I-I Made Dall-E Ukulungiselela Izithombe ze-Children Ku-Images Ehlukile

A Personal Journey Through Childhood Drawing, Imagination, and the Magic of Generative AI

TL;DR: Yini uma i-AI ingakwazi ukwenza ngaphezulu kunazo - ukuba ungayifumana? Kule ukucubungula kwelinye, ngithole imodeli ye-AI generative DALL·E ukuguqulwa ama-sketches amancane, abesifazane - kufaka ne-sketches eyenziwe yi-child real - ku-livid, imiphumela artwork. Nge-way, ngithole ukuthi i-AI ayikho kuphela ukulibazisa umzimba wabantu; kungase ngempumelelo.

Ukusuka ekubunjweni kwe-pattern kuya ku-"hallucinations" ezihambayo ezihambelana nezimo, lo mkhuba likhuluza ukuthi i-Generative AI ingangena njengesikhumulo enhle kakhulu yokwenza konke - ikakhulukazi abantu abacwaningi ukufundisa ukuchithwa.

Where does imagination begin - and can a machine ever truly share in it?

Izithombe zabo zihlanganisa, zikhuthazayo, futhi ngokuvamile zihlanganisa abantu abadala - kodwa ngokuvamile zihlanganisa nezimo.

Yini uma I-Generative AI, isixhobo esekelwe ku-logic kanye ne-data, ingathola kwelanga elincinane le-pencil lines kanye nama-dreams eyenziwe?Waze?

Why Do Kids Love to Draw?

A lot of children love drawing. It is intuitive, creative, futhi fun. Kodwa lapha is a catch: amaninzi abesifazane ukujabulela ngokushesha kanye nomthombo ngenxa yokusebenza yabo.

Qiniseka umdlavuza ukuthi ukucindezelwa kwelanga ngokufanele kuxhomekeke iminyaka ye-skill-building. Akukho elula.

Izingane abanezingane zihlole ngokugcwele izakhiwo zayo abstract - izibonelelo zabo zihlole izixazululo. Ukuze izibane abanezingane, kubalulekile kuphela umugqa, kodwa kubo, kuyinto izindiza rocket, umamafutha, noma i-unicorn esifundeni.

Ngaphezu kwalokho, abantu abaneminyaka engaphezu kwengu-20 ngeveki zihlanganisa imifanekiso yayo. Njengomfundi, izigulane zihlanganisa izinga ezahlukile - futhi ezinye zihlanganisa kakhulu kunazo.

Drawing Is a Superpower for Child Development

Izinzuzo ze-drawing ye-children zihlanganisa kakhulu - futhi zihlanganiswa ngenzululwazi.

Ukuthatha, umzekeloUkusebenzaWazeDr. Richard Jolley and Dr. Sarah RoseUmbhali we-Staffordshire University, umbhali we-Staffordshire University, umbhali we-Staffordshire University, umbhali we-Staffordshire University, umbhali we-Staffordshire University, umbhali we-Staffordshire University, umbhali we-Staffordshire University, umbhali we-Staffordshire University, umbhali we-Staffordshire University, umbhali we-Staffordshire University, umbhali we-Staffordshire University, umbhali we-Staffordshire University.

"Ukuhlola kungasiza izidakamizwa. Izifundo zibonisa ukuthi ukusetshenziswa kwe-drawing njenge-activity yokufundisa kungasiza ukucaciswa kwama-children eminye izindawo, njenge-science."

"Ukuhlola kungasiza ukwandisa ukunakekelwa kwezingane. Ukuhlolwa kwandisa ukuthi izigulane zithumela ulwazi olungaphezulu malunga ne-event eduze ngaphambi kokufika lapho kubhalwe ukuba zithole mayelana nawe ngenhloso."

"Ukuhlola kungasiza izidakamizwa. Izifundo zibonisa ukuthi ukusetshenziswa kwe-drawing njenge-activity yokufundisa kungasiza ukucaciswa kwama-children eminye izindawo, njenge-science."

"Ukuhlola kungasiza ukwandisa ukunakekelwa kwezingane. Ukuhlolwa kwandisa ukuthi izigulane zithumela ulwazi olungaphezulu malunga ne-event eduze ngaphambi kokufika lapho kubhalwe ukuba zithole mayelana nawe ngenhloso."

