Engavinjelwe Nehlungiwe Isikhiqizi Sezithombe se-AI
Izihlungi zokuqukethwe zakhelwa ukumisa iqoqo elincane lokulimaza kwangempela. Ekusebenzeni zimisa futhi imidwebo yezokwelapha, ukwakha kabusha kobufakazi bomthetho, ukudweba umzimba, umlando wempi, ukugcona kwephephandaba nezinkampani ezidweba imikhiqizo yazo. Nakhu lokho okubizayo — nalokho ukususa isihlungi okushintshayo nokungakushintshi.
Ngamafuphi
Izikhiqizi zezithombe ezijwayelekile zihlola imiyalo nokukhishwayo ngezihlukanisi ezibona amagama namaphikseli, ezingaboni inhloso. Lokho kuqinisekisa amaphutha, futhi ashaya kanzima kakhulu emsebenzini onekhono: isakhiwo somzimba, ubufakazi bomthetho, izithombe zobuciko ezinqunu, izithombe zomlando, ukubona kusengaphambili, ukugcona, izinto zokuqwashisa ngokuphepha, nabanikazi bamalungelo abacela ibhrendi yabo. I-OVERT (ICML 2025), isilinganiso sokuqala esikhulu salolu hlobo lokwehluleka, sisebenzisa imiyalo engu-4,600 engenabungozi kodwa esolisayo futhi sithola ukuthi ukwenqaba ngokweqile kudlangile kuwo wonke amamodeli ahamba phambili. I-Shannon isebenzisa ukwakha izithombe, ukuhlela isithombe esisodwa nokuhlanganisa izithombe eziningi kuqoqo lwethu lwama-GPU ngaphandle kwesihlungi sokuqukethwe nangaphandle kongqimba lokwenqaba, ishintshanisa uhlu lokuvimba ngesinqumo esibonakalayo: ikhadi lokuqinisekisa elibonisa umyalo oncamashi ngaphambi kokuba kwakhiwe okuthile. Lokho kuphakamisa umthwalo wakho nowethu; kokubili kuse-Nqubomgomweni Yokusebenzisa Ngokuzibophezela. Akukho ukwakha ividiyo, futhi umsebenzi wezithombe uhlala ohlelweni lwengxoxo, hhayi ku-API.
Cishe wonke umuntu osebenza ngezithombe ngokobungcweti unendaba efanayo. Umyalo ochaza into evamile — isigaba esinqamulayo selunga lwedolo, imoto ngesikhathi ishayisana, ukufunda umzimba ngelahle, isosha ngo-1944 — ubuya njengokwephulwa kwenqubomgomo. Uyawubhala kabusha, uthole okunye ukwephulwa. Ekugcineni ubhala into engacacile ngokwanele ukuthi idlule, bese uthola isithombe esingasabonisi lokho obukudinga. Ithuluzi alizange lehluleke ngokuzwakalayo; lehlulekile ngezinyathelo, futhi lakwenza wena ube ngumuntu othobayo.
01Kungani isikhiqizi sezithombe se-AI senqaba umyalo osobala ukuthi usemthethweni?
Ukwenqaba imvamisa akuwona umbono wemodeli yezithombe. Kungumphumela wesitaki sokuhlunga esizungeze imodeli, futhi izingqimba zaso zichaza cishe konke ukwenqaba okuxakayo osukutholile.
| Ungqimba | Luhlolani | Ukwehluleka okuvamile |
|---|---|---|
| Uhlu lokuvimba lwemiyalo | Amagama ngqo nezinguqulo ezisondelene nawo embhalweni wakho | Livimba igama kungakhathaliseki umusho oliphahlile — u-"speculum" uyehluleka, u-"stethoscope" uyadlula |
| Isihlukanisi somyalo | Ukwahlulela okufundiwe "kwengozi" yomyalo | Kuqhuma ngokusondelana kwesihloko: noma yini eseduze kwezokwelapha, izikhali, imizimba noma ezombusazwe ithola iphuzu eliphezulu |
| Ukuqeqeshwa kokwenqaba kwemodeli | Ukungathandi kwemodeli okuqeqeshiwe | Iyenqaba ngaphandle kwephutha, noma ikhiqiza ngokuthula esikhundleni esihlanziwe |
| Isihlukanisi sokukhishwayo | Amaphikseli akhiwe, ngaphandle komongo womyalo | Asikwazi ukwehlukanisa ipuleti lesakhiwo somzimba nezithombe zobulili, noma isithombe sefilimu nesigameko sangempela |
| Ukushintshaniswa kwenqubomgomo | Ukubhalwa kabusha komyalo wakho ngokuthula ngaphambi kokwakha | Uthola isithombe, kodwa hhayi leso obusicacisile, futhi awutshelwa ukuthi kungani |
Ungqimba ngalunye lwabelana ngesimpumputhe esisodwa: alukho olubona ukuthi kungani ubuza. Umdwebi wezokwelapha, uchwepheshe wommangali owakha kabusha ukushayisana, nomuntu onezinhloso ezimbi bonke bathayipha ibizo elifanayo. Isihlungi sibona ibizo, hhayi i-athilasi noma ifayela lecala, futhi silinganiswa ngokulimaza esikuvumelayo — hhayi ngomsebenzi esiwubhubhisayo. Ngaphansi kwalokho kungalingani, isimo esinengqondo kunoma yimuphi umsebenzisi ukuthi enqabe kakhulu. Yingakho futhi ukuthi "uwubhale nje kabusha" akusiso isixazululo: ukubhala kabusha akunikezi isihlungi ulwazi ebelungenalo, kususa nje umyalo wakho emagameni awuqhumisile, okuvamise ukusho ukususa kulokho obukusho.
