BlenderWorld Uyasebenza

I-BlenderWorld iyibhentshimakhi yokuhlola ama-Agent okudala i-3D eyethulwe ngokubambisana yiQwen team, iZhejiang University, iChinese University of Hong Kong, iFudan University kanye neNational University of Singapore. Iqukethe imisebenzi yangempela engu-223, futhi ihlola ukuthi i-AI ingakwazi yini ukwakha nokuletha iphrojekthi ye-3D ehlelekayo ku-Blender isebenzisa izithombe, amavidiyo noma imininingwane yokuklama. I-BlenderWorld isebenzisa ukuhlolwa kwamathrekhi amabili, okuhlanganisa ukunikezwa kwezithombe kusuka kuma-engeli amaningi kanye nokuhlolwa kwephrojekthi yomdabu, ikala ukuqedwa kwezidingo nekhwalithi yokubukeka kusetshenziswa izindinganiso zokwamukelwa ezingu-2,651. Ezinhlelweni ezingu-13 ezihloliwe, i-GPT-6-Astra enamandla kakhulu nayo yaphumelela ngokuphelele emisebenzini engu-42.2% kuphela.

Ukuhlolwa kwenqubo ephelele yokudala i-3D: I-Agent kumele iqale ngezidingo nezinto eziyizithenjwa, iqedele ukumodela, izinto zokwakha, ukukhanyisa, ikhamera, ukugqwayiza nezinye izigaba, bese iletha iphrojekthi ye-.blend ehlelekayo.
Ukusekelwa kwezinto eziyizithenjwa eziningi: Imisebenzi ihlanganisa izindlela ezintathu zokufaka, okuyizithombe, amavidiyo namafayela emininingwane emsulwa.
Ukumbozwa kwemisebenzi emikhakheni ebanzi: Kukhavwe izigaba ezinkulu ezingu-9 nezigaba ezingaphansi ezingu-22, okuhlanganisa izakhiwo nezangaphakathi, ukukhiqizwa kwemikhiqizo, izimoto nemishini, abalingiswa nezilwane, izindawo zemvelo, ukugqwayiza, ukulingisa kanye nokubonwa kwesayensi.
Ukuhlolwa okuzenzakalelayo kwamathrekhi amabili: I-VLM ihlola izithombe namavidiyo anikeziwe, kuyilapho i-Agent ehlola iphrojekthi iqinisekisa ukwakheka nezici zangaphakathi zefayela le-.blend.
Ukukhiqizwa kobufakazi obuvela kuma-engeli amaningi: Kukhethwa ngokuzenzakalelayo ama-engeli angu-8 ahambisanayo ephrojekthi ethunyelwe, bese kusetshenziswa i-Cycles ukunikeza ubufakazi obengeziwe bokubukeka.
Ukulinganisa okuphelele ngezilinganiso ezimbili: Isikolo esiphelele sibalwa ngo-0.5 × Rubric + 0.5 × Dimension, futhi kubikwe izinga lokuphasa le-Perfect.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI avulekile
Ukwakhiwa kwemisebenzi nezindinganiso zokwamukelwa: Abacwaningi basebenzisa amamodeli anezindlela eziningi zokufaka ukuze bakhiqize futhi bahlele kabusha izidingo zemisebenzi, bese abantu behlola ukufundeka kwezinamathiselo nokuvumelana phakathi kombhalo nezithombe. Izimo zokwamukelwa zikhishwa ezidingweni, ezintweni eziyizithenjwa nasezinamathiselweni. Ngemva kokuhlolwa kokuqala nokuhlolwa okuphambanayo okwenziwa okungenani amamodeli amabili azimele okuhlola, lezi zimo ziqiniswa zibe yizindinganiso ezimbili ezingu-6—22, zabelwe izisindo ngokubaluleka kwazo (3/2/1), futhi azibonakali ku-Agent ekhiqizayo ukuze kuvinjelwe ukukhohlisa kokuhlolwa.
Ukuqoqa ubufakazi ngokunikezwa kwama-engeli amaningi: Indawo yokuhlola ezimele ifunda iphrojekthi ye-.blend ethunyelwe. I-Automatic Agent Render ikhetha ngokuzenzakalelayo ama-engeli angu-8 ahambisanayo ngokusekelwe emsebenzini nasezimweni zokwamukelwa, bese isebenzisa i-Cycles ukuyinikeza ngokulungiswa okungu-1024×1024 kanye nama-samples angu-64. I-Agent ehlolayo ingalungisa ikhamera nezilungiselelo zokunikezwa, kodwa ayivunyelwe ukulungisa okuqukethwe kwephrojekthi. Imidiya ekhishwe ekuqaleni iyona ethathwa njengeyisisekelo, kanti imisebenzi yamavidiyo yengezwa ngamafreyimu athathwe ngokulandelana njengobufakazi bokunyakaza.
Indlela yokuhlola yamathrekhi amabili: I-VLM-as-Judge isebenzisa izithombe namavidiyo ukwahlulela ukubukeka, ukuhlelwa kwendawo nokunyakaza. I-Agent-as-Judge ihlola ngokuqondile ukwakheka komdabu nezici zephrojekthi ye-.blend isebenzisa amathuluzi, futhi ibhekisa kubufakazi bokubukeka kuphela lapho kudingeka. Lokhu kuqinisekisa ubukhulu obuqondile, ukwakheka kwangaphakathi nolunye ulwazi olungabonakali ezithombeni ezinikeziwe.