Futhi iziphumo zithunyelwe ezingenalutho. Njengoba ngifundisa lokhu ibhizinisi ngaphezulu, ngifumana izifundo ezininzi ezivamile zihlanganisa ukuthi ukucindezelwa kusiza ukucindezeleka, ukucindezeleka, ukucindezeleka, kanye nokucindezeleka kwamadoda emzimbeni kwamadoda.

A Lightbulb Moment

Lezi zibonelelo zangaphendula umqondo yami. Umbuzo elilodwa, enhle wabhala:

If drawing is both enjoyable and beneficial - but often frustrating for kids - can I make it more magical? Could I motivate children to draw more often and with greater satisfaction?

Yini uma ungayitholela indoda yami ukuba i-artistNgamunye?

Ngingathanda kahle kakhulu ukuze zibonele. Kodwa ingcindezi - yes, itUkulungeleYenziwe

Enter Generative AI: Magic With a Sketch

Ngokusekelwa kwami ku-generative AI tools likeNgena ngemvumeNgiyaxolisa okuhle ngokushesha:

Umhlahlandlela kufuneka ukuguqulwa isikhwama elula kwizithombe ephakeme, enhle - uma kusetshenziselwa imiyalelo elula.

Why?Ngenxa yalesi amamodeli abafundiswe ku-datasets amancane abacwaningi ukuthi abacwaningi bafundisa indlela abantu abacwaningi amamodeli futhi abacwaningi.

Yinto eyodwa esithathwe kuyinto umyalezo olufanele (isib. Umyalezo olufanele).

My First Experiment - A Sunny Day

Ukubonisa umqondo yami, ngifake isikhwama esisheshayo ku-iPad yami - kodwa i-pencil kanye ne-paper ezivamile bakwazi ukusebenza kahle. Ngitholile ngempumelelo, kodwa iqukethe izinto eziningana eziningana ukuhlangabezana nezidingo ze-model.

Umhlahlandlela wama-skiz ye-rough, engaphelile, futhi ongaphandle kwemibala - kakhulu ku-spirit of a child's drawing. Nakuba lokhu inguqulo wama-digital, i-skiz ye-spontaneous, i-imaginative output ye-your could see from a young child.

Ukuze udlulisele imodeli, ngisebenzisa ucwaningo elilodwa - eyodwa owaziwa ngokuqondile emhlabeni wonke ukuhlolwa:

“Generate a realistic image based on this sketch, placing each element exactly where it appears in the original drawing.”

Umphumela wabanjwa okungenani okuhambisa.

More Than a Copy

Thola ukuyifaka ngokugqithisileyo umphumela ukuze ufunde ukuthi kungcono kakhulu.

Konke elihlanganisi evela ku-skiz ye-original ayikho kuphela ku-image eyenziwe - kodwa ibekwe ngokucacileyo kwe-spatial precision.Njengesikhathi esifundeniI-model ayikwazanga kuphela ama-components; i-respective yokufaka kanye ne-size, ngokuvumelana ne-composition cishe ngokucacileyo njenge-drawn.

Ukuze kwenziwe okuhle, ngifake i-grid elihlanganisa ukuvuselelwa phakathi kwe-skitch ne-output ekupheleni:

Kodwa thina siphinde kakhulu - ngoba DALL-E akuyona kuphela imiyalezo,UkukhishwaWaze

Ikhaya manje iqukethe ifenisha kanye nezindiza - izibuyekezo yangaphakathi, kodwa izongezelelo ezinhle kakhulu. I imodeli ngokucacisa ukuthi "isakhiwo" kufanele kuhlanganise futhi zihlanganisa izinga.

Kuhlolwe futhi emangalisayo emangalisayo emangalisayo emangalisayo - izinketho ezivamile, ezisebenzayo. Ngokuvamile, zonke imibala zihlanganisa ngokugcwele: i-solar is yellow, ama-clouds is white, futhi i-rainbow i-multicolor, ngokufanayo neyaziwa. Lokhu akuyona i-casuality - kuyinto imiphumela ye-training enhle, esisekelwe embhobho mayelana nokuthumela kanye nokuthanda ulwazi lomsindo.

