1. Gabatarwa

Wannan takarda ta zayyana wani aikin PhD da ke gudana wanda ke bincika haɗa Cibiyoyin Adawar Samarwa (GANs) cikin ayyukan haɗin kai na zane-zanen tufafi. Babban jigo shi ne cewa GANs, maimakon maye gurbin ƙirƙirar ɗan adam, na iya zama abokan haɗin gwiwa waɗanda ke haɓaka tsarin ƙira. Aikin yana a tsakiyar haɗuwa tsakanin Mu'amalar Mutum-Kwamfuta (HCI), koyon na'ura mai samarwa, da nazarin ƙira. Yana neman amsa: "Ta yaya za a iya amfani da GANs a cikin haɗin kai, kuma ta yin haka, ta yaya za su iya ba da gudummawa ga tsarin zane-zanen tufafi?" Ta hanyar amfani da tsarin haɗin kai mai gauraye, binciken yana nufin fassara kaddarorin algorithm na GAN zuwa cikin mu'amala mai sauƙi, mai ma'amala wanda ke haɓaka haɗin gwiwa mai haɗin kai tsakanin mai zane da AI.

2. Bayan Fage & Ayyukan Da Suka Danganta

Aikin ya ginu akan wasu muhimman fagage na bincike da suka wanzu.

2.1. GANs a Cikin Yankunan Ƙirƙira

GANs sun nuna iyawa mai ban mamaki wajen samar da ingantattun abubuwa, sababbin abubuwa a yankuna kamar fasaha, fuskoki, da tufafi. Samfura kamar StyleGAN da CycleGAN sun kasance masu mahimmanci. Misali, tsarin CycleGAN don fassarar hoto zuwa hoto mara biyu, kamar yadda aka yi cikakken bayani a cikin babban takardunsa na Zhu et al. (2017), yana ba da tushen fasaha don aikace-aikacen canja salo waɗanda suka dace da tufafi sosai.

2.2. Kalubalen Akwatin Baƙi & Rashin Tabbaci

Babban cikas ga amfani da GAN a cikin ƙira na ƙwararru shi ne rashin fahimtarsu na asali. Rukunin sararin latent mai rikitarwa, mai haɗaka yana sa mai tsara ya yi wahala ko ya fahimta ko sarrafa tsarin samarwa bisa tsammani. Masu bincike kamar Benjamin et al. suna ba da shawarar kula da rashin tabbaci na koyon na'ura a matsayin kayan ƙira, suna nuna cewa "rashin tabbaci" na hanyoyin sadarwar jijiyoyi na iya zama tushen ƙwaƙwalwar ƙirƙira maimakon aibi da za a kawar.

2.3. Haɗin Kai Mai Gauraye

Wannan tsarin HCI yana mai da hankali kan tsarin da sarrafawa ke rabawa cikin sauri tsakanin wakilan mutum da kwamfuta, kowannensu yana ba da gudummawar ƙarfinsa na musamman. Manufar ba cikakken sarrafa kansa ba ne amma ƙarfafawa, inda AI ke sarrafa ƙirar ƙira da samarwa a sikeli, yayin da ɗan adam ke ba da babban niyya, hukunci na ado, da fahimtar mahallin.

3. Tsarin Aikin & Hanyoyin Bincike

3.1. Babban Tambayoyin Bincike

  • Ta yaya kaddarorin fasaha na GANs (misali, tsarin sararin latent, rugujewar yanayi) suke bayyana a cikin saitin haɗin kai mai ma'amala?
  • Wadanne tsarin mu'amala (misali, zane, na'urorin sifili na ma'ana, gyara bisa misali) suka fi dacewa wajen cike gibin tsakanin niyyar mai zane da samarwar GAN?
  • Ta yaya haɗin kai tare da GAN ke tasiri tsarin zane-zanen tufafi, ƙirƙirar mai zane, da sakamakon ƙarshe?

3.2. Tsarin Haɗin Kai da Ake Shawarawa

Tsarin da ake tsammani yana bin madauki mai maimaitawa: 1) Mai zane yana ba da shigarwar farko (zane, allon yanayi, umarni na rubutu). 2) GAN yana samar da jerin ƙirar ƙira. 3) Mai zane yana zaɓar, yin suka, da inganta 'yan takara, yana iya amfani da kayan aikin mu'amala don sarrafa sararin latent. 4) Ingantaccen fitarwa yana ba da labari ga zagayowar samarwa na gaba ko kuma an kammala shi.