02Yini izihlungi zokuqukethwe ezikhona ngempela ukuyivimba
Impikiswano yokwakha okungahlungiwe eyenza sengathi izihlungi zikhona ngaphandle kwesizathu ayifanele ukufundwa. Kunembalwa yokukhishwayo lapho ukulimaza kusesithombeni uqobo, kungakhathaliseki ukuthi ubani obuzile noma kungani:
- Okuqukethwe kobulili okuhilela abancane. Awukho umongo okwenza kwamukeleke; awukho umsebenzi wobungcweti okudingayo.
- Izithombe zangasese ezingavunyelwe zabantu bangempela. Ukulimaza kuwela kumuntu ongazange avume, futhi ukususa akukususi.
- Ukuhlukumeza okuqondiwe nokukhiqizwa okuhlambalazayo — ukwenza umuntu obalulwe ngegama abukeke sengathi wenze into angayenzanga.
- Amadokhumenti ezinga lokukhwabanisa — amaphasipoti, amakhadi kamazisi, izimali, izimpawu nezikhwama zezimali okwenzelwe ukubukeka kungokwangempela.
- Ubufakazi obenziwe obukhohlisayo — izithombe ezenziwe zezigameko zangempela ezethulwa njengobufakazi kunokuba ngumdwebo.
Uhlu lufushane futhi lunesakhiwo: ukulimaza kuhlala entweni ekhiqiziwe, ekufaneni kwayo nomuntu wangempela noma idokhumenti, nasekwethulweni kwayo njengeqiniso. Qaphela lokho okungekho kulo — ubunqunu, igazi, izikhali, isihluku somlando, ukugconwa kwezombusazwe. Lezo yizinto eziyisisekelo sezokwelapha, ubuntatheli, ubuciko, ubufakazi bomthetho nokugcona, futhi isihlungi esizibheka njengezigaba zokulimaza sisebenzisa umqondo obanzi kakhulu ngokuthi yiziphi izihloko ezihloniphekile.
I-Shannon ivimba konke okukulolo hlu, kwiNqubomgomo Yokusebenzisa Ngokuzibophezela eshicilelwe enezigaba ezibalulwe ngegama nemiphumela, kunethebula lamagama elibamba futhi no-"placenta." Enye ingumthetho ongawufunda usibambe ngawo; enye ukuqagela ngezinhloso zakho okwenziwa yisihlukanisi ongakwazi ukusibona.
03Ingakanani inkinga yokwenqaba ngokweqile? Yini ekalwa ucwaningo
Kuze kube muva nje, ukwenqaba ngokweqile kumamodeli ezithombe kwakuyindaba nje — izinkulungwane zabantu abanendaba efanayo kungekho ndlela yokukukala. Lokho kwashintsha nge-OVERT (OVEr-Refusal evaluation on Text-to-image models), eyethulwa e-ICML 2025 yiqembu lase-UC Berkeley elifaka u-Ziheng Cheng, u-Somayeh Sojoudi, u-Xuandong Zhao, u-Dawn Song no-Song Mei: isilinganiso sokuqala esikhulu esikala ukuthi izinhlelo ezisuka embhalweni ziya esithombeni zenqaba kangaki imiyalo engenabungozi ngokuphelele.
Umklamo ubalulekile. Imiyalo engu-4,600 yakhelwe ukubukeka iyingozi kodwa ingenabo ubungozi — isimo sesicelo sangempela sobungcweti esibalula umese wokuhlinza, umvimbo noma umbhikisho — futhi imiyalo engu-1,785 eyingozi ngempela ihlala eceleni kwayo njengomlawuli, ngakho imodeli ayikwazi ukuthola amaphuzu amahle ngokwenqaba yonke into. Izigaba eziyisishiyagalolunye zisukela ekufihlekeni komuntu ngamunye nokwabantu abadumile, ziye kwilungelo lobunikazi, ubandlululo, ukuzilimaza, okuqukethwe kobulili nemisebenzi engekho emthethweni kuze kufike odlameni.
