FLUX 3 Image Klein Uyasebenza

I-FLUX 3 Image Klein iyimodeli yokukhiqiza izithombe ye-AI enomthombo ovulekile ekhishwe yi-Black Forest Labs, enamapharamitha angu-5.6B kuphela futhi ekwazi ukusebenza ngekhadi elilodwa lezithombe elisetshenziswa ngabathengi. I-FLUX 3 Image Klein isekela ukukhiqiza izithombe kusuka embhalweni, ukuhlela izithombe ngemiyalelo, nokukhiqiza izithombe eziningi kusetshenziswa izithombe eziyizethenjwa ezingafika kwezi-10. Iphinde inikeze ngokukhipha okungokwendabuko kwe-4K, ukuhlela okunembile ezingeni lephikseli kanye nokulawula ukwakheka kusetshenziswa amabhokisi ezixhumanisi. Ezivivinyweni zangaphakathi eziyimpumputhe ezisemthethweni, ibe sendaweni yokuqala kuzo zontathu izigaba: ukukhiqiza izithombe zesiShayina, ukuhlela kanye nezithenjwa eziningi, yadlula i-Qwen-Image 2.1 ne-Krea 2.

Ukukhiqiza kusuka embhalweni: Sekela ukukhiqizwa kwezithombe ezisezingeni eliphezulu ngokulungiswa okungokwendabuko okufika ku-4K.
Ukuhlela ngemiyalelo: Sebenzisa ulimi lwemvelo ukushintsha noma iyiphi ingxenye yesithombe, kuyilapho amaphikseli ezindawo ezingashiwongo ehlala engashintshiwe.
Ukukhiqiza ngezithombe eziningi eziyizethenjwa: Yamukela izithombe eziyizethenjwa ezingafika kwezi-10 ngesikhathi esisodwa, igcine ukufana komlingiswa, kwesitayela noma komkhiqizo.
Ukulawula ukwakheka ngamabhokisi ezixhumanisi: Sebenzisa amabhokisi ezixhumanisi zegridi engu-0–1000 ukuze ucacise ngokunembile indawo yezinto noma unciphise indawo ezohlelwa.
Isakhiwo esinesisindo esisodwa: Imisebenzi yokukhiqiza, ukuhlela kanye nezithenjwa eziningi isebenzisa isisindo esifanayo samapharamitha angu-5.6B, ngaphandle kokushintsha amamodeli.
Ukusebenza kuhadiwe yabathengi: Inguqulo ye-INT8 quantized ingasebenza ekhadini elinememori yevidiyo engu-8GB, kanti inguqulo distilled ikhiqiza isithombe ngezinyathelo ezi-4.
Imvelo yokwenza fine-tuning: Izisindo ezivulekile zifaka ukusekelwa kokuqeqeshwa kwe-LoRA, okuvumela ukwenziwa ngokwezifiso kwesitayela, komkhiqizo noma kokugeleza komsebenzi.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI anomthombo ovulekile
Indlela yokukhiqiza ye-Rectified Flow: Ngokungafani namamodeli okusabalalisa endabuko afunda indlela engahleliwe negobile esuka kumsindo iye esithombeni, umndeni we-FLUX ufunda indlela ecishe iqonde exhumanisa umsindo nedatha. Isampula iqedwa ngokuxazulula i-equation evamile yokuhlukanisa, hhayi i-equation engahleliwe yokuhlukanisa. Njengoba indlela iqonda kakhulu, kudingeka izinyathelo ezimbalwa zokuhlanganisa; yingakho inguqulo distilled ingathola imiphumela esezingeni eliphezulu ngesampula yezinyathelo ezingu-16 noma ezisezingeni lezinyathelo ezi-4.
Umzimba we-MMDiT Transformer onemifudlana emibili: Imodeli isekelwe ku-multimodal diffusion Transformer. Ama-token ezithombe nama-token ombhalo acutshungulwa ndawonye kumabhulokhi okuqala anemifudlana emibili, ashintshane ulwazi ngokunaka okwabelwana ngakho, bese ecwengwa ndawonye kumabhulokhi anomfudlana owodwa. Lokhu kwenza imiyalelo yombhalo ihambisane ngokujulile nokukhiqizwa okubukwayo kuzo zonke izingqimba, okuyisisekelo sesakhiwo sokulandela iziyalezo eziyinkimbinkimbi nokunikeza umbhalo.
Isakhiwo esizimele somfaki khodi wombhalo: Amapharamitha angu-5.6B abhekisela kuphela kumzimba oyinhloko wesithombe; ukuqonda umbhalo kwenziwa ngumfaki khodi wangaphandle we-Qwen3-VL-4B-Instruct. Ukwehlukaniswa kwalezi zingxenye kuvumela ikhono lokuqonda incazelo nomzimba oyinhloko wokukhiqiza ukuthi kuthuthukiswe ngokuzimela, okusho ukuthi ukusetshenziswa kwememori yevidiyo ekusetshenzisweni kwangempela kufanele kufake nezindleko zomfaki khodi.
Ukuqeqeshwa okuhlangene kwezindlela eziningi nokudluliswa kolwazi: I-Klein izuza ngokuqondile emiphumeleni yokuqeqeshwa kwangaphambili yemodeli enkulu ye-FLUX 3, eqeqeshwe ngokuhlanganyela ezindleleni ezintathu: izithombe, ividiyo nomsindo. Indlela yokusebenzisa izindlela ezahlukene ukuze zibe yizivimbelo komunye nomunye yenza imodeli encane ithole ikhono lokuqonda okubukwayo elidlula ubukhulu bamapharamitha ayo.

