Isifundo sikaPian Pian Ge sezingubo zabesifazane: Ukwakha igalari yezingubo ye-AI enezilimi ezimbili nesevisi yokuzizama nge-virtual

Isingeniso sephrojekthi

拍片哥女装 (COSTUME-DB-01) iyigalari eyimithombo evulekile yezithenjwa zemfashini esebenzisa i-AI, enetimu emnyama. Iqukethe izithombe zezimpahla zokugqoka ezikhiqizwe yi-AI ezingu-1407, ezingenabantu ababonakalayo futhi ezinesilinganiso esingu-2:3. Ingaphambili yiwebhusayithi emile esebenza ngaphandle kokwakhiwa, ngakho ungaphequlula, useshe futhi uhlunge ngokuqondile; uma udinga ukuzizamela izingubo, ungaphinde usebenzise i-backend ye-GoFrame v2 etholakala ku-backend/.

Qala ngokuvula inguqulo eku-inthanethi ukuze uzame imisebenzi yegalari yamahhala, noma uthole ikhodi yomthombo ku-repo ye-GitHub ukuze uyiqhube endaweni yangakini futhi uyithuthukise.

Imisebenzi esemqoka

• Amasethi ezimpahla angu-1407: ahlukaniswe abe yizigaba ezingu-8: intwasahlobo, ihlobo, ikwindla, ubusika, izingubo zangaphansi, amaphathi, izingubo zokubhukuda kanye nezingubo ezihehayo.
• Izilimi ezimbili, isiShayina nesiNgisi: isiShayina siboniswa ngokuzenzakalela, kodwa ungashintshela ku-English ekhoneni eliphezulu kwesokudla sekhasi.
• Ukusesha ngamagama angukhiye ngezindlela ezimbili: kusekela amagama esiShayina afana no-“白、蕾丝、连衣裙”, kanye namagama esiNgisi afana ne-white, lace ne-dress.
• Ukuphequlula izithombe: kunokuskrola okungapheli, ukubuka izithombe ezinkulu, ukushintsha ngezinkinobho zemicibisholo yesokunxele nesokudla, ukuvala nge-Esc kanye nokulanda i-PNG ngosayizi wokuqala.
• Indawo yokuhlanganisa yokuzizamela izingubo: ungalayisha isithombe somuntu, ukhethe ingubo bese uyibuka kuqala, uyihudule, uyikhulise noma uyinciphise futhi uyihlanganise ngemva kokususa ingemuva.
• Ukushintsha izingubo nge-AI: nge-backend ozikhethela yona, ungashayela i-Gemini, i-Doubao, i-GPT Image, amamodeli ahambisana ne-OpenAI Images noma i-ComfyUI yendawo.
• Ukugcinwa kobumfihlo endaweni: izilungiselelo zamamodeli zigcinwa esipheqululini samanje; amarekhodi akhiqizwe ngempumelelo agcinwa ku-IndexedDB futhi awavumelaniswa phakathi kwamadivayisi.
• Ukusesha imikhiqizo efanayo ngesithombe: imininingwane yesithombe esikhulu nendawo yokuzizamela izingubo kungavula ukusesha kwezithombe kwe-Taobao.

Isinyathelo sokuqala: sebenzisa i-frontend emile

I-frontend yakhiwe kakhulu yi-index.html, i-data.js kanye ne-assets/; akudingeki ufake izixhomekeki noma wenze i-build. Ngemva kokuthola i-repo, qala noma iyiphi iseva yamafayela emile empandeni yephrojekthi.

Ukusebenzisa i-Python

python3 -m http.server 3000

Ukusebenzisa i-Node.js

npx serve -l 3000 .

Bese uvakashela ku-http://localhost:3000. Ungakwazi futhi ukuvula i-index.html ngokuqondile esipheqululini sanamuhla, kodwa ukusebenzisa iseva ye-HTTP yendawo kuvame ukufaneleka kakhulu ekuhloleni ukusebenzisana okuphelele.

Ukuphequlula igalari, ukuhlunga ngezigaba, ukusesha ngezilimi ezimbili, ukubuka izithombe ezinkulu kanye nendawo eyisisekelo yokuhlanganisa akudingi i-API Key futhi akudingi i-backend. Isamba sama-PNG ngosayizi wokuqala singaphezu kuka-1GB, ngakho-ke asifakwanga ku-repo; inkinobho yokulanda izokhomba ezinsizeni zesayithi lokuqala.