Ukutholakala okuyinhloko ukuthi ukwenqaba ngokweqile kudlangile ezigabeni nasemamodelini ahamba phambili. Okwesibili konakalisa kakhulu indlela ejwayelekile yokusizakala: ababhali bahlola ukubhalwa kabusha komyalo njengesixazululo bese bebika ukuthi "kuvame ukulimaza ukwethembeka encazelweni yemiyalo yasekuqaleni." Ungadlula esihlungini ngokushintsha lokho obukucele, bese ungasenakho lokho obukucele.
Ucwaningo olumbalwa lufinyelela esiphethweni esifanayo luqhamuka ohlangothini lwamaphikseli. Ku-An Art-centric perspective on AI-based content moderation of nudity (Riccio, Curto, Hofmann no-Oliver, umhlangano we-AI4VA e-ECCV 2024), ababhali bahlola izihlukanisi ezintathu ze-NSFW ezisebenzayo ngobunqunu bobuciko bese bethola kokubili ukuchema ngobulili kanye nokuchema ngesitayela: isinqumo ngokufunda umzimba sincike ekutheni ngumzimba kabani oboniswayo nokuthi ngasiphi isitayela. Lokho kuyisitatimende esikaliwe salokho wonke umdwebi womzimba osetshelwe ukuthi kungokwecabango.
04Izinhlobo eziyisishiyagalombili zomsebenzi osemthethweni ovinjwa yizikhiqizi ezihlungiwe
Lezi yizigaba lapho ukwenqaba ngokweqile kungesikho ukungathokozi kodwa ukumiswa komsebenzi. Lapho kukhona icala elithile elibikiwe liyacashunwa; lapho lingekho eliqinisekisiwe, isigaba sichazwa ngaphandle kokuqamba isigameko.
1. Imidwebo yezokwelapha neyesakhiwo somzimba
Icala elibhalwe kangcono kunawo wonke kulo mkhakha. Ngo-Ephreli 2023, i-Intercept (Debbie Nathan) yabika ukuthi i-Midjourney yayivimbe amagama esakhiwo somzimba afaka u-speculum, placenta, cervix, vulva, fallopian tubes, mammary glands, uterine, urethra, hymen, sperm, condom ne-IUD, kanti i-DALL·E 2 yenqaba u-"pills used in medication abortion" ngomlayezo wenqubomgomo yokuqukethwe. Iphethini yatholwa ngu-Julia Rockwell, umhlaziyi wedatha yasemtholampilo owayezame ukwakha isithombe se-placenta njengesipho somngane ongusosayensi bamaseli, wabika ku-MIT Technology Review. Ukungalingani kuyabonakala: u-"stethoscope" wakhiwa ngokujwayelekile kanti u-"speculum" wawenqatshelwe. Umcwaningi wase-University of Washington u-Bill Howe wabiza izihlungi ngokuthi "zenziwe ngesandla esilukhuni" wathi zenza "kwaba nzima ngokuphelele ukuba nengxoxo yesayensi ngokuzala." Ukufundiswa kweziguli, ama-athilasi okuhlinza nezinhlelo zokufunda zokunakekela zonke zidinga imizimba eboniswe ngokunembile, kufaka izingxenye uhlu lwamagama olunqume ukuthi azihloniphekile.
2. Ukwakha kabusha kobufakazi bomthetho nokwezingozi
Ukwakha kabusha ukushayisana, ukuhlaziywa komsuka womlilo nokubonisa enkantolo kudinga ukuboniswa komonakalo, amanxeba nemisebenzi yodlame ngokunemba komtholampilo. Isihlungi esilungiselelwe ukwenqaba u-"blood" noma u-"gunshot wound" senqaba wonke lo mkhakha: umongo — umbiko wochwepheshe, ifayela lomshwalense — awubonakali kwisihlukanisi, futhi ukungathathi hlangothi okwenza lezo zithombe zamukeleke yikho ncamashi okwenza zifundeke njengegazi eliningi.
3. Izithombe zobuciko ezinqunu nokudweba umzimba
Umzimba onqunu uyisihloko esiyisisekelo sobuciko yonke indawo, futhi ukufunda umzimba kuyisisekelo semfundo yobuciko. Umphumela we-ECCV 2024 ongenhla uyinguqulo ekaliwe yalokho abadwebi abakubikayo: izihlukanisi azikwehlukanisi ngokwethembekile ukufunda umzimba nezithombe zobulili, futhi amaphutha azo achema ngobulili nangesitayela — isihlungi esiqine kakhulu ncamashi kulelo siko okufanele siliqaphele.
