WeVisCut Uyasebenza

I-WeVisCut iyi-Agent yokuhlela ngokuzenzakalela amavidiyo amade ekhishwe yi-Tencent WeChat Vision njengephrojekthi yomthombo ovulekile. Ifaka umbhalo wokulandisa, bese isu elincane le-9B lisebenzisa i-runtime enesimo sokukhiqiza esabelwana ngaso ukuze liqede ngokuzenzakalela ukwakhiwa kwezigcawu, ukuqoshwa kwezwi, ukusesha izinto, ukuhlanganisa nokulungisa ikhwalithi, likhiphe ngokuqondile ividiyo eqediwe yemizuzu eminingana. I-WeVisCut isebenzisa iphayiphi ye-editing inversion + verified replay ukuze ibuyise ngaphezu kuka-130,000 wemininingwane yokuqeqesha kusuka kumavidiyo aqediwe angu-3,750. Ekuhlolweni, ukusebenza kwayo kulingana no-Claude Opus 5, konga ama-token okukhiqiza cishe izikhathi ezingu-33, kanti izindleko ngomzuzu wevidiyo eqediwe zingama-$0.09 kuphela.

Ukuguqula umbhalo ube yividiyo ende: Faka umbhalo wokulandisa noma wolwazi, bese ikhipha ngokuzenzakalela ividiyo ephelele yemizuzu eminingana ngaphandle kokwakha izigcawu ngesandla.
Isu elilodwa elithatha izinqumo kulo lonke uchungechunge: Imodeli eyodwa ye-9B idlala izindima ezintathu zokuhlela, ukukhiqiza nokugxeka, ithathe indawo yokugeleza komsebenzi kwama-Agent ezigaba eziningi.
Isimo sokukhiqiza esabelwana ngaso: I-harness igcina ngokuqhubekayo isimo sezigcawu, izinto nomugqa wesikhathi, bese imiphumela yethuluzi lesinyathelo sangaphambilini ibopha ngokuqondile izinqumo zesinyathelo esilandelayo.
Ukuhlola nokulungisa okuvalwe ngesiyingi: Isu lihlola ngokuzenzakalela ividiyo eqediwe, lithole ukwehluleka futhi lenze ukulungisa okulinganiselwe endaweni ethile, ngaphandle kokuphinda lenze izingxenye eziphumelele.
Ukuhlela ngezingxoxo nomqondisi: Ngemva kwevidiyo yokuqala, ungashintsha izinto, umculo nemibhalo engezansi ngengxoxo; isu lihlela kabusha ingxenye ethile futhi lilethe umphumela omusha.
Ukulethwa komugqa wesikhathi ohlelekayo: I-runtime ye-Studio inikeza umugqa wesikhathi onamathrekhi angu-8, osekelwa ukunikezwa kabusha kwesigcawu ngasinye nokubuyiselwa kwezinguqulo.
Ukusetshenziswa kwangasese ngezindleko eziphansi: Umzuzu ngamunye okhiqizwe ngempumelelo ubiza cishe u-$0.09 kanye nemizuzu engu-2.1 yesikhathi sewashi; izisindo ze-9B zomthombo ovulekile zingafakwa endaweni yangakini.

