Yeqa uye kokuyinhloko
Izifundo ze-AI

Isifundo se-Lanzi: Dala umlingiswa ongaguquki usebenzisa izithombe eziyizinkomba zokuma nezingubo, bese uyixhuma ku-DeepSeek Harness

Lesi sifundo sichaza ukuthi uyifaka futhi uyisebenzisa kanjani i-Lanzi, ukhethe izithombe eziyizinkomba phakathi kwemidwebo yokuma engu-1439 nezingubo ezingu-1407, ulungise i-Seedream, i-Gemini noma isixhumi esihambisana ne-OpenAI ukuze ukhiqize izithombe, bese uxhuma isitudiyo esigcwele namathuluzi okukhiqiza ku-DeepSeek Harness.

Isifundo se-Lanzi sokudala umlingiswa ongaguquki ngezithombe zokuma nezingubo nokuxhuma ku-DeepSeek Harness

Iyini i-Lanzi

I-Lanzi iyisitudiyo sokuma esisebenza sodwa kanye ne-plugin ye-DeepSeek Harness. Ikhetha ukuma okukodwa noma eziningi emiqoqweni engu-288 enemidwebo yokuma engu-1439 eqoqwe ku-ai-pose-gallery-zh, bese ihlanganisa izithombe eziyizinkomba zomlingiswa nezingubo ukuze ikhiqize izithombe zemodeli efanayo ezenza izenzo ezahlukene.

Ngokuzenzakalela, iphrojekthi isebenzisa umlingiswa omdala we-Whale Girl womphakathi, Da Feiyu, njengemodeli. Unezinwele ezinde eziluhlaza ezigoqene, izindlebe zomkhomo nomsila womkhomo oluhlaza, futhi ugqoka iyunifomu yencekukazi eluhlaza okwesibhakabhaka nemhlophe. Ungangenisa futhi izithombe ezifika kwezintathu zamanye amamodeli ukuze ushintshe ubuso nezingubo zomlingiswa.

Imidwebo yokuma isetshenziswa kuphela ukuchaza ukuma komzimba; ngokuzenzakalela, umzimba ongenalutho osemidwebeni ngeke udwetshwe esithombeni sokugcina. Imidwebo ayigcinwa endaweni yokugcina ikhodi yephrojekthi, kodwa ifundwa lapho idingeka ivela kusevisi yomphakathi yokugcina izithombe.

Imisebenzi esemqoka

  • Iqoqo elibanzi lokuma: Linikeza amaqoqo angu-288 nemidwebo yokuma engu-1439, ngokukhetha okukodwa noma okuningi.
  • Ukushintshwa kwezingubo: Khetha ingubo eyodwa kwezingu-1407 eziku-COSTUME-DB-01, kuyilapho kugcinwa ubuso, isitayela sezinwele, izindlebe nomsila womkhomo womlingiswa.
  • Imodeli eyenziwe ngokwezifiso: Ngenisa izithombe zamamodeli ezifika kwezintathu, noma uqhubeke usebenzisa uDa Feiyu ozenzakalelayo.
  • Izixhumi eziningi zokukhiqiza izithombe: Isekela i-Volcano Ark Seedream, i-Gemini kanye nezixhumi ezihambisana ne-OpenAI.
  • Ukuhamba komsebenzi ngaphandle kokhiye: Ngisho noma ungakalungisi ukhiye we-API, kusakhiwa umyalo ophelele ongawukopisha uwusebenzise ku-dsh-image-gen.
  • Ukuhlanganiswa ne-Harness: Ngemva kokufaka i-plugin, ungavula isitudiyo esigcwele kubha eseceleni, noma uvumele i-Agent ukuthi iseshe ukuma, izingubo futhi ikhiqize izithombe.
  • Isixhumi esibonakalayo esiphendulayo: Izikrini ezibanzi zisebenzisa amakholomu amaningi, kanti izikrini ezincane zishintshela kukholomu eyodwa futhi zigcine inkinobho yokukhiqiza ezansi.

