBEPOJOUIFTIBQFPGEBSUI WBEFS

    IPMEJOHBTXPSE "TVQFSNBUIXJ[BSEDBU SJDIMZUFYUVSFEPJMQBJOUJOH

  21. 2VBMJUBUJWF3FTVMUT ࠂ੟ೠ ஶࣆী ؀ೠ ؀਽ מ۱ ಣо "OBTUSPOBVUSJEJOHBIPSTF DIJMESFOESBXJOHTUZMF "OBTUSPOBVUSJEJOHBIPSTF

    DIJMESFOESBXJOHTUZMF

  22. 2VBMJUBUJWF3FTVMUT 1IPUPSFBMJTUJDࢤࢿ оמ ৈࠗ ಣо

  23. 2VBMJUBUJWF3FTVMUT 1IPUPSFBMJTUJDࢤࢿ оמ ৈࠗ ಣо "OBTUSPOBVUSJEJOHBIPSTF DIJMESFOESBXJOHTUZMF BQIPUPPGBCBCZQVQQZXBJUJOHGPSIFSNPN

  24. 2VBMJUBUJWF3FTVMUT 1IPUPSFBMJTUJDࢤࢿ оמ ৈࠗ ಣо 1IPUPPGBZFBSPME XIJUFNBO TJMWFSIBJS OFBUCFBSE XFBSJOHBSFEHJMFU

    UIJDLSJNNFEHMBTTFT

  25. 2VBMJUBUJWF3FTVMUT 1IPUPSFBMJTUJDࢤࢿ оמ ৈࠗ ಣо "OBTUSPOBVUSJEJOHBIPSTF DIJMESFOESBXJOHTUZMF BQPSUSBJUPGBOPMENPOL

  26. 2VBMJUBUJWF3FTVMUT *NBHFWBSJBUJPOоמ ৈࠗ ಣо

  27. 2VBMJUBUJWF3FTVMUT *NBHFWBSJBUJPOоמ ৈࠗ ಣо

  28. അ੤ ݽ؛ ѐࢶਸ ਤೠ োҳ ೐۽ં౟ 0VS "QQSPBDIFT UP *NQSPWF

    5XPTUBHF *NBHF (FOFSBUJPO 'SBNFXPSL

  29. 5XP4UBHF*NBHF(FOFSBUJPO'SBNFXPSL  "0PSE 07JOZBMT BOE,,BWVLDVPHMVt/FVSBM%JTDSFUF3FQSFTFOUBUJPO-FBSOJOHu/FVS*14s

  30. 5XP4UBHF*NBHF(FOFSBUJPO'SBNFXPSL  "0PSE 07JOZBMT BOE,,BWVLDVPHMVt/FVSBM%JTDSFUF3FQSFTFOUBUJPO-FBSOJOHu/FVS*14s

  31. 5XP4UBHF*NBHF(FOFSBUJPO'SBNFXPSL  "0PSE 07JOZBMT BOE,,BWVLDVPHMVt/FVSBM%JTDSFUF3FQSFTFOUBUJPO-FBSOJOHu/FVS*14s

  32. 5XP4UBHF*NBHF(FOFSBUJPO'SBNFXPSL  "0PSE 07JOZBMT BOE,,BWVLDVPHMVt/FVSBM%JTDSFUF3FQSFTFOUBUJPO-FBSOJOHu/FVS*14s 4UBHF

  33. 5XP4UBHF*NBHF(FOFSBUJPO'SBNFXPSL  "0PSE 07JOZBMT BOE,,BWVLDVPHMVt/FVSBM%JTDSFUF3FQSFTFOUBUJPO-FBSOJOHu/FVS*14s 4UBHF"3PSOPO"3NPEFMJOH

  34. അ੤ ݽ؛ ѐࢶਸ ਤೠ োҳ ೐۽ં౟  %-FF $,JN 4,JN

    .$IP BOE8)BOt"VUPSFHSFTTJWF*NBHF(FOFSBUJPOVTJOH3FTJEVBM2VBOUJ[BUJPOt$713s  5:V 4,JN $,JN %-FF BOE#)BOt-PDBMMZ)JFSBSDIJDBM"VUPSFHSFTTJWF.PEFMJOHGPS*NBHF(FOFSBUJPOt/FVS*14s  %-FF $,JN 4,JN .$IP BOE8)BOt%SBGUBOE3FWJTF&GGFDUJWF*NBHF(FOFSBUJPOXJUI$POUFYUVBM325SBOTGPSNFSt/FVS*14s 327"& $713s )27"& /FVS*14s %SBGUBOE3FWJTF /FVS*14s