Qinisekisa ihadiwe: Inguqulo distilled ye-BF16 idinga imemori yevidiyo engu-11.2GB; amakhadi angu-8GB kufanele akhethe inguqulo ye-INT8/FP8 quantized, enosayizi ongu-5.6GB kuphela ku-FP8.
Landa izisindo: Donsa izisindo ezinembayo ezidingekayo endaweni yokugcina ye-black-forest-labs/FLUX.3-Klein ku-Hugging Face, bese ulanda nomfaki khodi wombhalo ohambisanayo we-Qwen3-VL-4B.
Yamukela ilayisensi: Chofoza “Agree and access repository” endaweni yokugcina ukuze wamukele i-FLUX Kommunity License ngaphambi kokulanda; ngaphambi kokusetshenziswa kwezentengiselwano, hlola ukuthi imali engenayo yonyaka ingaphansi yini komkhawulo wama-dollar aseMelika angu-500,000.
Khetha indlela yokusetshenziswa: Abasebenzisi be-ComfyUI bangafaka i-plugin esemthethweni ye-BFL ku-Manager bese bekhetha imodeli ye-Klein kugrafu yezinodi, noma basebenzise i-Diffusers ngokuyilayisha ngemigqa embalwa yekhodi ye-Python.
Bhala imiyalelo: Ekukhiqizeni okujwayelekile kusuka embhalweni, bhala incazelo ngokuqondile; ekwakhekeni okuyinkimbinkimbi, kunconywa ukunika amandla i-Prompt Upsampler esemthethweni ukuze ithuthukise imiyalelo ngokuzenzakalelayo.
Sebenzisa ukwakheka ngamabhokisi ezixhumanisi: Emiyalweni, maka izinto ezifana nombhalo nomzimba ngamabhokisi ezixhumanisi athi [y_min, x_min, y_max, x_max]; imodeli izobeka izinto ngokunembile ngokwegridi.
Ukuhlela izithombe: Layisha isithombe sokuqala bese uchaza izinguquko ngolimi lwemvelo, njengokuthi «Yenza lo mgwaqo ube indawo eneqhwa»; amaphikseli ezindawo ezingashiwongo azohlala engashintshiwe.
Ukukhiqiza ngezithombe eziningi eziyizethenjwa: Ku-API noma ku-UI, layisha izithombe eziyizethenjwa ezingafika kwezi-10 kanye nombhalo ochazayo; imodeli izokhiqiza isithombe esisha igcine ukufana kwezinto eziyizethenjwa.
Lungisa imingcele bese ukhiqiza: Inguqulo distilled ayidingi i-CFG futhi ingasetshenziswa ngokuqondile; enguqulweni ye-Base kunconywa i-CFG 3.5 nezinyathelo ezingu-50. Uma imemori yevidiyo inganele, yehlisa ukulungiswa noma uvule i-quantization.