Ukuqalisa isitudiyo sendawo

1. Lungisa indawo yokusebenza

I-Lanzi idinga i-Node.js 22.19 noma inguqulo entsha. Ngemva kokuqinisekisa inguqulo ye-Node.js, faka okuncikile bese uqala iseva yokuthuthukisa enkombeni yephrojekthi:

npm install
npm run dev

Ngemva kokuqala, vula http://127.0.0.1:43123 esipheqululini.

2. Khetha ukuma

Khetha ukuma endaweni emaphakathi yesitudiyo. Ungakhetha ukuma okuningi ngesikhathi esisodwa ukuze uhlelo lukhiqize izithombe ezihambisanayo ngokwehlukana. Ukuma komzimba ogcwele okujwayelekile kuboniswa ngokuzenzakalela; ukuma okuvusa inkanuko nokuma kwabantu ababili akuboniswa kuze kuvulwe ngokucacile iswishi Yabantu abadala.

3. Khetha izingubo

Ngemva kokushintshela ekhasini Izingubo, ungakhetha ingubo eyodwa. Izigaba zasentwasahlobo, zasehlobo, zasekwindla nezasebusika, kanye nezephathi, zibonakala ngokuzenzakalela; izigaba zezingubo zangaphansi, ezokubhukuda nezivusa inkanuko zisebenzisa iswishi efanayo nezokuma zabantu abadala.

Ngemva kokukhetha ingubo, umphumela wokukhiqiza uzosebenzisa leyo ngubo futhi uzame ukugcina ubuso, isitayela sezinwele nezinye izici zomlingiswa. Ngalesi sikhathi, inketho ezenzakalelayo ethi Gcina izingubo izobekelwa eceleni okwesikhashana yingubo ekhethiwe.

Isitudiyo se-Lanzi esinokuma nezingubo ezikhethiwe nesithombe esikhiqiziwe

4. Lungisa noma shintsha imodeli

Indawo engakwesokunxele isetshenziselwa ukuphatha imodeli. Ungaqhubeka usebenzisa uDa Feiyu ozenzakalelayo, noma ungenise izithombe zamamodeli akho ezifika kwezintathu. Izithombe eziningi eziyizinkomba zisiza ukunikeza ulwazi ngobuso, izinwele nezingubo zomlingiswa, kodwa umphumela wangempela usancike kumodeli yokukhiqiza izithombe esetshenzisiwe.

5. Lungisa isixhumi sokukhiqiza izithombe

Vula Izilungiselelo zokukhiqiza izithombe ekhoneni eliphezulu kwesokudla, bese ugcwalisa okhiye abadingwa yi-Gemini noma isixhumi esihambisana ne-OpenAI. Ukhiye ugcinwa kuphela kulesi siphequluli futhi awubhalwa endaweni yokugcina njengedatha yephrojekthi.

I-Lanzi isekela lezi zilungiselelo ezizenzakalelayo:

  • Seedream: Isebenzisa ikheli le-Volcano Ark https://ark.cn-beijing.volces.com/api/v3, imodeli ezenzakalelayo ingu-doubao-seedream-5-0-flash-260915.
  • Gemini: Imodeli ezenzakalelayo ingu-gemini-2.5-flash-image.
  • Isixhumi esihambisana ne-OpenAI: Kuqala sibiza images/edits; uma isixhumi singakusekeli ukuhlela, sibuyela ekukhiqizeni ngombhalo kuphela futhi sibonise emiphumeleni ukuthi izithombe eziyizinkomba azithunyelwanga.

I-Seedream ithumela ndawonye isithombe semodeli, isithombe sezingubo esingakhethwa kanye nomdwebo wokuma. Uma kukhethwe izingubo, ukuhleleka kwezithombe eziyizinkomba kuba umlingiswa, isithombe sezingubo esingenamuntu, umdwebo wokuma.

6. Qala ukukhiqiza

Ngemva kokuqeda ukulungisa imodeli, ukuma, izingubo nesixhumi, chofoza inkinobho yokukhiqiza. Ngisho ngaphandle kokhiye olungisiwe, indawo engakwesokudla isazokhiqiza umyalo wokuma ngakunye okukhethiwe. Ungayikopisha le miyalo uyinike i-dsh-image-gen efakiwe ukuze ikhiqize.