Landela i-WeChat uphendule ngokuthi “开源” ukuze ujoyine iqembu lokuxoxa ngamaphrojekthi e-AI omthombo ovulekile
Incazelo yenkinga: Ukuphathwa kwesimo ekukhiqizeni okunochungechunge olude. Ubunzima bokuguqula umbhalo wevidiyo ende ube yividiyo busekuhlobananeni kokukhethwa kwezinto, ukusetshenziswa kwamathuluzi nezinqumo zokulungisa; ukunemba kwengxenye ngayinye yesigcawu akusho ukuthi ividiyo yonke inesigqi. Uma ihlukaniswa yaba yizigaba eziningi zama-Agent ajwayelekile, isigaba ngasinye kufanele siphinde silandele umongo womhlaba wonke, okwenza izindleko zama-token nezokubiza amathuluzi zikhule ngokulingana nenani lezigcawu. Umqondo oyinhloko we-WeVisCut ukusebenzisa isu elilodwa le-9B eligcina sonke isimo sokukhiqiza, liguqule “ukuqonda konke” kusukela entweni ephindaphindwa esinyathelweni ngasinye kube yinto eqhubekayo eyenziwa kanye.
I-Harness enesimo: Isimo sokukhiqiza esabelwana ngaso. I-Runtime igcina isimo sokukhiqiza esiphikelela, futhi ibhala kuphela imiphumela yamathuluzi eqinisekisiwe kuleso simo. Isu lithumela ucingo olulodwa lwethuluzi ngesikhathi, likhethe isenzo esisemthethweni esimweni samanje; izingcingo ezingekho emthethweni ziyenqatshwa njenge-blocker enohlobo, okuphoqa ukuthi kucwaliswe kuqala okuncikile kukho bese kuzanywa futhi. Isimo nomongo wemodeli kwehlukanisiwe, ngakho isu alidingi ukuphatha umlando ophelele ngokuphindaphindiwe.
Isu elilodwa elinezindima ezintathu: Isu elinamapharamitha abelwana ngawo lisebenza njengomhleli, umkhiqizi nomgxeki ngokusebenzisa ukuqondiswa kwezindima. Isikhala sezenzo sigxilile kumathuluzi angu-12, ngaphandle komshini wezigaba obhalwe kanzima; umzila wonke unqunywa yisu ngokuzimela, kuyilapho izivikelo zivimba kuphela izingcingo ezingekho emthethweni.
Idatha yokuqeqesha: I-Editing Inversion + Verified Replay. Ividiyo eqediwe ikhombisa umphumela kuphela, ayivezi inhloso yokuhlela. Ababhali basebenzisa iphayiphi enezigaba ezine ukubuyisa ukuqondisa: kuqala babuyisa umbhalo nenhloso yesigcawu kusuka kuvidiyo eqediwe, bese beyidlala kabusha ku-runtime ukuze baqinisekise ukuthi iyasebenza, balungise amaphutha ngokulinganiselwe lapho ukudlalwa kubonisa izinkinga futhi bagcine imikhiqizo ephumelele, ekugcineni baguqule amarekhodi okwenziwa abe yizibonelo zokuqeqesha ezinempendulo yamathuluzi. Amavidiyo aqediwe angu-3,750 aguqulwe aba yidatha yokuqondisa esebenzayo engu-134,771, okwenza imodeli ikwazi ukufunda ukubuyela esimweni esifanele ngemva kokwehluleka.
Ukuqeqeshwa kwezigaba ezintathu: I-SFT ehlanganisiwe kanye ne-RFT enokuqinisekiswa yi-runtime kuqalisa ukusetshenziswa kwamathuluzi nokuziphatha kokululama; i-Guided-OPD yethula ukungenelela kukathisha ukuze kuqeqeshwe izinqumo zokululama, kuncishiswe kancane kancane ukungenelela njengoba ukuqeqeshwa kuqhubeka; i-RCF-DAPO iqhathanisa imiphumela engaba khona endimeni efanayo nasemongweni ofanayo wokukhiqiza ukuze ithuthukise ikhwalithi yezinqumo. Lezi zigaba ezintathu zikhuphule i-Overall isuka ku-74.5 yesisekelo yaya ku-90.5.

Thola ikhodi: Klona i-repository ku-GitHub: github.com/Tencent/WeVisCut.
Lungisa imvelo: Faka i-runtime ye-Studio kanye nokuncika kwayo ngokwemiyalelo ye-repository.
Download imodeli: Thola izisindo ze-9B ze-tencent/WeVisCut ku-HuggingFace.
Qalisa isevisi: Sebenzisa iskripthi sokuqalisa ukuze ulayishe isu nerejista yamathuluzi, bese uvula isixhumi esibonakalayo sewebhu se-Studio.
Faka umbhalo: Namathisela noma ulayishe umbhalo omude wokulandisa noma wolwazi kusixhumi esibonakalayo.
Buka ukukhiqizwa: I-Agent ihlukanisa ngokuzenzakalela izahluko nezigcawu, iqophe izwi, iseshe izinto futhi yakhe umugqa wesikhathi ngesikhathi sangempela; inqubo iyabonakala.
Hlela ngengxoxo: Ngemva kwevidiyo yokuqala, phakamisa izinguquko efasiteleni lengxoxo nomqondisi ukuze kulethwe kabusha ingxenye ethile.
Thumela ividiyo eqediwe: Qinisekisa isigcawu ngasinye kusixhumi esibonakalayo somugqa wesikhathi bese uthumela ividiyo yokugcina, noma ubuyisele inguqulo ethile futhi unikeze kabusha njengoba kudingeka.