Yini siphinde akuyona kuphela umfanekiso eyenziwe nge-AI - kuyinto umbhalo enhle, enhle, eyenziwa ku-skizche enhle.

Ukuze umfanekiso, ukukhiqiza umbhalo kuleli ngempumelelo. Futhi kunjalo, lapha kuyinto - enhle, ephelele, futhi yes, DALL-E ngisho akufanele ukufinyelela ngaphakathi.

The Science Behind the Magic

I-DALL-E inokukwazi ukukhiqiza umbhalo enhle kusuka ku-sketch enhle ngoba i-DALL-E iyatholakala ku-datasets eningi eyenziwe nge-millions of images ne-textual descriptions ehlanganisiwe.

Nge-training yayo, umzila ufundise ukujabulela imizila - akuyona kuphela indlela izindawo zihlanganisa, kodwa futhi indlela zihlanganisa futhi zihlanganiswa ngabantu.

Ngakho-ke, uma i-skiz, ngisho enhle, i-DALL-E ingathola ukuthi izakhi zihlanganisa (i-quadrat with a triangle on the top enokuthi inikeza "isakhiwo") futhi usebenzisa izazi zayo ezijoliswe ukuze zihlele imibuzo yobuchwepheshe ngempumelelo, ngokuvumelana ngempumelelo. I-DALL-E akufinyelela kuphela; inikeza, ukwandisa, futhi zihlanganisa umbhalo ngokuvumelana nezakhiwo ze-human-created content.

(Uma abafazi mayelana nokusebenza kanjani i-DALL-E 3, ngihambisa ngaphezulu ku-mechanics ye-model training ku-DALL-E 3.Imininingwane yami)

Sishayele ukucubungula i-flow emzimbeni ye-AI generative umzila ukucubungula inqubo ngaphezulu:

Okwamanje, ngiyathanda ngokwenene - futhi ngithanda ukukhuthaza imodeli ngaphezulu.

My Second Experiment - Frosty the Snowman

I-sketch elilandelayo yinkimbinkimbi enhle kakhulu, njengomphumela isinyathelo esiyingqayizivele: i-snowman. Lesi-elementi eyodwa inikeza i-saison eyodwa - i-winter - okuvumela isilinganiso se-extra ye-model.

Umphumela? Akukwazi ukujabulela.

I-DALL-E iyakwaziNge-Infer the Time of the YearUkuphendula ngokuvumelana ngokuvumelana nge-snowman. It is a logical and contextually accurate choice given that snow is a fundamental requirement for building a snowman. It is a logical and contextually accurate choice since snow is a fundamental requirement for building a snowman.

Njengoba ku-sketch edlule, imodeli iye yasungulwa ekhaya ephelele nge ukhasi, ifenisha, kanye nesikhumba esebenzayo. I-Snowman was brought to life with two stick arms, a scarf, a hat, and a carrot nose. A i-pine tree, ngokwemvelo i-dusted nge-snow, futhi ibandakanya - enye ingxubevange futhi efanelekayo.

Izinzuzo ezincinane ziye zihlanganiswa: umbhalo wahlanganisa inani elilodwa we-clouds kusuka ku-skizche yokuqala.

Kulabo abafuna ukujabulela, kukhona enye ingcindezi okungenani - i-flower. I-interestingly, lo mhlophe lihlanganisa ngama-creation yethu edlule, kodwa nge-colour shift kusuka ku-yellow kuya ku-red. Nakuba akuyona ingxenye ye-skiz ye-akhawunti, akuyona ingcindezi esisodwa. Lokhu kubonisa uhlobo le-"memory leak" kusuka ku-session edlule - i-quirk engaziwayo yokusebenza kwe-model.

Nokho, ku-grand scheme of things, lokhu kuyinto isizukulwane esincane (noma kwaba?) ku-output enhle futhi enhle kakhulu.

Umbuzo omncane wahlala ngokushesha:Yini uma ngithole isithombe yesibili e-DALL-E chat window entsha -Ukuguqulwa okufanayo kwe-context, indlela emzimbeni lethu ngokwemvelo lapho ukuguqulwa kwe-focus?

My Third Experiment - The Ultimate Challenge

Ngiyaxesha isikhathi sokuphendula kokuqala: ukusetshenziswa umbhalo we-in-life child.