Ukulungisa okhiye ngasohlangothini lweseva

Uma ungafuni ukufaka okhiye esipheqululini, ungasebenzisa iziguquguquko zemvelo ukulungisela inqubo yeseva umhlinzeki, imodeli, ikheli lesixhumi nokhiye:

POSE_MUSE_PROVIDER=seedream \
POSE_MUSE_MODEL=doubao-seedream-5-0-flash-260915 \
POSE_MUSE_BASE_URL=https://ark.cn-beijing.volces.com/api/v3 \
POSE_MUSE_API_KEY=ukhiye wakho \
npm run dev

Le ndlela ifanele ukuthumela isitudiyo emshinini ongashintshi noma endaweni elawulwayo. Ngemva kokuqala, ingaphambili lingasebenzisa ngokuqondile izilungiselelo zokukhiqiza ezihlinzekwa yiseva, ngaphandle kokugcina ukhiye esipheqululini.

Ukufaka ku-DeepSeek Harness

Uma i-DeepSeek Harness isivele ifakiwe emshinini, ungangeza i-Lanzi njenge-plugin ku-web profile:

dsh plugin --profile web add gen-ai-pose-image

Igama lephakheji le-plugin ngu-gen-ai-pose-image. Uma iphakheji ingakakhishwa, noma udinga ukuhlola izinguquko zasendaweni, shintsha igama lephakheji ulibeke endleleni yendawo yokugcina ikhodi. Ngemva kokuqeda ukufaka, qala kabusha i-profile ehambisanayo.

Isitudiyo se-Lanzi ngaphakathi kwe-DeepSeek Harness

Ngemva kokuqalisa kabusha, ibha eseceleni engakwesokunxele izoba nokungena kwe-Lanzi / Isitudiyo, okuhambisana ne-Plugins kanye Nemisebenzi ezenzakalelayo. Uma uyivula, ungakhetha imodeli, ukuma nezingubo ekhasini elifanayo, ulungise izilungiselelo zokukhiqiza futhi ubuke imiphumela.

Ukuvumela i-Agent ibize amathuluzi e-Lanzi

I-plugin inikeza i-Agent amathuluzi amathathu:

  • search_poses: Isesha ukuma okuhambisana nezimo ezithile.
  • search_costumes: Isesha izingubo.
  • generate_pose_images: Ikhiqiza izithombe zokuma ngokusekelwe kokuqukethwe okukhethiwe.

Uma udinga ukushintsha izingubo, dlulisela inombolo yezingubo ku-costume_id. Uma inqubo ye-Harness ine-POSE_MUSE_API_KEY elungisiwe, ithuluzi lizokhiqiza izithombe ngokuqondile lizifake eseshini; ngaphandle kokhiye, ithuluzi lizobuyisa imiyalo elungisiwe ukuze iqhubeke inikezwe ku-generate_image noma edit_image ye-dsh-image-gen.

Ukuma nezingubo zabantu abadala akubonakali emiphumeleni yosesho ngokuzenzakalela. Kufanele udlulisele kuphela include_adult: true ku-search_poses noma ku-search_costumes lapho umsebenzisi ekucela ngokucacile.

Ukusebenzisa kumadivayisi eselula namaphaneli amancane

I-Lanzi isekela ukwakheka okuphendulayo. Esikrinini esincane, indawo yemodeli iyagoqeka iye phezulu, izigaba zokuma zishintshela ekukhetheni okwehlayo, kanti inkinobho yokukhiqiza igcinwa iqinile ezansi kwekhasi. Izigaba zezingubo zizisonga ngokuzenzakalela zibe imigqa eminingi, futhi izincazelo zamakhadi zihlala endaweni ebonakalayo.

Ikhasi lokukhetha izingubo le-Lanzi ebubanzini beselula

Kwiphaneli encane ye-Harness, isitudiyo ebesinamakholomu amaningi naso siyashintshela kukholomu eyodwa, ngakho akudingeki ushintshele enguqulweni ehlukile yeselula.