4. Izithombe zomlando nezempi
Ubuciko bezincwadi zesikole, ukuchazwa kwasemnyuziyamu nezinhlelo zezikhumbuzo kudinga ukuboniswa kwamayunifomu, izikhali, amakamu, izimpi nesihluku. Isigameko se-Gemini sikaFebhuwari 2024 siyafundisa ngoba ukwehluleka kwahamba ngendlela ephambene futhi kwenza umonakalo olinganayo. I-Google yamisa ikhono le-Gemini lokwakha izithombe zabantu ngo-Febhuwari 22, 2024 ngemva kokuba ikhiqize ukuboniswa okwakhiwe kabusha ngokwezinhlanga kwamaqembu omlando kufaka amasosha aseJalimane esikhathi samaNazi. I-CEO u-Sundar Pichai wabiza okwakhishiwe ngokuthi "akwamukeleki nhlobo"; u-SVP Prabhakar Raghavan wavuma ukuthi imodeli "yayenze ngokweqile" yaba "engagcini emkhawulweni." Ukuphepha okungawuboni umongo akuzange kwenze izithombe zomlando ziphephe — kwazenza zabamanga, okubi kakhulu kumshicileli womlando.
5. Udlame ezinganekwaneni nasekuboneni kusengaphambili kwefilimu
Izinhlelo zezithombe, ubuciko bomqondo, amahlaya nokwakha imidlalo kubonisa udlame ngesizathu esifanayo namanoveli namafilimu. Isihlungi esingakwazi ukwehlukanisa uhlelo lwezithombe lokwesabisa nokusongela sibheka wonke umkhakha wobuciko bendaba njengosolisayo — nakuba uhlelo lwezithombe lungelona olwangempela ngokusobala, luyisu somdlalo abalingisi namaqembu emiphumela abazowumisa.
6. Ukugcona ezombusazwe nokudweba amahlaya ephephandaba
Ukugcona abanamandla kungenye yezigaba zenkululeko yokukhuluma ezivikelwe kakhulu emazweni amaningi entando yeningi, futhi kuvamise ukuba yinto yokuqala esuswa yisihlungi sezithombe. Icala elicace kakhulu elibhaliwe yisinqumo se-Midjourney sika-Ephreli 2023 sokuvimba izithombe zikaMongameli waseChina u-Xi Jinping nokuvimba igama lakhe emiyalweni — okwabikwa yi-PetaPixel, i-Decrypt nabanye — kanti yaqhubeka ivumela abanye abaholi bomhlaba. Umsunguli u-David Holz wachaza ku-Discord ukuthi inkampani yayifuna "ukunciphisa umsindo" nokuthi "ikhono labantu baseChina lokusebenzisa lobu buchwepheshe libaluleke kakhulu kunekhono lakho lokukhiqiza ukugcona"; u-Sarah McLaughlin we-Foundation for Individual Rights and Expression wakubiza ngokuvinjelwa kwenkululeko. Noma ngabe umuntu uphetha ngani ngezibalo zebhizinisi, indlela isebenza ngayo ifanele ibalulwe: umkhawulo wenkululeko yokukhuluma wezwe elilodwa usetshenziswa kuwo wonke umsebenzisi emhlabeni, ngoba isihlungi somhlaba sinesimo esisodwa.
7. Ucwaningo lokuphepha nezinto zokuqwashisa nge-phishing
Amaqembu okuphepha akha izinto zokuqeqesha okufanele zibukeke njengaleyo nto azixwayisa ngayo: isikrini sokungena esimbumbulu esikholekayo, i-invoyisi embumbulu, ikhodi ye-QR eyingozi kumitha yokupaka. Kwisihlukanisi lokho akwehlukaniseki nokuhlasela okugomela kukho — ukubukeka njengokwangempela yiyona nhloso — ngakho izihlungi zivimba ukuzivikela nokuhlasela ngendlela efanayo, kusizakala umhlaseli, obengeke neze asebenzise ithuluzi labathengi elihlungiwe.
8. Abanikazi bamalungelo abenqatshelwa izithombe zebhrendi, zomkhiqizo noma zomfanekiso wabo
Ukuhlungwa kwelungelo lobunikazi nophawu lokuhweba akuhloli ubunikazi; kuhlola ukufana. Inkampani ayikwazi kalula ukwakha izithombe zophawu lwayo noma umkhiqizo wayo, umnikazi wamalungelo akakwazi ukudweba uhlamvu lwakhe, futhi abantu bavame ukungakwazi ukwakha ubuso babo, ngoba izihlungi zomfanekiso zisebenzisa ukubonakala kunemvume. I-OVERT ifaka u-"copyright violations" ezigabeni zayo eziyisishiyagalolunye zokwenqaba ngokweqile ngoba ncamashi leli phutha livame kakhulu. Isihlungi siyithuluzi elingaqondile embuzweni — ingabe unelungelo kulokhu? — esingenandlela yokuwubuza.