Isinyathelo sesibili: phequlula futhi useshe izingubo

1. Khetha isigaba: nciphisa ububanzi bokuphequlula ngokukhetha izigaba ezifana nentwasahlobo, ihlobo, ikwindla nobusika.
2. Faka igama elingukhiye: faka izici zengubo ngesiShayina noma ngesiNgisi ebhokisini lokusesha; imiphumela izobuyekezwa ngesikhathi sangempela.
3. Shintsha ulimi: sebenzisa inkinobho yolimi ekhoneni eliphezulu kwesokudla ukuze ushintshe phakathi kwesixhumi esibonakalayo sesiShayina nesiNgisi.
4. Buka isithombe esikhulu: ngemva kokuvula imininingwane yengubo, sebenzisa imicibisholo yesokunxele nesokudla yekhibhodi ukuze uphequlule izithombe eziseduze, bese ucindezela u-Esc ukuze uvale.
5. Landa noma uthole umkhiqizo ofanayo: sebenzisa isixhumanisi sokulanda ngosayizi wokuqala noma uvule ukusesha kwezithombe kwe-Taobao.

Isayithi futhi lisekela ukwabelana ngolimi oluthile nesimo sokusesha kusetshenziswa amapharamitha e-URL:

?lang=en
?q=lace
?lang=en&q=white

lang ingaba ngu-zh noma u-en, kanti i-q ichaza igama lokuqala lokusesha.

Isinyathelo sesithathu: sebenzisa indawo eyisisekelo yokuzizamela izingubo

Ngemva kokuvula umsebenzi wokuzizamela izingubo ku-frontend, khetha ingubo eyisithenjwa egalari bese ulayisha isithombe somuntu. Indawo eyisisekelo yokuhlanganisa isekela ukubuka kuqala ku-inthanethi, ukuhudula, ukukhulisa nokunciphisa kanye nokuhlanganisa ngemva kokususa ingemuva; ayibizi imodeli enkulu. Uma ufuna ingubo ihambisane ngokwemvelo nesimo somzimba, kufanele uqhubeke ngokufaka i-backend ye-AI yokushintsha izingubo.

Lapho ulayisha isithombe, i-frontend izonciphisa ngokuzenzakalela uhlangothi olude lube ngamaphikseli angu-2048, bese izama kuqala ukusiguqulela ku-WebP futhi isicindezele ngezinyathelo zekhwalithi ezisuka ku-0.90 ziye ku-0.75.

Isinyathelo sesine: sebenzisa i-Docker ukuze ufake i-backend ye-AI yokushintsha izingubo

I-backend itholakala ku-backend/. Kunconywa ukusebenzisa i-Docker Compose ukuze ungafaki indawo yokwakha i-Go emshinini osingathayo.

cp backend/.env.example backend/.env
docker compose up -d

Hlela i-backend/.env bese ufaka okhiye bamamodeli ohlela ukuwasebenzisa. Ngemva kokuqala kwesevisi, izolalela ngokuzenzakalela ku-http://localhost:8000. Uma i-frontend nayo ivulwe nge-localhost noma nge-127.0.0.1, izozama ngokuzenzakalela ukuxhuma ku-http://127.0.0.1:8000.

Ukusebenzisa i-Docker Run

cd backend
docker build -f manifest/docker/Dockerfile -t costume-tryon .
docker run -d --name costume-tryon -p 8000:8000 -e ARK_API_KEY="I-ArkKey yakho ye-Doubao" -e OPENAI_API_KEY="sk-OpenAIKey yakho" costume-tryon

Ngaphandle kokusebenzisa i-Docker

Abasebenzisi be-Windows bangasebenzisa i-backend/restart.bat. Abasebenzisi asebefake i-Go bangayiqala ngokuqondile:

cd backend
export ARK_API_KEY="Ukhiye wakho"
go run main.go

Kwi-Windows PowerShell, indlela yokubhala okuguquguqukayo kwemvelo ngu-$env:ARK_API_KEY="Ukhiye wakho".