Iziphakamiso zokusebenzisa ezithuthukile

Hlela kahle ukukhiqizwa kokuma okuningi

I-Lanzi ivumela ukukhetha ukuma okuningi ngesikhathi esisodwa. Uma udinga ukuqhathanisa izenzo, qala ngokugcina imodeli nezingubo kungashintshi, bese ukhetha iqoqo lokuma ukuze ukhiqize ngobuningi. Lokhu kwenza kube lula ukubona ukungaguquguquki komlingiswa ezimweni ezahlukene.

Qonda umsebenzi wezithombe eziyizinkomba zezingubo

Izithombe ezikwiqoqo lezingubo azinabantu futhi zisetshenziswa kakhulu ukunikeza ukubukeka kwengubo. Ngemva kokukhetha izingubo, i-Lanzi ihlanganisa izinkomba zomlingiswa, izinkomba zezingubo nomdwebo wokuma. Uma isixhumi singakusekeli ukuhlela izithombe bese sibuyela ekukhiqizeni ngombhalo kuphela, izithombe eziyizinkomba ngeke zithunyelwe, futhi ikhasi lemiphumela lizokuchaza ngokucacile lokhu.

Qinisekisa ukuhamba komsebenzi ngemiyalo kuqala

Uma okwamanje ungenawo ukhiye we-API osebenzayo, qala ngokukhetha ukuma, imodeli nezingubo, uhlole imiyalo ekhiqizwe ngakwesokudla, bese uyikopisha uyisebenzise kwamanye amathuluzi okukhiqiza izithombe. Lokhu kukuvumela uqinisekise ukuhamba komsebenzi ngaphambi kokulungisa isixhumi esisemthethweni.

Yenza ukuhlolwa kwephrojekthi

Ngaphambi kokushintsha ikhodi yendawo noma ukulungiselela ukuthumela, ungasebenzisa izivivinyo nokuhlolwa kwezinhlobo:

npm test
npm run typecheck

Amalungelo okukopisha nemikhawulo yokusebenzisa

Uhlelo lwe-Lanzi uqobo lusebenzisa i-MIT License, kodwa le layisensi ihlanganisa ikhodi yohlelo kuphela futhi ayizimbozi ngokuzenzakalela izithombe eziyizinkomba zokuma, izingubo nomlingiswa.

  • Amalungelo okukopisha emidwebeni yokuma nasemininingwaneni yokuqala angaka-ギア; incazelo yomthombo ithi lezi zithombe azivunyelwe ukusetshenziswa kwezohwebo.
  • Izithombe zezingubo zivela ku-衣装集 COSTUME-DB-01 ka-Gia; indawo yokugcina ikhodi igcina kuphela izinombolo namagama angukhiye.
  • Izithombe zokumisa uDa Feiyu zivela ku-deepseek-whale-girl, umbhali ngu-TreapGoGo, futhi zisetshenziswa ngaphansi kwe-MIT License yendawo yokugcina ehambisanayo.
  • UDa Feiyu ungumlingiswa womphakathi futhi akamele i-DeepSeek; i-DeepSeek iwuphawu lokuhweba lwe-DeepSeek, futhi ilayisensi yephrojekthi ayinikezi amalungelo alolo phawu.

Ngaphambi kokushicilela noma ukusebenzisa imiphumela ekhiqiziwe, hlola shared/license.mjs kanye ne-LICENSE kuphrojekthi, futhi ulandele amalayisensi azimele nemikhawulo yokusebenzisa yazo zonke izinto.

Isifinyezo

I-Lanzi ihlanganisa izinkomba zomlingiswa, imidwebo yokuma, izithombe zezingubo ezingenabantu nezixhumi zokukhiqiza izithombe kube yisitudiyo esisodwa. Ungakhetha ngokushesha ukuma nezingubo endaweni yangakini, noma ufake i-plugin ku-DeepSeek Harness ukuze i-Agent iseshe izinto futhi ikhiqize izithombe. Ukuze uqale, sebenzisa isitudiyo sendawo uqinisekise ukuhamba komsebenzi, bese ulungisa okhiye esipheqululini noma ohlangothini lweseva ngokwesixhumi osisebenzisayo.