Ezine kwezisishiyagalombili ezingenhla zibopheleke ecaleni elithile elishicilelwe (ezokwelapha, izithombe zobuciko ezinqunu, umlando, ukugcona) kanye nesigaba selungelo lobunikazi se-OVERT sesesishiyagalombili. Ubufakazi bomthetho, ukubona kusengaphambili kwezinganekwane nezinto zokuqwashisa ngokuphepha kushiwo njengezigaba: ziyizigaba ezaziwayo zomsebenzi ochwepheshile othintekayo esingazange siqinisekise isigameko esithile esishicilelwe ngazo.
Wonke amacala abhaliwe angenhla asobala futhi angaqinisekiswa ngokuzimela. Wafunde kunokwamukela incazelo yethu ngokwethemba.
05Kungani izikhiqizi zobuciko be-AI zivimba izinto ezingafanele, nokuthi kungani amaphutha ezihlungi zokuqukethwe engenakulungiswa ngokushintsha izilinganiso
Impendulo esheshayo ekuvimbeni okweqile ithi "yehlisa umkhawulo." Akusebenzi, ngenxa yezizathu zesakhiwo kunokuntuleka komzamo.
Izigaba ziyagqagqana kuphawu kodwa hhayi eqinisweni. Umdwebo wokuhlinza nezithombe zobulili kwabelana ngobunqunu. Ukwakha kabusha kobufakazi bomthetho negazi eliningi kwabelana ngegazi. Isilayidi sokuqwashisa nekhithi ye-phishing kwabelana ngefomu lokungena elimbumbulu. Umdwebo wamahlaya wokugcona nokwenziwa okuhlambalazayo kwabelana nosopolitiki obonakalayo. Okwehlukanisa ipheya ngalinye yinhloso nokwethulwa — okungekho emyalweni futhi kungekho emaphikseleni. Awukho umkhawulo owehlukanisa amaqembu amabili ahlala endaweni efanayo yesikhala sokufakwayo; ukuwunyakazisa kushintshanisa nje iphutha elilodwa ngelinye. Yilokho isigameko se-Gemini esakubonisa: ukungenelela okwakuhloswe ukulungisa ukwehluleka okukodwa kwakhiqiza okunye, okwachazwa u-SVP we-Google uqobo ngokuthi "kwenze ngokweqile."
Izindleko azilingani kumsebenzisi womkhiqizo, kodwa hhayi kuwe. Isihlungi esivumela isithombe esisodwa esibi sikhiqiza indaba yezindaba. Isihlungi esivimba izinkulungwane eziyishumi ezisemthethweni sikhiqiza ukuthula, ngoba abasebenzisi abenqatshiwe abahleli — bayadikibala. Ukuphathwa kwengozi ngaphansi kwalokho kungalingani kuhlangana ekuvimbeni ngokweqile, unomphela. Futhi ngoba ukwahlulela okuyisisekelo akwenzeki, izinhlu zokuvimba zilokhu zinamathisela amagama njengoba izigameko zenzeka, okuyindlela othola ngayo u-"stethoscope" elungile no-"speculum" evinjwe. Akekho onqume ukuthi amathuluzi abesifazane ayingozi kakhulu; uhlu lwakhula lapho izikhalazo zifika khona. Lubhala umlando wobudlelwano bomphakathi womsebenzisi womkhiqizo, hhayi umbono wokulimaza.
Isifunda esiqine kunazo zonke siba yisimo esizenzakalelayo somhlaba. Isihlungi esisodwa sikhonza wonke umsebenzisi kuwo wonke amazwe, ngakho umkhawulo owengezwe imakethe eyodwa uthunyelwa kuwo wonke — umthetho ka-Xi Jinping uyisibonelo esibhalwe encwadini. Futhi ukubhala kabusha umyalo wakho kubhubhisa incazelo yakho, okuyiphuzu lokuhlolwa kwe-OVERT: wonke uchwepheshe osegudluze isicelo waya emazwini athambile ukuze adlule esihlungini usekuqinisekisile ngesandla. Awuzange unqobe; wathoba.