  35. 4FRVFODF-FOHUIBOE"3.PEFM

  36. 4FRVFODF-FOHUIBOE"3.PEFM

  37. 3FTJEVBM2VBOUJ[FE7"& 327"&  %-FF $,JN 4,JN .$IP BOE8)BOt"VUPSFHSFTTJWF*NBHF(FOFSBUJPOVTJOH3FTJEVBM2VBOUJ[BUJPOt$713s

  38. 3FTJEVBM2VBOUJ[FE7"& 327"&  %-FF $,JN 4,JN .$IP BOE8)BOt"VUPSFHSFTTJWF*NBHF(FOFSBUJPOVTJOH3FTJEVBM2VBOUJ[BUJPOt$713s

  39. 3FTJEVBM2VBOUJ[FE7"& 327"&  %-FF $,JN 4,JN .$IP BOE8)BOt"VUPSFHSFTTJWF*NBHF(FOFSBUJPOVTJOH3FTJEVBM2VBOUJ[BUJPOt$713s

  40. 3FTJEVBM2VBOUJ[FE7"& 327"&  %-FF $,JN 4,JN .$IP BOE8)BOt"VUPSFHSFTTJWF*NBHF(FOFSBUJPOVTJOH3FTJEVBM2VBOUJ[BUJPOt$713s

  41. 3FTJEVBM2VBOUJ[FE7"& 327"&  %-FF $,JN 4,JN .$IP BOE8)BOt"VUPSFHSFTTJWF*NBHF(FOFSBUJPOVTJOH3FTJEVBM2VBOUJ[BUJPOt$713s

  42. 3FTJEVBM2VBOUJ[FE7"& 327"&  %-FF $,JN 4,JN .$IP BOE8)BOt"VUPSFHSFTTJWF*NBHF(FOFSBUJPOVTJOH3FTJEVBM2VBOUJ[BUJPOt$713s $PEF3FTPMVUJPO )

    Y 8 Y % $PEFCPPL4J[F S'*% 72("/ YY   m  QBQFS YY          327"& YY    YY   

  43. 3FTJEVBM2VBOUJ[FE5SBOTGPSNFS 325SBOTGPSNFS  %-FF $,JN 4,JN .$IP BOE8)BOt"VUPSFHSFTTJWF*NBHF(FOFSBUJPOVTJOH3FTJEVBM2VBOUJ[BUJPOt$713s

  44. 325SBOTGPSNFSز੘ ߑध

  45. 3FTJEVBM2VBOUJ[FE5SBOTGPSNFS 325SBOTGPSNFS  %-FF $,JN 4,JN .$IP BOE8)BOt"VUPSFHSFTTJWF*NBHF(FOFSBUJPOVTJOH3FTJEVBM2VBOUJ[BUJPOt$713s

  46. 3FTJEVBM2VBOUJ[FE5SBOTGPSNFS 325SBOTGPSNFS  %-FF $,JN 4,JN .$IP BOE8)BOt"VUPSFHSFTTJWF*NBHF(FOFSBUJPOVTJOH3FTJEVBM2VBOUJ[BUJPOt$713s

  47. )JFSBSDIJDBM2VBOUJ[FE7"& )27"&  5:V 4,JN $,JN %-FF BOE#)BOt-PDBMMZ)JFSBSDIJDBM"VUPSFHSFTTJWF.PEFMJOHGPS*NBHF(FOFSBUJPOt/FVS*14s

  48. )JFSBSDIJDBM2VBOUJ[FE7"& )27"&  5:V 4,JN $,JN %-FF BOE#)BOt-PDBMMZ)JFSBSDIJDBM"VUPSFHSFTTJWF.PEFMJOHGPS*NBHF(FOFSBUJPOt/FVS*14s

  49. )JFSBSDIJDBM2VBOUJ[FE7"& )27"&  5:V 4,JN $,JN %-FF BOE#)BOt-PDBMMZ)JFSBSDIJDBM"VUPSFHSFTTJWF.PEFMJOHGPS*NBHF(FOFSBUJPOt/FVS*14s

  50. $POUFYUVBM325SBOTGPSNFS  %-FF $,JN 4,JN .$IP BOE8)BOt%SBGUBOE3FWJTF&GGFDUJWF*NBHF(FOFSBUJPOXJUI$POUFYUVBM325SBOTGPSNFSt/FVS*14s

  51. $POUFYUVBM325SBOTGPSNFS  %-FF $,JN 4,JN .$IP BOE8)BOt%SBGUBOE3FWJTF&GGFDUJWF*NBHF(FOFSBUJPOXJUI$POUFYUVBM325SBOTGPSNFSt/FVS*14s

  52. $POUFYUVBM325SBOTGPSNFS  %-FF $,JN 4,JN .$IP BOE8)BOt%SBGUBOE3FWJTF&GGFDUJWF*NBHF(FOFSBUJPOXJUI$POUFYUVBM325SBOTGPSNFSt/FVS*14s

  53. $POUFYUVBM325SBOTGPSNFS  %-FF $,JN 4,JN .$IP BOE8)BOt%SBGUBOE3FWJTF&GGFDUJWF*NBHF(FOFSBUJPOXJUI$POUFYUVBM325SBOTGPSNFSt/FVS*14s

  54. ߊ಴ܳ ݃஖ݶࢲ

  55. ଵҊ ޙ೴  +)P "+BJO BOE1"CCFFMt%FOPJTJOH%JGGVTJPO1SPCBCJMJTUJD.PEFMTu /FVS*14s  "3BNFTI 1%IBSJXBM

    "/JDIPM $$IV BOE.$IFOu)JFSBSDIJDBM5FYU$POEJUJPOBM*NBHF(FOFSBUJPOXJUI$-*1-BUFOUTuBS9JWW  "0PSE 07JOZBMT BOE,,BWVLDVPHMVt/FVSBM%JTDSFUF3FQSFTFOUBUJPO-FBSOJOHu/FVS*14s  %-FF $,JN 4,JN .$IP BOE8)BOt"VUPSFHSFTTJWF*NBHF(FOFSBUJPOVTJOH3FTJEVBM2VBOUJ[BUJPOt$713s  5:V 4,JN $,JN %-FF BOE#)BOt-PDBMMZ)JFSBSDIJDBM"VUPSFHSFTTJWF.PEFMJOHGPS*NBHF(FOFSBUJPOt/FVS*14s  %-FF $,JN 4,JN .$IP BOE8)BOt%SBGUBOE3FWJTF&GGFDUJWF*NBHF(FOFSBUJPOXJUI$POUFYUVBM325SBOTGPSNFSt/FVS*14s

  56. &0%