Izindleko ziphansi kakhulu: Umzuzu ngamunye wevidiyo okhiqizwe ngempumelelo udinga ama-token okukhiqiza angu-2,377 kuphela kanye no-$0.09; imodeli encane ye-9B ingafakwa endaweni yangasese.
Ukulethwa okuthembekile: I-Overall engu-90.5 ezinkomba ezinhlanu idlula isisekelo ngamaphuzu angu-16; izinga lokufeza izinhlelo, ukuhambisana kwesikhathi eside nezinga lokukhiqizwa kokunikezwa kuhamba phambili ngokuphelele.
Isu elilodwa liphatha yonke inqubo: Imodeli eyodwa ye-9B idlala izindima zokuhlela, ukukhiqiza nokugxeka, ithathe indawo yokugeleza komsebenzi kwama-Agent ezigaba eziningi futhi igweme ukwakhiwa kabusha komongo okuphindaphindiwe.
Ukuphathwa kwesimo okuqhubekayo: Isimo sokukhiqiza esabelwana ngaso senza umphumela wethuluzi langaphambilini ubophe ngokuqondile isinqumo esilandelayo, ukuze uchungechunge olude lungalahlekelwa isigqi futhi lungaphindi luqonde umongo wonke.
Iyakwazi ukuzilungisa: Idatha yokuqeqesha iqukethe imizila yokululama ngemva kokwehluleka; isu lingahlola ikhwalithi, lithole inkinga futhi lenze ukulungisa okulinganiselwe endaweni ethile, ligcine izingxenye eziphumelele.
Ukuhlela okusebenzisanayo: Ngemva kokukhiqizwa kwevidiyo, ungasebenzisa ingxoxo yolimi lwemvelo ukuze wenze izinguquko; kuhlelwa kabusha ingxenye ethile ngaphandle kokuqopha kabusha.

Iwebhusayithi esemthethweni yephrojekthi: https://tencent.github.io/WeVisCut/
Indawo yokugcina ye-GitHub: https://github.com/Tencent/WeVisCut
Indawo yemodeli ye-HuggingFace: https://huggingface.co/tencent/WeVisCut
Iphepha lobuchwepheshe le-arXiv: https://arxiv.org/pdf/2610.38352

Ukukhiqizwa ngobuningi kwamavidiyo olwazi: Guqula ngokuzenzakalela imibhalo emide yolwazi noma yokulandisa ibe amavidiyo achazayo emizuzu eminingana, alungele i-MCN nabadali abazimele abakhipha okuqukethwe nsuku zonke.
Ukuqeqeshwa kwezinkampani nokukhiqizwa kwezifundo: Guqula ngokushesha izincwadi zokuqeqesha nokulandisa kwezifundo kube amavidiyo aqediwe, unciphise izindleko zokukhiphela ngaphandle amavidiyo okuqeqesha angaphakathi.
Ukuboniswa kwezindaba nolwazi: Hlanganisa ngokuzenzakalela izithombe nomculo emibhalweni yezindaba zezimali nezobuchwepheshe, okuzivumelanisa nokushicilelwa okusheshayo kwama-akhawunti ezindaba.
Ukumaketha nokuchaza imikhiqizo: Khiqiza ngokuchofoza okukodwa inguqulo yokuqala yevidiyo yokukhangisa kusuka embhalweni womkhiqizo, bese abasebenzi bezokumaketha belungisa izinto nesitayela ngengxoxo.
Ukuyisa okuqukethwe ezimakethe zezilimi eziningi: Ngemva kokuhumusha umbhalo, khiqiza kabusha ngokuqondile izinguqulo zokuqoshwa kwezwi nemibhalo engezansi, uphinde usebenzise iphayiphi efanayo yokukhiqiza ngobuningi.