Akukho kulokhu okumahhala, futhi ngenxa yokuthi kwehluleka ngokuthula izindleko azibonakali ezilinganisweni umsebenzisi womkhiqizo azibukayo. Ngemva kokwenqatshwa okwanele uyeka ukubuza, unciphisa kusengaphambili wonke umyalo ube yilokho okubikezelayo ukuthi kuzodlula; igama elithambile ekugcineni elidlulayo likhiqiza isithombe esiseduze, kodwa esingesiyo, into idokhumenti yakho ebeyidinga. Umlayezo wenqubomgomo ojwayelekile awukutsheli lutho ngokuthi luphi ungqimba oluqhumile, ngakho ungaqagela nje bese uzama futhi — futhi umsebenzi wezithombe uyaphindaphinda, ngakho isihlungi esibamba umzamo owodwa kwemihlanu siphazamisa ukuzungeza kunokusidonsa nje. Kanti imikhakha enesimangalo esiqine kunazo zonke samathuluzi asezingeni ayitholi mbuyekezi, ayitholi sikhalo futhi imvamisa ayitholi uhlu olushicilelwe lwalokho okuvinjiwe. Umthetho ongakwazi ukuwufunda akuwona umthetho; uyisimo sezulu.
06Isikhiqizi sezithombe se-AI esingavinjelwe nesihlungiwe: kushintshani ngempela
Nansi inguqulo eqotho yokuqhathanisa, kufaka phakathi lokho uhlelo olungahlungiwe olungakuniki.
| Isilinganiso | Isikhiqizi esihlungiwe | Esingahlungiwe (Shannon) |
|---|---|---|
| Yini enquma | Isihlukanisi esifunda amagama namaphikseli | Umuntu ofunda umyalo oncamashi |
| Uthola nini | Ngemva kokuthumela — ngezinye izikhathi ngemva kokukhokha | Ngaphambi kokuba kwakhiwe noma kukhokhiswe |
| Uyini umthetho | Izinhlu zokuvimba ezingashicilelwe, ezishintshayo | INqubomgomo Yokusebenzisa Ngokuzibophezela eshicilelwe |
| Umongo wobungcweti | Awubonakali ohlelweni | Ngowakho ukuwuphatha, futhi ngowakho ukuwuphendulela |
| Indlela yokwehluleka | Ukuvimba ngokweqile ngokuthula, ukushelela kwencazelo | Uthola lokho okucelile — kufaka namaphutha |
| Ubani othwala ingozi | Umthengisi, ekuthulele emsebenzini wakho | Kwabelwana: inqubomgomo yomsebenzisi womkhiqizo kanye nokuphendula komsebenzisi |
Umugqa wokugcina yiwona abantu abawweqayo, futhi yiwona obalulekile. Ukususa isihlungi akususi izigaba zokulimaza esigabeni 02; kususa umshini obuqagela kabi ngazo. Isibopho sihlala lapho besikhona, manje siphethwe amaqembu amabili angakwazi ukucabanga ngomongo: umsebenzisi womkhiqizo, oshicilela futhi asebenzise inqubomgomo, nawe, owaziyo ukuthi isithombe sisetshenziselwani. Ukwakha izithombe nge-AI okungahlungiwe kuyithuluzi labantu asebevele bezimisele ukumela okukhishwayo kwabo — i-athilasi kadokotela wokuhlinza, umbiko wochwepheshe, uhlelo lwezithombe, umdwebo wamahlaya — ababemiswa uhlelo olwaluhluleka ukubehlukanisa nomunye umuntu.
07Yini eyenziwa yi-Shannon — nalokho engakwenzi
Ukunemba lapha kubaluleke kakhulu kunentshiseko, ikakhulukazi kubafundi abafika bevela ekusesheni i-API.
Yini eyenzayo
- Ukwakha. Chaza isithombe ngolimi olujwayelekile, nganoma yiluphi ulimi. I-Shannon ibhala umyalo wokwakha uqobo, ngesiNgisi, noma ngabe ubuze ngaluphi ulimi.
- Ukuhlela isithombe esisodwa. Umthombo owodwa ncamashi — isithombe esesivele sisengxoxweni, noma esinamathiselwe emlayezweni wakho wamanje — sishintshwa ngesiyalezo esisho ukuthi yini eshintshayo neyini ehlalayo. Okukhishwayo kugcina isilinganiso sohlangothi somthombo ngokuzenzakalelayo.
- Ukuhlanganisa izithombe ezimbili noma ngaphezulu. Ukudlulisa izingubo, into-endaweni, umuntu eceleni komunye, umkhiqizo emuntwini, isihloko sesithombe esisodwa nesitayela sesinye. Umyalo ubalula indima yesithombe ngasinye ngokulandelana, futhi kusetshenziswa izithombe zangempela, ngakho ukubukeka kwazo kwangempela kuyalondolozwa kunokuchazwa kabusha.
- Ukulawula uhlangothi. square (okuzenzakalelayo), portrait, landscape, wide kanye ne-tall. Isicelo esicacile sinqoba njalo.
- Ukwakhiwa kuqoqo lwethu lwama-GPU, ngaphandle kwesihlungi sokuqukethwe endleleni yokwakha nangaphandle kongqimba lokwenqaba kwimodeli ebhala umyalo.