Faka i-ModelScope SDK: Qalisa u-pip install -U modelscope ukuze ufake ithuluzi lokulanda amasethi edatha.
Landa isethi yedatha yebhentshimakhi: Shayela i-dataset_snapshot_download(dataset_id="Qwen/BlenderWorld", local_dir="./BlenderWorld-data") ukuze ulande imisebenzi engu-223 endaweni yangakini.
Xhuma i-Agent yakho ye-3D: Vumela i-Agent yakho ifunde izidingo zomsebenzi nezinto eziyizithenjwa, yenze umsebenzi wokudala endaweni ye-Blender, bese ikhipha iphrojekthi ye-.blend namafayela emidiya ahambisanayo.
Thumela umsebenzi ukuze uhlolwe: Thumela iphrojekthi endaweni yokuhlola ezimele, lapho uhlelo luzokwenza ngokuzenzakalelayo ukunikezwa kwama-engeli amaningi nokuhlolwa kwamathrekhi amabili.
Buka imiphumela yokulinganisa: Thola isikolo sokwamukelwa se-Rubric, isikolo sokubukeka se-Dimension, isikolo esiphelele kanye nezinga lokuphasa le-Perfect, bese usebenzisa imikhondo yokwehluleka ukuthola izindawo ezibuthakathaka ze-Agent.

Izinhloso zokuhlola zisondelene nezidingo zangempela zokukhiqiza: Kudingeka iphrojekthi ye-.blend ehlelekayo, hhayi isithombe esisodwa somphumela; uma iphrojekthi ingathunyelwanga, kunikezwa amaphuzu aziro.
Izindinganiso zokwamukelwa ziqinile futhi zivimbela ukukopela: Izindinganiso ezimbili ezingu-2,651 ziqiniswa ngemva kwemijikelezo eminingi yokuhlolwa okuphambanayo, futhi azibonakali ku-Agent ekhiqizayo, ngaleyo ndlela kuncishiswe kakhulu amathuba okukhohlisa okuqondisiwe.
Ukuhlolwa kwamathrekhi amabili kugcwalisa izikhala zokubona: Ukuhlanganisa ukunikezwa kwama-engeli amaningi nokuhlolwa kwephrojekthi yomdabu kwenza kube nokwenzeka ukuthola amaphutha ekwakhekeni kwangaphakathi, ebukhulu obuqondile, ekucushweni kokugqwayiza nakwezinye izici ezingavezwanga yisithombe esisodwa esivela ku-engeli eyinhloko.
Uhlelo lokulinganisa lubheka kokubili ukuqedwa nekhwalithi yokubukeka: I-Rubric ne-Dimension ngayinye inesisindo esingu-50%, futhi kunenkomba ye-Perfect ehlukanisa “ukumbozwa kwezidingo okuphezulu” “nokulethwa okuphelele nokuthembekile”.
Imisebenzi ibanzi futhi izinhlobo zezinto eziyizithenjwa ziyahlukahluka: Imisebenzi engu-223 ihlanganisa izigaba ezinkulu ezingu-9 nezigaba ezingaphansi ezingu-22, futhi isekela izindlela ezintathu zokufaka, okuyizithombe, amavidiyo namafayela emininingwane emsulwa.

Iwebhusayithi esemthethweni: https://blenderworld.github.io/
Inqolobane ye-GitHub: https://github.com/QwenLM/BlenderWorld
Inqolobane yamamodeli ye-HuggingFace: https://huggingface.co/datasets/Qwen/BlenderWorld
Iphepha lobuchwepheshe: https://github.com/QwenLM/BlenderWorld/blob/main/docs/BlenderWorld__Benchmarking_Visual_Agents_for_3D_Creation_from_Real_World_References.pdf

Ukuhlola amakhono e-3D Agent: Kunikeza amaqembu athuthukisa ama-Agent azenzakalelayo e-Blender indawo yokuhlola ejwayelekile yokukala amakhono amathathu okudala: ukwakha kabusha, ukulungisa nokuklama.
Ukuqhathanisa nokukhetha amamodeli amakhulu: Ngaphansi kwezindinganiso zokwamukelwa ezifanayo ezingu-2,651, kuqhathaniswa ukuthembeka kokulethwa kwe-3D kwamamodeli ahlukahlukene avaliwe navulekile ukuze kusekelwe ukukhethwa kobuchwepheshe.
Ukuhlolwa kokubuyela emuva ekuphindaphindweni kwamamodeli: Ibhentshimakhi iqhutshwa ngemva kokuthuthukiswa ngakunye kwe-Agent noma kwemodeli eyisisekelo, kusetshenziswa isikolo se-Perfect nemikhondo yokwehluleka ukuthola ukwehla noma ukuthuthuka kwamakhono.
Ukuhlaziya ukwehluleka nokuqinisa amakhono: Ubufakazi obuvela kuma-engeli amaningi nemikhondo yokwenza kusetshenziswa ukuhlonza amaphutha ajwayelekile, kuqondiswe ukuthuthukiswa kwengqondo yokuzihlola namasu okulungisa.
Ibhentshimakhi yocwaningo nohlu lomphakathi: Njengebhentshimakhi yezemfundo evulekile, isekela ukuhlolwa kwamaphepha, amazinga omphakathi kanye nokwakhiwa kwedatha yokuqeqesha emkhakheni wama-Agent okubona i-3D.