LTX-2.5 Uyasebenza

I-LTX-2.5 iyimodeli eyisisekelo ye-AI yokukhiqiza amavidiyo evulekile ye-LTX, enamapharamitha angu-22B futhi enesisindo esivulekile kusukela ngosuku lokukhishwa. Imodeli isekela ukukhiqizwa kwamashothi amaningi ngokwemvelo, i-4K HDR kanye nezindlela zokusebenza zobuchwepheshe ze-RAW/EXR, futhi ingakhiqiza umsindo ngesikhathi esifanayo, inamakhono anembile okuhlela amavidiyo. Njengoba amamodeli avaliwe eqhubeka nokubiza kakhulu, i-LTX-2.5 inikeza enye indlela evulekile esezingeni lobuchwepheshe emkhakheni wamavidiyo e-AI, isebenzisa isu le-ecosystem evulekile.

Ukukhiqizwa kwamashothi amaningi ngokwemvelo: Ekukhiqizeni okukodwa, ingakhipha amashothi aqhubekayo anezindawo zekhamera namabanga ahlukene. Igcina ukufana okuphezulu kokubukeka kwabalingiswa, umoya wendawo, izimo zokukhanya nomsindo lapho kushintshwa amashothi, ixazulule inkinga eyinhloko yokuqhubeka kwamashothi ekudalweni kwezindaba ze-AI.
Ukureyindwa okugcinayo ukwethembeka: Amandla okubala anikezwa ngokuzenzakalelayo ngokobunkimbinkimbi besigcawu. Amashothi anokunyakaza okuyinkimbinkimbi athola izinsiza ezengeziwe, kuyilapho izigcawu ezimile noma ezilula zonga izinsiza ngokuzenzakalelayo, kuthuthukiswe ukusebenza kahle ngaphandle kokwehlisa ikhwalithi yesithombe.
Ukuhlela amavidiyo okunembile: Kugcinwa umkhondo wokunyakaza nesakhiwo sendawo sezinto ezithwetshuliwe, bese kwenziwa ukuguqulwa kwesitayela, ukudwetshwa kabusha kwesigcawu nokushintshwa kwabalingiswa ngokunemba okuphezulu, ukuze kudalwe ngendlela elawulekayo “ngohlaka lokuthwebula kwangempela + ukumbozwa kwesitayela se-AI”.
Ukubikezela ubude ngokuzenzakalelayo: Imojuli ye-Duration Head eyakhelwe ngaphakathi ifunda izincazelo zezenzo ku-prompt bese inquma ngokuzenzakalelayo ubude besiqeshana obufanele kakhulu, ngaphandle kokuthi umdali azame izikhathi eziningi.
Amafomethi okukhipha ezinga lobuchwepheshe: Isekela okukhipha kwe-4K HDR, RAW ne-EXR okungalahlekelwa ikhwalithi, futhi ingaxhunywa ngokuqondile ezinqubweni zobuchwepheshe zokulungisa imibala, ukuhlanganisa imiphumela nokuhlela ngemuva.
Ukukhiqizwa komsindo okuvumelanayo: Isebenzisa imojuli ezimele ye-Audio VAE ukukhiqiza ithrekhi yomsindo ehambisana nevidiyo ngesikhathi esifanayo.
Ukulawula okunembile kwe-IC-LoRA: Inikeza izindlela ezintathu zokulawula amavidiyo ayireferensi: imiphetho ye-Canny, ukujula kwe-Depth nokuma komzimba kwe-Pose, ukuze imiphumela ekhiqiziwe ilandele ngokuqinile ukwakheka, isakhiwo sendawo noma ukunyakaza kwabalingiswa okusezintweni eziyireferensi.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokushintshisana ngamaphrojekthi e-AI avulekile
I-encoder yombhalo ye-Gemma 4 12B: I-encoder yemodeli yolimi enkulu eyenziwe yi-Lightricks ngokukhethekile i-LTX, ekwazi ukugcina ngokuphelele ubudlelwano bezinto eziningi, imininingwane yezenzo, izincazelo zokukhanya nesiqondiso sekhamera kuma-prompt amade ayinkimbinkimbi, ngaphandle kokulahlekelwa ulwazi lwezigatshana lapho i-prompt iba yinde.
Prompt Enhancer: Imodeli esizayo engasindi eyandisa ngokuzenzakalela i-prompt emfushane yomsebenzisi ibe yiziqondiso ezinemininingwane ze-cinema, yehlise umkhawulo wobunjiniyela bama-prompt futhi isebenzise izindleko zokubala eziphansi kakhulu.
I-Diffusion Video Decoder: Isakhiwo esisha esithatha indawo ye-VAE decoder yendabuko, esilungiselelwe ngokukhethekile izigcawu ezihamba ngokushesha. Sehlisa kakhulu ukufiphala nezinto ezingamanga esithombeni, senze ubuso bucace futhi umbhalo ufundeke kalula.
I-Audio VAE: I-encoder ezimele yesikhala esifihlekile somsindo, enomsebenzi wokuguqula okuqukethwe kwevidiyo kube ukumelwa komsindo okufanayo, ukuze kukhiqizwe isithombe nomsindo ngokuvumelanayo.
Ukureyindwa kwe-Diffusion Fidelity: Kusetshenziswa isu lezigaba ezimbili elithi “qala ngesakhiwo, bese kuba imininingwane”. Esigabeni sokuqala, ukunyakaza, ukwakheka nohlaka lwekhamera kwakhiwa esikhaleni esifihlekile esicindezelwe ngesikhathi ngo-8×, bese kukhiqizwa ama-keyframe athembeke kakhulu ngokuzenzakalelayo; esigabeni sesibili, kusetshenziswa isakhiwo nama-keyframe ndawonye ekusakazeni ukuze kureyindwe ividiyo yokugcina, kuhlaziywe ukuthungwa, impahla nemininingwane yobuso ngokunemba kwezinga lamaphikseli.
I-Duration Head: Imojuli yokubikezela efakwa ngaphambi kwenqubo ye-diffusion. Ngokuqonda incazelo yesenzo ku-prompt, ibikezela ngokuzenzakalelayo futhi ikhiphe ubude bevidiyo obufanele kakhulu, inikeze imodeli ukuqonda kokuqala kwesigqi sokulandisa.