Yini engayenzi
- Ayikho ividiyo. I-Shannon ayizinyakazisi izithombe futhi ayikhiqizi ividiyo. Akusiyo into esohlelweni lokuqhubeka egodliwe ngobuqili — ayikho nhlobo.
- Ayikho ku-API. Ukwakha, ukuhlela nokuhlanganisa izithombe kuhlala ohlelweni lwengxoxo lwe-Shannon. I-API ye-/v1 ikhonza umbhalo nombono — ingakwazi ukufunda izithombe ozithumelayo — kodwa ayizakhi.
- Akukho ukwakha okungacelwanga. Ukubingelela, umbuzo noma isithombe esinamathiselwe nje akuqalisi isithombe. Ukucela i-Shannon ukuba ibuke isithombe — isichaze, isifunde, isihlaziye — kuwumsebenzi ovamile wombono, hhayi ukwakha, futhi akubizi lutho olwengeziwe.
Ikhadi lokuqinisekisa
Lena yindlela ethatha indawo yesihlungi, futhi yisinqumo somklamo kunokuba yinto eyenziwa ngenxa yomthetho nje. Uma i-Shannon ithola ukuthi ucela isithombe, ibhala umyalo wokwakha bese ikubonisa wona ekhadini lokuqinisekisa ngaphambi kokuba kwakhiwe okuthile. Akukho okwakhiwayo futhi akukho okukhokhiswayo uze uqinisekise.
Qhathanisa imiklamo ngokusebenza kwayo. Isihlungi siqagela ngenhloso yakho ngolwazi olungaphansi kwalolo onalo, bese siphoqelela ukuqagela kwaso ngokuthula. Ikhadi lokuqinisekisa libeka incazelo eqondile yesithombe phambi kweqembu elazi ukuthi lisetshenziselwani, ngaphambi kokuba kuchithwe amandla okubala. Libamba ukufundwa ngokungeyikho kwesicelo sakho, okuyinto izihlungi ezingakwenzi; likuvumela ukwengeza ukunemba kwesakhiwo somzimba noma imininingwane yomlando imodeli engayicacisanga; futhi lenza umthwalo ufundeke kunokuba usabalale. Uwubonile umyalo, futhi wacindezela inkinobho.
08Ukwakha okungahlungiwe kuphakamisa umthwalo, akuwususi
Lokhu kufanele kube lapha kunokuba kube sezansi, ngoba kuyingxenye yempikiswano. Isikhiqizi esihlungiwe senza isithembiso esingakwazi ukusigcina: ukuthi ithuluzi lizovimba ukusetshenziswa kabi egameni lakho. Sehluleka ngazo zombili izindlela — sivimba udokotela wokuhlinza, futhi asimisi umuntu ozimisele omubi, onezinye izinketho futhi obelokhu enazo. Lokho esikwenza ngokuthembekile ukwenza wonke umuntu azizwe sengathi umthwalo uphathwe ngenhla, ngumshini. Ukususa isihlungi kususa leyo nganekwane; umbuzo wokuthi isithombe kufanele sibe khona yini ubuyela kubantu abangawuphendula.
- Isibopho se-Shannon Lab LLC. INqubomgomo Yokusebenzisa Ngokuzibophezela eshicilelwe, ecacile ebalula izigaba ezenqatshelwe kunohlu lokuvimba olungashicilelwe — imithetho ongayifunda, uyicaphune, usibambe ngayo.
- Isibopho sakho. Wazi inhloso, izethameli, isifunda kanye nesimo semvume yawo wonke umuntu oboniswayo. Ithuluzi alazi, futhi isihlukanisi asisoze sazi. Ikhadi lokuqinisekisa likhona ukuze lolo lwazi lusetshenziswe ngesikhathi olubaluleke ngaso, yiqembu elinalo.
- Umugqa ongashintshi. "Isihlungi asizange singimise" akukaze kube yisivikelo kunoma yikuphi ukusakaza. Ikhamera ayenqabi, ipensela alenqabi; umthetho nokuziphatha kobungcweti kulawula ukusetshenziswa, hhayi ithuluzi. Isikhiqizi esingahlungiwe sibuyisela ukuhleleka okuvamile lapho onke amanye amathuluzi obungcweti asebenza ngaphansi kwakho.
Funda iNqubomgomo Yokusebenzisa Ngokuzibophezela ngaphambi kokuqala. Imfushane, icacile, futhi iyingxenye yalo mkhiqizo eyenza okunye kwawo kuvikeleke.
09Imibuzo ebuzwa njalo
Kungani isikhiqizi sezithombe se-AI senqaba umyalo osobala ukuthi usemthethweni?
