Slide 1

Slide 1 text

4DBMBCMF%JGGVTJPO.PEFMTXJUI 5SBOTGPSNFST ܚጯٛक़େֶ ਿӜ޸໌ݚڀࣨ#ീౡେ஍ 8JMMJBN1FFCMFT 4BJOJOH9JF 6$#FSLFMFZ /FX:PSL6OJWFSTJUZ *$$7 8JMMJBN1FFCMFT 4BJOJOH9JFl4DBMBCMF%JGGVTJPO.PEFMTXJUI5SBOTGPSNFSTzJO*$$7 QQ

Slide 2

Slide 2 text

എܠɿ֦ࢄϞσϧʹΑΔಈը૾ੜ੒ FH 4PSB ͷൃల IUUQTXXXZPVUVCFDPNXBUDI W),Z%"1/@

Slide 3

Slide 3 text

ؔ࿈ݚڀɿ֦ࢄϞσϧͷόοΫϘʔϯͱͯ͠6/FU͕ଟ༻ • 6/FUͷ.VMUJTDBMFTLJQDPOOFDUJPOTˠ ෆཁͳܭࢉࢿݯͷ࢖༻ 手法 概要 DALL-E 2 [Ramesh+, 22] CLIPを用いてテキストと画像のAlignmentを行う Stable Diffusion [Rombach+, CVPR22] 潜在拡散モデル 6/FU<3POOFCFSHFS .*$$"*> 4UBCMF%JGGVTJPO<3PNCBDI $713>

Slide 4

Slide 4 text

ఏҊख๏ɿ%JGGVTJPO5SBOTGPSNFS %J5 • જࡏ֦ࢄϞσϧ -%. <3PNCBDI $713>Λϕʔεʹߏங • 7JTJPO5SBOTGPSNFS 7J5 <%PTPWJUTLJZ *$-3>ػߏΛಋೖ • $POEJUJPOJOHʹΑΔ৚݅৘ใͷೖྗ

Slide 5

Slide 5 text

ఏҊख๏ ɿજࡏ֦ࢄϞσϧͱֶͯ͠शͤ͞Δ • ߴ࣍ݩͷըૉۭؒͰ֦ࢄϞσϧΛֶश ͤ͞Δ͜ͱ͸ܭࢉྔతʹࠔ೉ • -%.ͱֶͯ͠शͤ͞Δ͜ͱͰ ܭࢉྔΛ࡟ݮ • ըૉۭؒͷ֦ࢄϞσϧͰ͋Δ "%.<%IBSJXBM /FVS*14> ͷ໿ͷͷ(GMPQTͰֶशՄೳ

Slide 6

Slide 6 text

ఏҊख๏ ɿೖྗϊΠζΛQBUDIʹ෼ղ • "VUPFODPEFS͔ΒಘΒΕͨ /PJTFE-BUFOU YY Λ 7J5ͱಉ༷ʹE࣍ݩͷτʔΫϯ5ʹม׵ • 1BUDIαΠζQΛ൒෼ʹ͢Δͱ5͸ഒ ʹͳΓUSBOTGPSNFS૚ͷ (GMPQT͸গͳ͘ͱ΋ഒҎ্

Slide 7

Slide 7 text

ఏҊख๏ ɿ৚݅෇͖ೖྗ $POEJUJPOJOH ͷॲཧ • ৚݅෇͖֦ࢄϞσϧͰ͸ϊΠζΛؚΉը૾ͱͱ΋ʹ௥Ճ৘ใ͕Ճ͑ΒΕΔ FH UJNFTUFQɼΫϥεϥϕϧɼࣗવݴޠ FUD • ຊݚڀͰ͸͜ΕΒͷ৚݅෇͖ೖྗΛॲཧ͢ΔͨΊʹҎԼͷͭͷҟͳΔઃܭΛఏҊ • *ODPOUFYUDPOEJUJPOJOH • $SPTT"UUFOUJPOCMPDL • "EBQUJWFMBZFSOPSN BEB-/ CMPDL • BEB-/;FSPCMPDL