Khetha indlela: Nquma ukuthi uzosebenzisa isevisi ye-API yamafu noma ukufakwa kwendawo nge-ComfyUI.
Ukuxhuma nge-API: Faka isicelo se-API Key kuwebhusayithi esemthethweni ye-LTX, bese ukhetha izinga le-Fast noma le-Pro ukuze usebenzise i-interface.
Faka ama-node: Ekufakweni kwendawo, buyekeza i-ComfyUI ibe yinguqulo yakamuva, bese usebenzisa i-Manager ukufaka iphakheji yama-node ye-ComfyUI-LTXVideo.
Landa imodeli: Esipheqululini samatemplate se-ComfyUI, sesha “LTX-2.5″, khetha i-workflow ehambisanayo bese ulanda wonke amafayela emodeli ngokuchofoza okukodwa.
Hlela amapharamitha: Setha ububanzi nobude bube ukuphindaphindwa kuka-32, inani lamafreyimu libe ngu-1+ ukuphindaphindwa kuka-8, izinga lamafreyimu libe ngu-24/25/48/50fps, bese usetha isinqumo sokuqala sibe uhhafu wesinqumo esihlosiwe ukuze unike amandla ukukhulisa ngo-2×.
Bhala ama-prompt: Sebenzisa ipharagrafu ehambisanayo echaza ububanzi beshothi, ukukhanya kwesigcawu, izenzo zabalingiswa, ukunyakaza kwekhamera nomsindo; gwema ukunqwabelanisa amathegi noma i-syntax yezisindo.
Qalisa ukukhiqiza: Chofoza u-run. Umbhalo uye kuvidiyo ukhiqiza ividiyo ngokuqondile; isithombe uye kuvidiyo kudinga ukulayishwa kwefreyimu yokuqala; imodi yamafreyimu okuqala nawokugcina idinga ukulayishwa kwawo womabili.
Ukulawula okunembile: Abasebenzisi abathuthukile bangalayisha i-IC-LoRA yemodi ye-Canny, Depth noma Pose, baxhume ividiyo eyireferensi ukuze bakhiye ukwakheka noma ukunyakaza.
Ukukhetha imodeli: Sebenzisa imodeli ye-distilled ekuphindaphindeni okusheshayo, bese usebenzisa imodeli ephelele ye-dev ekukhiqizeni kokugcina ukuze uthole imininingwane ecwebe kakhulu.