Isinyathelo sesihlanu: lungisa amamodeli okushintsha izingubo

Ulayini wokulungiselela amamodeli utholakala ku-backend/manifest/config/config.yaml, kanti isifanekiso ngu-config.example.yaml. Imodeli ngayinye ingachaza ngokwehlukana iphrothokholi, i-API Key, igama lokuguquguquka kwemvelo, i-Base URL kanye negama langempela lemodeli.

llm:
  models:
    gpt-image:
      provider: openai
      envKey: "OPENAI_API_KEY"
      baseURL: "https://api.openai.com/v1"
      model: "gpt-image-2"

    doubao:
      provider: doubao
      envKey: "ARK_API_KEY"
      baseURL: "https://ark.cn-beijing.volces.com/api/v3"
      model: "doubao-seedream-3-0-t2i-250415"

    gemini:
      provider: gemini
      envKey: "GEMINI_API_KEY"
      baseURL: "https://generativelanguage.googleapis.com/v1beta"
      model: "gemini-3.1-flash-image"

Okuguquguqukayo kwemvelo kunokubaluleka okuphezulu kune-apiKey esekucushweni. Uma ufuna ukuxhuma kumasevisi afana ne-Tongyi Wanxiang, i-Zhipu CogView noma i-SiliconFlow ahambisana nephrothokholi ye-OpenAI Images, engeza ingxenye yemodeli ene-provider: openai bese ugcwalisa i-Base URL negama lemodeli elihambisanayo; akudingeki ushintshe ikhodi yokuhlanganisa.

Isinyathelo sesithupha: xhuma i-frontend ku-backend

I-frontend izosesha isevisi ye-AI yokushintsha izingubo ngokwalokhu okulandelayo:

1. “Ikheli le-API ye-backend” eligcinwe esikhungweni sezilungiselelo.
2. Isiguquguquki somhlaba wonke window.AI_API_URL esisethwe ngaphambi kokulayishwa kwekhasi.
3. Ipharamitha ye-tryon eku-URL.
4. Lapho uvakashela i-frontend endaweni yangakini, ibuyela ku-http://127.0.0.1:8000.

Indlela elula ukuvula isikhungo sezilungiselelo ekhoneni eliphezulu kwesokudla sekhasi bese ufaka http://localhost:8000. Ungasebenzisa futhi leli kheli lokwabelana:

https://Ikheli_le-frontend_yakho/?tryon=http://localhost:8000

Isikhungo sezilungiselelo sivumela futhi ukugcina iphrothokholi, igama lemodeli, i-Base URL, i-API Key kanye ne-Token ezimele yokufinyelela i-backend. Ukulungiselelwa kwemodeli yomuntu kugcinwa kuphela esipheqululini samanje, futhi kudluliselwa ku-backend okwesikhashana ngesicelo esisodwa ngesikhathi sokukhiqiza; akubhalwa kusizindalwazi se-backend, kumafayela okulungiselela noma kumalogi ebhizinisi. Ngemva kokusebenzisa idivayisi eyabiwe, kufanele ususe ngokwakho ukulungiselelwa kwasendaweni.

Ukushintsha izingubo ungaxhunyiwe ku-inthanethi nge-ComfyUI yendawo

Uma ikhompuyutha inekhadi lemifanekiso elizimele, noma kunomshini we-ComfyUI kunethiwekhi yendawo, ungasebenzisa indlela yendawo engadingi i-API Key. Iphrojekthi isekela i-Qwen-Image-Edit-2511 Int8, i-Lightning LoRA yezinyathelo ezingu-4 kanye ne-workflow ye-MiniMax H3, futhi ingabuyela ku-H3 lapho i-workflow ekhethwayo ingakalungeli.

1. Qala i-ComfyUI; ikheli elizenzakalelayo ngu-http://127.0.0.1:8188.
2. Landela i-backend/README.MD ukuze ubeke imodeli ye-Qwen-Image-Edit-2511 ne-Lightning LoRA.
3. Lungisa ingxenye yemodeli ye-ComfyUI namafayela e-workflow.
4. Esikhungweni sezilungiselelo sekhasi lewebhu, khetha “ComfyUI yendawo”, uhlole ukuthi uxhumano luphumelele bese ulugcina.

comfyui:
  provider: comfyui
  baseURL: "http://127.0.0.1:8188"
  model: "qwen-image-edit-2511"
  workflowPath: "workflows/comfyui_qwen_image_edit_2511_api.json"

Amaqhinga athuthukile nezincomo zokufakwa ekukhiqizeni

Khetha uhlelo lwemodeli olufanele

• ComfyUI: ifanele abasebenzisi abanekhadi lemifanekiso lendawo, abazisa ubumfihlo futhi abafuna ukugwema izindleko ze-API.
• Gemini: i-README ibeka i-Nano Banana 2, okungukuthi i-gemini-3.1-flash-image, njengenketho ezenzakalelayo enconyiwe yamafu.
• Doubao Seedream: ifanele abasebenzisi abadinga inketho esebenza kumanethiwekhi ase-China.
• GPT Image 2: isebenzisa iphrothokholi ye-OpenAI Images, futhi ingalungiselelwa nekheli lokudlulisa lenkampani yangaphandle elihambisanayo.