Ukusebenza kahle kwamapharamitha: Ngamapharamitha angu-5.6B kuphela, idlula i-Qwen-Image 2.1 engu-7B ne-Krea 2 engu-12.9B ezivivinyweni eziyimpumputhe ezisemthethweni, futhi idlule i-Ideogram V4.5 evaliwe.
Imfuneko yehadiwe ephansi: Inguqulo ye-INT8 quantized ingasebenza ekhadini elivamile labathengi elinememori yevidiyo engu-8GB, okuyenza ibe yisixazululo somthombo ovulekile esinezindleko eziphansi kakhulu zokusetshenziswa phakathi kwamakhono asezingeni eliphezulu afanayo.
Imisebenzi emithathu ngesisindo esisodwa: Ukukhiqiza kusuka embhalweni, ukuhlela izithombe kanye nokukhiqiza ngezithombe eziningi eziyizethenjwa kusebenzisa isisindo esisodwa, ngaphandle kokushintsha noma ukuhlanganisa amamodeli emisebenzi ehlukene.
Ikhono lokulawula ukwakheka: Amabhokisi ezixhumanisi alawula ngokunembile indawo yezinto; ngaphambilini lokhu bekubonakala kuphela kumamodeli amakhulu avaliwe, kanti emkhakheni womthombo ovulekile akukabikho imbangi esezingeni elifanayo.
Ukwethembeka kokuhlela ezingeni lephikseli: Ngesikhathi sokuhlela, amaphikseli ezindawo ezingashiwongo awashintshi nhlobo, okuhlangabezana nezimo ezifana ne-e-commerce nomklamo lapho kudingeka kungalimazi isithombe sokuqala.
Ukukhipha okungokwendabuko kwe-4K: Khiqiza ngokuqondile ngokulungiswa kwe-4K ngaphandle kochungechunge lokukhulisa i-super-resolution, ugweme ukulahleka kwekhwalithi okubangelwa ukucutshungulwa kwesibili.

Iwebhusayithi yephrojekthi: https://bfl.ai/models/flux-3-image-klein
Indawo yamamodeli ku-HuggingFace: https://huggingface.co/collections/black-forest-labs/flux-3-image-klein

Ukukhiqizwa ngobuningi kwezithombe eziyinhloko ze-e-commerce ewela imingcele: Layisha izithombe eziyizethenjwa ezingu-10, sebenzisa amabhokisi ezixhumanisi ukucacisa izindawo zomkhiqizo, zamaphuzu okuthengisa nezamathegi entengo, bese ukhiqiza ngesikhathi esisodwa izinguqulo zezilimi eziningi zezithombe eziyinhloko zemikhiqizo ze-4K.
Ukuhlela izithombe zemikhiqizo ngokushintsha kwesizini: Esithombeni somkhiqizo esivele sikhona, faka umyalo othi «Yenza lo mgwaqo ube indawo eneqhwa»; shintsha ngokuphelele izizinda zesizini, kuyilapho amaphikseli omkhiqizo uqobo egcinwa iphuzu nephuzu, kukhiqizwe inguqulo yasebusika ngaphandle kwezindleko zokuhlela izithombe.
Ukuhlelwa okusheshayo kwamaphosta nama-Banner: Thola ngokulandelana indawo yesihloko, ye-Logo neyomfanekiso oyinhloko ngaphakathi kwamabhokisi ezixhumanisi zegridi engu-0–1000, bese ukhiqiza ngokuqondile iphosta yomcimbi lapho umbhalo ungaphambuki endaweni yawo.
Ukuklama amakhava ezincwadi noma omagazini: Sebenzisa amabhokisi ezixhumanisi ukulungisa indawo ehlobene yesihloko sencwadi nomfanekiso; ukukhipha okungokwendabuko kwe-4K kuhlangabezana ngokuqondile nokulungiswa kokuphrinta, kunciphise isidingo sokulungisa kabusha kusofthiwe yokuhlela.
Ukwenza ngokwezifiso i-Brand LoRA: Qeqesha i-LoRA ngezithombe zomkhiqizo wakho; ngemva kwalokho zonke izithombe ezikhiqizwayo zizokhiya ngokuzenzakalelayo imibala yomkhiqizo, ifonti nokubukeka komkhiqizo, kukhiqizwe ngobuningi izinto zemidiya yokuxhumana ezinesitayela esifanayo.