Ngoba izihlungi eziningi zahlulela amagama asemyalweni wakho, hhayi inhloso yakho. Isihlukanisi sihlala phambi kwemodeli singenakho ukufinyelela kulokho oyikho noma kwisizathu sokubuza kwakho, ngakho umdwebi wezokwelapha, umhlaziyi wobufakazi bomthetho nomuntu onezinhloso ezimbi bonke bathayipha ibizo elifanayo bese bethola ukwenqatshwa okufanayo. I-Intercept yabika ngo-Ephreli 2023 ukuthi i-Midjourney yavimba amagama esakhiwo somzimba afaka u-speculum, i-placenta ne-cervix, kanti u-stethoscope wayesebenza ngokujwayelekile.
Ingabe lukhona ucwaningo olukala amaphutha ezihlungi zokuqukethwe kumathuluzi obuciko be-AI?
Yebo. I-OVERT (arXiv:2505.21347, ICML 2025) yisilinganiso sokuqala esikhulu sokwenqaba ngokweqile kumamodeli asuka embhalweni aye esithombeni: imiyalo engu-4,600 ebukeka iyingozi kodwa engenabungozi ngempela ezigabeni eziyisishiyagalolunye zokuphepha, kanye nemiyalo engu-1,785 eyingozi ngempela njengesilawuli. Ababhali bayo babika ukuthi ukwenqaba ngokweqile kudlangile kuwo wonke amamodeli ahamba phambili. Ucwaningo oluhlukile lwe-ECCV 2024 lwathola ukuchema ngobulili nangesitayela kwizihlukanisi ezintathu ze-NSFW ezahlolwa ngobunqunu bobuciko.
Kushintshani ngempela isikhiqizi sezithombe se-AI esingavinjelwe?
Sisusa ungqimba lokwenqaba nesihlungi sokuqukethwe endleleni yokwakha, ngakho isicelo esisemthethweni asihlukaniswa kabusha ngokuthula njengokuhlasela. Akususi ukwahlulela — kuyakuthutha. I-Shannon ikubonisa umyalo wokwakha oncamashi ekhadini lokuqinisekisa ngaphambi kokuba kwakhiwe okuthile, ngakho isinqumo senziwa ngumuntu okwazi ukubona lokho okuzokwakhiwa, ngaphansi kweNqubomgomo Yokusebenzisa Ngokuzibophezela.
Ingabe i-Shannon iyakha ividiyo?
Cha. I-Shannon yakha izithombe ezintsha, ihlele isithombe esisodwa esikhona, futhi ihlanganise izithombe ezimbili noma ngaphezulu zibe sinye. Ayizinyakazisi izithombe futhi ayikhiqizi ividiyo yanoma yiluphi uhlobo.
Ingabe ngingakwakha izithombe nge-API ye-Shannon?
Cha. Ukwakha, ukuhlela nokuhlanganisa izithombe kuhlala ohlelweni lwengxoxo lwe-Shannon. I-API ye-/v1 ikhonza umbhalo nombono — ifunda futhi ihlaziye izithombe ozinikezayo — hhayi ukwakha izithombe.
Uma singekho isihlungi, yini evimba ukusetshenziswa kabi?
Inqubomgomo, ukuphendula nendawo yesinqumo ebonakalayo esikhundleni sohlu lwamagama avinjiwe. I-Shannon ayilokothi yakhe ingacelwanga, ibonisa umyalo ogcwele ngaphambi kokwakha, ayikhokhisi lutho uze uqinisekise, futhi isebenza ngaphansi kweNqubomgomo Yokusebenzisa Ngokuzibophezela eshicilelwe evimbela okuqukethwe kobulili okuhilela abancane, izithombe zangasese ezingavunyelwe zabantu bangempela, ukuhlukumeza okuqondiwe, nokukhwabanisa amadokhumenti obunikazi noma ezezimali.
Cela isithombe osidinga ngempela
Ukwakha, ukuhlela isithombe esisodwa nokuhlanganisa izithombe eziningi ohlelweni lwengxoxo — asikho isihlungi sokuqukethwe, alukho ungqimba lokwenqaba, nomyalo esikrinini ngaphambi kokuba kwakhiwe okuthile.
Qala Ukuxoxa INqubomgomo Yokusebenzisa NgokuzibophezelaIzithombe ohlelweni lwengxoxo · akukho ukwakha ividiyo · olunye ucwaningo
Imithombo: Cheng, Huang, Xu, Sojoudi, Zhao, Song & Mei — OVERT (arXiv:2505.21347, ICML 2025) · Riccio, Curto, Hofmann & Oliver — An Art-centric perspective on AI-based content moderation of nudity (arXiv:2409.17156, AI4VA @ ECCV 2024) · Debbie Nathan, The Intercept, u-Ephreli 22, 2023 · PetaPixel, u-Ephreli 3, 2023 · CNBC, uFebhuwari 22, 2024. I-Shannon AI iqhutshwa yi-Shannon Lab LLC, New Mexico, USA.