Nika amandla imikhawulo yokuphepha yokufinyelela esidlangalaleni

Lapho ushicilela i-backend ku-inthanethi yomphakathi, akufanele uveze isevisi ezenzakalelayo kuphela. Iphrojekthi inikeza ukuqinisekiswa kwe-Bearer Token, uhlu olumhlophe olunembile lwe-CORS, ukulinganiselwa kwe-Token Bucket, izabelo zansuku zonke kanye nokulawula ukuvumelana komhlaba wonke nokwesihloko ngasinye. Amanani azenzakalelayo abhalwe ku-README ngu-10 ngomzuzu, 100 ngosuku, ukuvumelana komhlaba wonke okungu-2 kanye nokuvumelana kwesihloko esisodwa okungu-1; bheka i-backend/DEPLOY.md ukuze uthole ukulungiselelwa okuqondile kokukhiqiza.

I-backend yamukela kuphela okuqukethwe kwesithombe sangempela kwe-JPEG, PNG ne-WebP. Usayizi omkhulu wesithombe esisodwa ngu-8 MiB, kanti izithombe ezimbili zihlangene azidluli ku-16 MiB; ubude nobubanzi besithombe kumele kube phakathi kwamaphikseli angu-256 no-8192, futhi inani lamaphikseli selilonke alidluli ezigidini ezingu-20. Le mikhawulo isiza ukunciphisa ubungozi bamafayela omgunyathi, ukufakwa kwekhodi eyingozi kanye namabhomu amaphikseli avela ekucindez/ekuvulweni kwamafayela.

Ukuqonda idatha yendawo nemingcele yobumfihlo

Imiphumela ephumelele ye-AI yokushintsha izingubo igcinwa ku-IndexedDB yesiphequluli samanje, futhi ingabukwa, ilandwe, isuswe ngayinye noma isulwe yonke. Amarekhodi awalayishwa kuseva futhi awavumelaniswa ngokuzenzakalelayo namanye amadivayisi. I-Base URL yomuntu siqu ivinjelwe ngokuzenzakalelayo ukuthi iqondise ku-localhost, kunethiwekhi yangaphakathi noma emakhelini agodliwe, ukuze kuvinjelwe i-backend esesidlangalaleni ukuthi isetshenziswe njengommeleli wezicelo zangaphakathi kwenethiwekhi.

Ukwandisa idatha yezingubo

Idatha yegalari igcinwe ku-data.js. Irekhodi ngalinye liqukethe id, カテゴリ, キーワード, サムネ, 画像 kanye ne-原寸. Lawa magama ezinkambu esiJapane agcinwe ukuze ahambisane nedatha yokuqala; i-frontend izokwenza ngokuzenzakalelayo ukuhumusha kwezigaba namagama angukhiye kube isiShayina noma isiNgisi. Lapho unweba idatha, qhubeka ulandela isakhiwo sezinkambu esikhona kanye nemithetho yezindlela zezithombe.

Isiphetho

Izingubo zabesifazane zika拍片哥 zingasetshenziswa njengegalari yezinkomba zezingubo ze-AI enezilimi ezimbili, esebenza ngokushesha ngaphandle kokulungiselelwa okwengeziwe, noma njengesisekelo se-frontend ne-backend sephrojekthi yokuzama izingubo ngokubona kusetshenziswa amamodeli amaningi. Ekusetshenzisweni kokuqala, kudingeka kuphela kuqaliswe i-frontend emile; uma kudingeka ukushintshwa kwezingubo kwangempela nge-AI, khetha imodeli yefu ye-Docker noma isixazululo sendawo se-ComfyUI ngokuya ngesidingo. Iphrojekthi isebenzisa i-MIT License, evumela ukusetshenziswa kwezohwebo, ukuguqulwa, ukuhlanganiswa nokusabalalisa, kodwa kufanele kugcinwe izitatimende ezidingekayo zelayisensi, futhi uqaphele ukuthi zonke izithombe zegalari ziyimiklamo yezingubo zomqondo ekhiqizwe yi-AI, hhayi izithombe zemikhiqizo yangempela yemikhiqizo edumile.