Slide 8

Slide 8 text

ఏҊख๏ ɿBEB-/;FSPCMPDL • 7J5ͷTFMGBUUFOUJPOCMPDLʹରͯ͠"EB-/ػߏΛಋೖ • "EB-/ͷεέʔϧ܎਺ ͓Αͼ ࢒ࠩ઀ଓͷલͷεέʔϧ܎਺ Λύϥϝʔλͱͯ͠௥Ճ ˠ৚݅৘ใΛը૾ʹΑΓڧ͘൓ө • "EB-/;FSPCMPDLͰ͸ͦΕΒΛθϩʹॳظԽ ˠֶशͷॳظஈ֊͸߃౳ؔ਺ʹ͍ۙಇ͖ ˠ ֶशͷ҆ఆԽ

Slide 9

Slide 9 text

࣮ݧઃఆ • σʔληοτ • $MBTT$POEJUJPOBM*NBHF/FUY Y<%FOH $713> • ΞʔΩςΫνϟ • 7J5ͱಉ༷ʹͭͷϞσϧͷେ͖͞ 4 # - 9- Λ༻ҙ • QBUDITJ[FQ • %%1.TBNQMJOHTUFQT • ධՁई౓ • '*% T'*% *4 1SFDJTJPO 3FDBMM • (GMPQT • ֶश • 516WQPE #BUDITJ[F

Slide 10

Slide 10 text

ఆྔత݁Ռɿ6/FUϕʔεͷख๏Λ্ճͬͨ

Slide 11

Slide 11 text

ఆੑత݁Ռ • 1BUDITJ[FΛখ͘͞ɼϞσϧΛେ͖͘͢ΔͱΑΓࣗવͳը૾͕ग़ྗ͞ΕΔ ˠ%J5Ͱ͸(GMPQT਺͕େ͖͍΄Ͳग़ྗը૾ͷ඼্࣭͕͕Δ

Slide 12

Slide 12 text

௥ࢼ͓ΑͼΤϥʔ෼ੳ ఆੑత݁Ռ ɿࣦഊྫ • ಛఆͷMBCFMʹରͯ͠ෆࣗવͳը૾͕ੜ੒͞ΕΔ • ྫɿJOQVUMBCFM UPZQPPEMF %%1.TBNQMJOHTUFQ • ϥϕϧʹΑͬͯTUFQͰ͸ੜ੒ը૾͕ෆ҆ఆˠ ਪ࿦࣌ͷTUFQ਺Λಈతʹมߋ

Slide 13

Slide 13 text

ॴײ • 4USFOHUI • ֦ࢄϞσϧʹUSBOTGPSNFSΛಋೖ • ܭࢉࢿݯͱग़ྗը૾ͷ඼࣭ʹ͍ͭͯͷߟ࡯ • 8FBLOFTT • ਺͕ࣜগͳ͔ͬͨ • Τϥʔ෼ੳ͕ͳ͍

Slide 14

Slide 14 text

·ͱΊ • എܠ • ֦ࢄϞσϧʹΑΔಈը૾ੜ੒ FH 4PSB ͷൃల • ֦ࢄϞσϧʹ͓͚ΔUSBOTGPSNFSͷར༻͕গͳ͍ • ఏҊख๏ • USBOTGPSNFSϕʔεͷ֦ࢄϞσϧͰ͋Δ%JGGVTJPO5SBOTGPSNFS %J5 ΛఏҊ • ݁Ռ • %J5͸εέʔϥϏϦςΟ͕ߴ͘ɼ(GMPQT͕େ͖͍΄Ͳ'*%͕௿Լ ˠ ܭࢉࢿݯͱग़ྗը૾ͷ඼࣭ʹڧ͍૬ؔؔ܎ • %J59-Ϟσϧ͕ɼ$MBTT$POEJUJPOBM*NBHF/FUʹ͓͍ͯ ैདྷͷ6/FUϕʔεͷ֦ࢄϞσϧΛ্ճͬͨ

Slide 15

Slide 15 text

"QQFOEJYɿ%FOPJTJOH%JGGVTJPO1SPCBCJMJTUJD.PEFM %%1. ֶश

Slide 16

Slide 16 text

"QQFOEJYɿ$MBTTJGJFSGSFFHVJEBODF • ৚݅෇͖֦ࢄϞσϧͰ͸ΫϥεϥϕϧΛϥϯμϜʹυϩοϓ ˠ αϯϓϦϯάͷਫ਼౓Λ޲্ • #BZFTͷఆཧΑΓ • ֦ࢄϞσϧͷग़ྗΛείΞͱͯ͠ղऍ͢Δͱਪఆ͢ΔϊΠζ͸ҎԼͷΑ͏ʹͳΔ TɿΨΠμϯεεέʔϧ

