Kugadzikana AI yakaburitsa modhi yekudzidza yemuchina, Yakagadzikana Vhidhiyo Diffusion, inogona kuburitsa mavhidhiyo mapfupi kubva pamifananidzo. Iyo modhi inowedzera kugona kweiyo Stable Diffusion purojekiti, iyo yaimbove yakaganhurirwa pakugadzira mifananidzo isingachinji. Iyo kodhi yeiyo neural network yekudzidziswa uye maturusi ekugadzira mifananidzo yakanyorwa muPython uchishandisa iyo PyTorch chimiro uye yakaburitswa pasi peMIT rezinesi. Iwo mamodheru akadzidziswa akavhurika akavhurika pasi peCreative ML OpenRAIL-M rezinesi rezinesi, rinobvumira kushandiswa kwekutengesa.
Mhando mbiri dzakasiyana dziripo kudhawunirodha: SVD (Yakagadzikana Vhidhiyo Diffusion), iyo inogadzira gumi nemana mafuremu pa576 × 1024 resolution yakavakirwa pane yakapihwa static mufananidzo, uye SVD-XT, iyo inogadzira makumi maviri neshanu mafuremu. Kugadzirwa kwevhidhiyo kunogoneka, kungave kusingafambi kana nekunonoka kutenderera kwekamera, isingapfuuri masekonzi mana. Kudzora kwakananga kweiyo modhi kunoenderana nemutauro wechisikigo hakusati kwatsigirwa, asi zvinokwanisika kutanga wagadzirira iyo sosi yemufananidzo uchishandisa yekare Stable Diffusion 2.1 modhi wozoishandura kuita vhidhiyo uchishandisa iyo SVD modhi.
Vhidhiyo mhando haisati yapa yakakwana photorealism kana vimbiso yechokwadi yekupihwa zviso nevanhu. Panyaya yekuita, iyo yakavhurwa-sosi modhi inodarika evaridzi analogi kubva kuRunway nePika Labs. Iyo modhi inogona kuchinjika nyore kune akasiyana mabasa, sekugadzira matatu-dimensional manhamba.
Chinocherechedzwa zvakare kuburitswa kweVhidhiyo-LLaVA muchina wekudzidza wekushandisa, iyo inogonesa kugadzirwa kweyakabatana inomiririra inomiririra yechinhu, inoumbwa nekushandisa zvese mafoto uye vhidhiyo kurekodha kwechinhu panguva yekudzidziswa. Iyo sisitimu inogona kushandiswa, semuenzaniso, kuziva kuvepo kwezvinhu zvakafanana mumifananidzo nemavhidhiyo. Iyo kodhi yakanyorwa muPython uye yakagoverwa pasi peiyo Apache 2.0 rezinesi.
Source: opennet.ru