4. Tushen Fasaha & Cikakkun Bayanai

4.1. Tsarin GAN & Sararin Latent

Aikin mai yiwuwa yana amfani da tsarin GAN mai sharadi ko na salo (misali, StyleGAN2) wanda aka horar da shi akan babban tarin hotunan tufafi. Babban ɓangaren shi ne sararin latent Z, ƙaramin yanki mai girma inda kowane batu z yayi daidai da hoton da aka samar. Kewayawa wannan sarari yana da mahimmanci ga sarrafawa.

4.2. Tsarin Lissafi

Babban manufar GAN shine wasan minimax tsakanin mai samarwa G da mai nuna bambanci D:

$\min_G \max_D V(D, G) = \mathbb{E}_{x \sim p_{data}(x)}[\log D(x)] + \mathbb{E}_{z \sim p_z(z)}[\log(1 - D(G(z)))]$

Don aikace-aikacen haɗin kai, hankali yana karkata zuwa koyon aikin taswira f daga shigarwar mai amfani (misali, zane-zane, sifofi) zuwa yankuna a cikin sararin latent: z' = f(Iuser), yana ba da damar samarwa mai jagora.

5. Tsarin Bincike & Misalin Lamari

Yanayi: Zana Tarin "Tufafin Maraice Mai Dorewa."

  1. Shigarwa: Mai zane ya loda allon yanayi tare da hotuna na laushi na halitta, silhouettes na lallausan, da palette na launuka na ƙasa. Sun kuma shigar da umarnin rubutu: "mai ladabi, ƙirar mara sharar gida, biophilic."
  2. Sarrafa AI: GAN mai nau'i-nau'i (misali, haɗa CLIP don rubutu da StyleGAN don hotuna) yana ɓoye waɗannan shigarwar zuwa cikin vector latent da aka haɗa, yana samar da bambance-bambancen ƙira 20 na farko.
  3. Inganta Dan Adam: Mai zane ya zaɓi bambance-bambance 3 masu ban sha'awa. Ta amfani da mu'amala tare da na'urorin sifili don sifofi kamar "tsari da gudana" ko "matakin kayan ado," suna daidaita hanyoyin latent da suka dace da waɗannan fasalulluka, suna ƙirƙirar sabbin gauraye.
  4. Fitarwa & Maimaitawa: Zaɓin ƙarshe shine babban ƙirar ƙira na sabbin ƙirar tufafi waɗanda ke haɗa niyyar ado ta farko tare da abubuwan tsari na AI da ba a zata ba, suna haɓaka lokacin ra'ayi.

6. Sakamakon Da Ake Tsammani & Hanyar Gwaji

6.1. Bayanin Ƙirar Samfuri

Samfurin mu'amala da aka ba da shawarar zai ƙunshi: zane don shigarwar farko/gyara; ɗakin zane na bambance-bambancen AI; panel tare da sarrafawa masu fassara don sarrafa sararin latent (misali, gano na'urorin sifili); da mai bin tarihi don ganin tafiyar haɗin kai.

6.2. Ma'aunin Kimantawa

Za a auna nasara ta hanyoyin gauraye:

  • Ƙididdiga: Lokacin kammala aiki, adadin maimaitawa zuwa ƙirar da ke gamsarwa, bambancin abubuwan da aka samar.
  • Mahimmanci: Tambayoyin mai zane da ke kimanta goyon bayan ƙirƙira da ake ganin, jin ikon yin aiki, da amfanin shawarwarin AI, ana nazarin su ta hanyar nazarin jigo.

7. Aikace-aikacen Gaba & Jagorori

Tasirin ya wuce HCI na ilimi. GANs masu haɗin kai masu nasara za su iya kawo juyin juya hali ga tufafi ta hanyar:

  • Democratizing Design: Rage shinge ga shiga ga masu zane masu zaman kansu.
  • Aiki Mai Dorewa: Ba da damar ƙirar ƙira ta zamani cikin sauri, rage sharar samfuran jiki.
  • Tufafi Na Musamman: Ƙarfafa dandamali na keɓancewa tare da taimakon AI akan buƙata.
  • Faɗaɗa Tsakanin Fannoni: Tsarin yana aiki ga ƙirar samfuri, gine-gine, da fasahar dijital.
Bincike na gaba dole ne ya magance rabuwar sararin latent don ingantaccen sarrafawa, mu'amala mai nau'i-nau'i (murya, karimci), da nazarin dogon lokaci kan yadda waɗannan kayan aikin ke sake fasalin aikin ƙwararru.

