Kugadzikana AI yakaburitsa yakagadzirira-yakagadzirwa modhi yeStable Diffusion muchina wekudzidza sisitimu, inokwanisa kugadzira uye kugadzirisa mifananidzo zvichibva pakutsanangurwa kwemavara mumutauro wechisikigo. Mamodheru anopihwa rezenisi pasi pemvumo yeCreative ML OpenRAIL-M rezinesi rekushandiswa kwekutengesa. Kudzidzisa sisitimu, sumbu re4000 NVIDIA A100 Ezra-1 GPUs uye LAION-5B muunganidzwa, kusanganisira 5.85 bhiriyoni mifananidzo ine tsananguro yemavara, yakashandiswa. Pakutanga, iyo kodhi yezvishandiso zvekudzidzisa neural network uye kugadzira mifananidzo yaive yakavhurika-yakavhurika pasi peMIT rezinesi.
Kuvepo kweiyo yakagadzirira-yakagadzirwa modhi uye zvine mwero masisitimu zvinodiwa izvo zvinobvumira munhu kutanga kuyedza paPC ine akajairwa maGPU zvakakonzera kubuda kwehuwandu hwezvirongwa zvine hukama:
- textual-inversion (code) - add-on iyo inokutendera kuti ugadzire mifananidzo ine yakapihwa hunhu, chinhu kana chimiro. Mune yekutanga Yakagadzika Diffusion, zvinhu zviri mumifananidzo yakasanganiswa hazvina kurongeka uye hazvidzoreki. Iyo yakakurudzirwa yekuwedzera inokubvumira kuti uwedzere zvako zvinhu zvinoonekwa, zvisungirire kumazwi akakosha uye uzvishandise mukubatanidza.
Semuenzaniso, mune yenguva dzose Yakagadzikana Diffusion unogona kubvunza sisitimu kuti igadzire mufananidzo ne "katsi muchikepe". Pamusoro pezvo, iwe unogona kujekesa hunhu hwekatsi uye chikepe, asi hazvifungidzike kuti ndeipi katsi uye chikepe zvichagadzirwa. Textual-inversion inokutendera iwe kudzidzisa sisitimu pamufananidzo wekatsi yako kana chikepe uye gadzira chifananidzo nekatsi kana chikepe. Nenzira imwecheteyo, inogona zvakare kutsiva zvinhu zvemufananidzo nezvimwe zvinhu, isa muenzaniso wechimiro chechimiro chekubatanidza, uye tsanangura pfungwa (semuenzaniso, kubva kune vakasiyana vese vanachiremba, unogona kushandisa yakanyanya kunaka uye yemhando yepamusoro sarudzo. muchimiro chaunoda).
- yakagadzika-diffusion-animation - kusikwa kwemifananidzo yehupenyu (inofamba) zvichienderana nekududzirwa pakati pemifananidzo inogadzirwa muStable Diffusion.
- stable_diffusion.openvino (code) - chiteshi cheStable Diffusion, chinoshandisa CPU chete yekuverenga, iyo inobvumira kuedza pamasisitimu asina maGPU ane simba. Inoda processor inotsigirwa muraibhurari yeOpenVINO. Zviri pamutemo, OpenVINO inopa mapulagi eIntel processors ane AVX2, AVX-512, AVX512_BF16 uye SSE ekuwedzera, pamwe neRaspberry Pi 4 Model B, Apple Mac mini uye NVIDIA Jetson Nano mabhodhi. Zvisizvo, zvinokwanisika kushandisa OpenVINO pane AMD Ryzen processors.
- sdamd ichiteshi che AMD GPUs.
- Kuitwa kwekutanga kwevhidhiyo synthesis.
- yakagadzikana-kupararira-gui, yakagadzikana-kupararira-ui, Artbreeder Collage, diffuse-the-rest - graphical interfaces yekugadzira mifananidzo uchishandisa Stable Diffusion.
- beta.dreamstudio.ai, Hugging Face Spaces, hlky Stable Diffusion WebUI - mawebhusaiti ekubatanidza mifananidzo uchishandisa Stable Diffusion.
- Plugins yekubatanidza Yakagadzikana Diffusion neGIMP, Figma, Blender uye Photoshop.
Pamusoro pezvo, tinogona kucherechedza kuburitswa neGoogle kwekodhi yeRawNeRF (RAW Neural Radiance Fields) yekudzidza muchina, iyo inobvumira, zvichibva pane data kubva kune akati wandei RAW mifananidzo, kunatsiridza kunaka kwemifananidzo ine ruzha inotorwa murima uye mukati. chiedza chisina kunaka. Pamusoro pekubvisa ruzha, maturusi akagadzirwa nepurojekiti anoita kuti zvikwanise kuwedzera ruzivo, kubvisa kupenya, kusanganisa HDR uye kushandura mwenje wakazara mumifananidzo, pamwe nekugadzirisa nzvimbo yemativi matatu ezvinhu uchishandisa akati wandei mafoto kubva kumakona akasiyana, shandura maonero, shandura kutarisa uye gadzira mifananidzo inofamba.
Source: opennet.ru