Ukuhambisana kokulandisa: Ikhono lemvelo le-Multishot lokukhiqiza amashothi amaningi ligcina ukubukeka kwabalingiswa, umoya wendawo, izimo zokukhanya nomsindo kufana kakhulu ekukhiqizeni okukodwa, lixazule ngokuphelele inkinga enkulu yokuqhubeka kwamashothi ekudalweni kwezindaba ze-AI.
Ukulawuleka okuthuthukisiwe: Ukuhlela amavidiyo okunembile kugcina umkhondo wokunyakaza nesakhiwo sendawo sezinto ezithwetshuliwe, futhi kuhlanganiswe nezindlela ezintathu zokulawula ze-IC-LoRA ze-Canny/Depth/Pose, kuguqule ukukhiqiza kusuka “ekuzameni ngenhlanhla” kube “uhlaka lokuthwebula kwangempela + ukumbozwa kwesitayela se-AI”.
Inqubo yokusebenza yezimboni: Okukhiphayo kwemvelo kwe-RAW/HDR/EXR kungaxhunywa ngokuqondile ezinqubweni zobuchwepheshe zokulungisa imibala, ukuhlanganisa imiphumela nokuhlela ngemuva, ngaphandle kokucindezela noma ukuguqula ifomethi, kuhlangabezane nesidingo esiqinile semboni yamafilimu sezinto ezingalahlekelwa ikhwalithi.
Ukusebenza okuzivumelanisa nezimo: Ubuchwepheshe be-Diffusion Fidelity Rendering bunikeza amandla okubala ngokuya ngobunkimbinkimbi besigcawu. Amashothi anokunyakaza okuyinkimbinkimbi areyindwa ngokuningiliziwe, kanti izigcawu ezimile nezilula zonga izinsiza ngokuzenzakalelayo, kuthuthukiswe ukusebenza kahle kokukhiqiza ngenkathi kugcinwa ikhwalithi yezinga lamaphikseli.

Iwebhusayithi yephrojekthi: https://ltx.io/model/ltx-2-5
Ilabhulali yamamodeli ye-HuggingFace: https://huggingface.co/Lightricks/LTX-2.5

Ukukhiqizwa kwamafilimu nokuhlelwa kwawo ngemuva: Okukhiphayo kwemvelo kwe-4K HDR, RAW ne-EXR okungalahlekelwa ikhwalithi kungaxhunywa ngokuqondile ezinqubweni zokulungisa imibala, ukuhlanganisa imiphumela nokuhlela, kuhlangabezane nezidingo zokukhiqizwa kwezinga lamabhayisikobho namasevisi okusakaza, ngaphandle kokubhekana nenkinga yokulungisa imibala ye-MP4 ecindezelwe.
Amadrama amafushane nama-comic dramas e-AI: Ukukhiqizwa kwamashothi amaningi kwe-Multishot kukhipha i-storyboard ehambisanayo ngesikhathi esisodwa, kuyilapho ukuhlela amavidiyo okunembile kusekela “uhlaka lokunyakaza lokuthwebula komuntu wangempela + ukushintshwa kwabalingiswa be-AI nokumbozwa kwesitayela”, kuxazulule ngokuphelele izinkinga ezimbili ezinkulu zokuqhubeka kwamashothi nokulawuleka kokunyakaza.
Izikhangiso nokuqukethwe kwemikhiqizo: Phindaphinda ngokushesha amavidiyo ebhizinisi asezingeni eliphezulu. Sebenzisa i-Prompt Enhancer ukwehlisa umkhawulo wobunjiniyela bama-prompt, imodeli ye-Distilled ibonise ngokushesha, bese imodeli ephelele ikhiqize ividiyo yokugcina, kuhlanganiswe ukusebenza kahle nekhwalithi.
I-Physical AI namarobhothi: I-Physical AI Checkpoint inikeza amaqembu obuhlakani obugxilile isisekelo semodeli yomhlaba, isize izinhlelo zamarobhothi zibone umhlaba obonakalayo nemithetho yokunyakaza, idlule umkhawulo wokudala okuqukethwe kuphela.