8. Ra'ayin Mai Bincike: Babban Fahimta & Zargi

Babban Fahimta: Wannan aikin ba game da gina mafi kyawun mai samar da hoto ba ne; bincike ne na dabaru cikin shawarwarin ikon yin aiki a zamanin ƙirƙirar AI. Ainihin samfurin shine sabon nahawu na mu'amala don haɗin gwiwar mutum-AI.

Kwararar Ma'ana: Hujja tana ci gaba da kyau daga gano matsala (yanayin akwatin baƙi na GANs) zuwa ba da shawarar tsarin mafita (haɗin kai mai gauraye) da takamaiman gwajin gwaji (tufafi). Ya gano daidai cewa ƙima ba ta cikin fitarwar AI kaɗai ba, amma a cikin tsarin da yake ba da damar.

Ƙarfi & Kurakurai: Ƙarfi: Mayar da hankali kan yanki mai ƙarfi, mai dacewa da kasuwanci (tufafi) yana da wayo. Yana kafa tambayoyin HCI na ka'ida a cikin aikin duniyar gaske. Yin amfani da tunanin "rashin tabbaci a matsayin fasali" shine sake fasalin ƙwararru na raunin ML na yau da kullun. Kurakurai Masu Muhimmanci: Shawarar tana da sauƙi a kan yadda za a cimma sarrafawa mai fassara. Kawai ambaton "haɗin kai mai gauraye" bai isa ba. Fannin yana cike da gazawar ƙoƙarin kayan aikin "ƙirƙirar AI" waɗanda masu zane suka watsar saboda mu'amalar ta ji kamar zato. Ba tare da wani ci gaba wajen sa sararin latent ya zama mai iya kewayawa ta ma'ana—watakila ta hanyar amfani da fasaha mai ƙirƙira kamar GANSpace (Härkönen et al., 2020) ko manufofin rabuwa a bayyane—wannan yana da haɗarin zama wani samfuri wanda bai kai ga amfani da ƙwararru ba. Bugu da ƙari, shirin kimantawa yana da alama na ilimi; yakamata ya haɗa da ma'auni daga masana'antar tufafi kanta, kamar daidaitawa tare da hasashen yanayin yanayi ko yiwuwar samarwa.

Fahimta Mai Aiki: Don wannan aikin ya sami tasiri, ƙungiyar dole ne:
1. Ba da fifiko ga Sarrafa fiye da Sabon Abubuwa: Yi haɗin gwiwa tare da masu zane-zanen tufafi masu aiki tun daga rana ɗaya don gina mu'amala mai maimaitawa wanda ya dace da samfuran tunaninsu, ba samfuran masu binciken ML ba. Kayan aikin dole ne su ji kamar kayan aiki masu daidaito, ba na caca ba.
2. Benchmark Dangane da Yanayin Fasaha: A kwatanta tsarin su na haɗin kai ba kawai da tushe ba, amma da kayan aikin kasuwanci kamar Adobe's Firefly ko dandamali masu tasowa kamar Cala. Wane ƙima na musamman ne tsarin su na ilimi ke bayarwa?
3. Shirya don Tsarin Halittu: Yi tunani fiye da samfuri. Ta yaya wannan kayan aikin zai haɗa kai cikin rukunin software na ƙira da suka wanzu (misali, CLO3D, Browzwear)? Hanyar karɓa ta hanyar haɗin kai mara tsangwama, ba aikace-aikacen kai da kai ba.

9. Nassoshi

  1. Goodfellow, I., et al. (2014). Generative Adversarial Nets. Advances in Neural Information Processing Systems 27.
  2. Zhu, J.-Y., et al. (2017). Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. IEEE International Conference on Computer Vision (ICCV).
  3. Karras, T., et al. (2020). Analyzing and Improving the Image Quality of StyleGAN. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).
  4. Benjamin, G., et al. (2021). Uncertainty as a Design Material. ACM CHI Conference on Human Factors in Computing Systems (CHI '21) Workshop.
  5. Härkönen, E., et al. (2020). GANSpace: Discovering Interpretable GAN Controls. Advances in Neural Information Processing Systems 33.
  6. Shneiderman, B. (2022). Human-Centered AI. Oxford University Press.
  7. Grabe, I., & Zhu, J. (2023). Towards Co-Creative Generative Adversarial Networks for Fashion Designers. CHI '22 Workshop on Generative AI and HCI. (The analyzed PDF).