Slide 17

Slide 17 text

"QQFOEJYɿ*ODPOUFYUDPOEJUJPOJOH • $POEJUJPOJOHͰ৚݅ͱͯ͠ೖྗ͞ΕͨτʔΫϯΛ ը૾τʔΫϯͷઌ಄ʹ௥Ճ • ͜ΕΒͷτʔΫϯ͸ը૾τʔΫϯͱಉ༷ʹѻΘΕɺ 7J5ʣʹ͓͚ΔDMTτʔΫϯͱࣅͨ໾ׂΛ࣋ͭ

Slide 18

Slide 18 text

"QQFOEJYɿ$SPTT"UUFOUJPOCMPDL • 4FMG"UUFOUJPOϒϩοΫͷޙʹ$SPTT"UUFOUJPOΛ ௥Ճͨ͠ઃܭ • <7BTXBOJ /*14>΍-%.ͱྨࣅͨ͠ΞʔΩςΫνϟ

Slide 19

Slide 19 text

"QQFOEJYɿ%J5CMPDLEFTJHO • %J59-Ϟσϧʹ͓͍ͯBEB-/;FSPΛ༻͍ͨ ৔߹͕࠷΋গͳ͍ܭࢉࢿݯͰ࠷΋ྑ͍ '*%,είΞΛୡ੒

Slide 20

Slide 20 text

"QQFOEJYɿ7JTJPO5SBOTGPSNFS <%PTPWJUTLJZ *$-3>

Slide 21

Slide 21 text

"QQFOEJYɿ*ODFQUJPO4DPSF *4 • *NBHF/FUͰࣄલֶशࡁΈͷ*ODFQUJPOOFUXPSLΛ༻͍ͨධՁࢦඪ • *ODFQUJPOOFUXPSL͕ࣝผ͠΍͘͢ɼࣝผ͞ΕΔϥϕϧͷଟ༷ੑ͕͋Δ΄Ͳ େ͖͘ͳΔࢦඪ <4[FHFEZ $713>

Slide 22

Slide 22 text

"QQFOEJYɿ'SFDIFU*ODFQUJPO%JTUBODF '*% • *NBHF/FUͰࣄલֶशࡁΈͷ*ODFQUJPOOFUXPSLΛ༻͍ͨධՁࢦඪ • ੜ੒͞Εͨը૾ͷಛ௃෼෍͕(5ը૾ͷಛ௃෼෍ͱͲͷఔ ౓ࣅ͍ͯΔ͔ΛධՁ͢Δࢦඪ • '*%͕খ͍͞΄Ͳੜ੒͞Εͨը૾ͷ඼࣭͕(5ը૾ʹ͍ۙͱߟ͑ΒΕΔ

Slide 23

Slide 23 text

"QQFOEJYɿ1SFDJTJPO3FDBMM • *NBHF/FUͰࣄલֶशࡁΈͷ7((<4JNPOZBO *$-3>Λ༻͍ͯ ಛ௃ϕΫτϧू߹ΛಘΔ

Slide 24

Slide 24 text

"QQFOEJYɿ(GMPQT • 'MPQTɿුಈখ਺఺ԋࢉͷճ਺ • (GMPQT 'MPQT • ը૾ੜ੒λεΫͰΞʔΩςΫνϟͷෳࡶ͞ΛධՁ͢Δࡍύϥϝʔλ਺Λ༻͍Δͷ ͕Ұൠత • ੑೳʹେ͖͘Өڹ͢Δը૾ղ૾౓ΛҰ੾ߟྀ͍ͯ͠ͳ͍ • Ϟσϧͷෳࡶ͞Λද͢ࢦඪͱͯ͠͸ෆे෼ͳ৔߹͕͋Δ

Slide 25

Slide 25 text

"QQFOEJYɿఆྔత݁Ռ

Slide 26

Slide 26 text

"QQFOEJY(GMPQTͱ'*%ͷ૬ؔ • ΑΓଟ͘ͷ(GMPQT਺Λ΋ͭϞσϧ͸'*%͕௿͘ͳΔ܏޲

Slide 27

Slide 27 text

"QQFOEJYɿϞσϧαΠζͱύοναΠζͷݕ౼