Ngesikhathi eside, i-daughter ye-better friend yami, i-Naomi, iyakwazi ukunceda - futhi ngokushesha i-skizki elandelayo ngokuhambisana nami.

Ngingathanda, ngingathanda kancane. I-drawing yaba abstract futhi ephelele, nge-concrete shapes noma izinhlamvu ezivamile. Yenza ngokuphelele kwi-imagination ka-Naomi - lezi zihlanganisi akuyona emoyeni enhle, ngakho-ke imodeli awukwazi ukuxhaswa kwezimpendulo ezithakazelisayo eyenziwa ku-training.This was uncharted territory.

Ngiyazi ukuthi lokhu kuyinto test enhle. Uma DALL-E angakwazi ukucubungula ukukhiqizwa kwe-Naomi kanye nokuvumela ama-characters ehamba emangalisayo, akuyona kuphela emangalisayo - i-moment lapho ubuchwepheshe ikakhulukazi ukujabulela ukujabulela kwezingane.

Umphumela we-DALL·E ihlukanise konke okungenani angakwazi ukucinga:

I-model ayikho kuphela ama-shapes nama-face features kusuka ku-sketch ye-Naomi - itUkuhlobisaumugqa omncane, abstract ku-characters emangalisayo, ezibonakalayo.Yenza into kusuka ku-nothing.

Kuyinto ukwakhiwa kwangempela: ukuguqulwa kwe-visual of an idea that had previously existed only in Naomi's imagination. Lezi zihlanganisi wahlanganyela, akukho umugqa - nje isikhunta yobuciko umfundisi, manje wahlanganyelwa ku-AI.

Ngohlabathi ye-Generative AI, uma imodeli ibonise yini okuyinto ayatholakala ngokuvumelana, sicela ushiye "hallucination. "I-terms inesibopho emibi ye-negative - enikeza ingozi, ingozi emibi, noma ukuhlangabezana nezidingo ze-user.

Thola, isibonelo, i-flower esifundeni esifundeni etholakalayo esifundeni yethu ye-Snowman. Ngokuvamile, lokhu ingahlanganiswa njenge-hallucination enhle: isinyathelo esincane, esifundeni esifundeni esifundeni esifundeni esifundeni.

Kodwa nge-drawing ka-Naomi, i-DALL-E iye yenza into emangalisayo kakhulu - umculo owenziwe ngempumelelo futhi enhle - nge-DALL-E.UkuhlobisaIt was not copying, it was not referencing; it was imagining.

Oku kuthatha umbuzo enhle kakhulu:Ngaphandle kwe-contexts, kungenzeka ukuthi okuziwa nangokuthi i-hallucination kuyinto enhle kakhulu - njengeUkukhanya?

A Final Reflection

Okuqala njenge-test elula kwangaphambili: isikhokelo lapho ubuchwepheshe lithathwe emangalisayo - akuyona ukuguqulwa, kodwa ukuguqulwa.

Ukubuyekeza izilinganiso ze-Naomi zihlanganisa ku-living, ama-characters ezikhuthaza kakhulu kunzima nje ngokucindezeleka - kubonakala emangalisayo.

Ngohlabathi lapho izidakamizwa zihlola indlela yokubonisa okuhlobene nokuthanda, i-AI ingasebenza njenge-bridge phakathi kwami-imagination ne-reality, enikeza imibuzo emikhulu kakhulu ukuze izindandatho ezincinane zihlanganisa kuphela. Futhi mhlawumbe okuhle kakhulu - ukuhlola ukuthi uma kusetshenziselwa ngokucacileyo, i-AI ayidinga ukujabulela umzimba wabantu; inikeza, isikhwama esisodwa ngexesha.

mayelana nami

I-Maria Piterberg - ingcebisi ye-AI enikezela iqembu le-software ye-Runtime e-Habana Labs (Intel) futhi umbhali we-semi-professional abasebenza phakathi kwezidakamizwa ezivamile nezidijithali. Ngizinhlelo zokusebenza zokusebenza ze-AI emikhulu, kuhlanganise i-libraries ye-communication (HCCL) ne-runtime optimization. I-Bachelor of Computer Science.