Stable Diffusion 2.0 Image Synthesis System Adayambitsidwa

Kukhazikika kwa AI yasindikiza kope lachiwiri la Stable Diffusion makina ophunzirira makina, omwe amatha kupanga ndikusintha zithunzi kutengera template yomwe akufuna kapena kufotokozera kwachiyankhulo chachilengedwe. Khodi ya maphunziro a neural network ndi zida zopangira zithunzi zalembedwa mu Python pogwiritsa ntchito dongosolo la PyTorch ndikusindikizidwa pansi pa layisensi ya MIT. Mitundu yophunzitsidwa kale ndi yotsegulidwa pansi pa chilolezo cha Creative ML OpenRAIL-M chololeza, chololeza kugwiritsidwa ntchito pamalonda. Kuphatikiza apo, jenereta ya zithunzi zapa intaneti ilipo.

Kusintha kwakukulu mu mtundu watsopano wa Stable Diffusion:

  • Chitsanzo chatsopano cha kaphatikizidwe kazithunzi kutengera kufotokozera kwalemba kwapangidwa - SD2.0-v, yomwe imathandizira kupanga zithunzi ndi kusamvana kwa 768x768. Chitsanzo chatsopanocho chimaphunzitsidwa pogwiritsa ntchito gulu la LAION-5B, lomwe limaphatikizapo zithunzi za 5.85 biliyoni zokhala ndi malemba. Mtunduwu umagwiritsa ntchito magawo omwewo monga mtundu wa Stable Diffusion 1.5, koma umasiyana pakusintha kugwiritsa ntchito encoder yosiyana kwambiri ya OpenCLIP-ViT/H, yomwe yasintha kwambiri mawonekedwe azithunzi zomwe zatuluka.
    Stable Diffusion 2.0 Image Synthesis System Adayambitsidwa
  • Mtundu wosavuta wa SD2.0-base wakonzedwa, wophunzitsidwa pazithunzi za 256 Γ— 256 pogwiritsa ntchito mtundu wakale wolosera zaphokoso ndikuthandizira m'badwo wa zithunzi ndikusintha kwa 512 Γ— 512.
    Stable Diffusion 2.0 Image Synthesis System Adayambitsidwa
  • Ndizotheka kugwiritsa ntchito ukadaulo wa supersampling (Super Resolution) kuti muwonjezere kusintha kwa chithunzi choyambirira popanda kuchepetsa mtundu, pogwiritsa ntchito makulitsidwe apakati komanso ma algorithms omanganso mwatsatanetsatane. Mtundu woperekedwa wokonza zithunzi (SD20-upscaler) umathandizira makulitsidwe kanayi, omwe amalola kupanga zithunzi ndi kusamvana kwa 2048x2048.
    Stable Diffusion 2.0 Image Synthesis System Adayambitsidwa
  • Mtundu wa SD2.0-depth2img umaperekedwa, poganizira zakuya ndi makonzedwe a malo a zinthu. Pakuyerekeza kwakuya kwa monocular, dongosolo la MiDaS limagwiritsidwa ntchito. Chitsanzochi chimakupatsani mwayi wopangira zithunzi zatsopano pogwiritsa ntchito chithunzi china ngati template, chomwe chingakhale chosiyana kwambiri ndi choyambirira, koma sungani zolemba zonse ndi kuya. Mwachitsanzo, mutha kugwiritsa ntchito mawonekedwe a munthu pachithunzi kupanga munthu wina yemwe ali pachithunzi chomwecho.
    Stable Diffusion 2.0 Image Synthesis System Adayambitsidwa
    Stable Diffusion 2.0 Image Synthesis System Adayambitsidwa
    Stable Diffusion 2.0 Image Synthesis System Adayambitsidwa
  • Mtundu wosinthira zithunzi wasinthidwa - SD 2.0-inpainting, yomwe imakulolani kuti musinthe ndikusintha magawo a chithunzicho pogwiritsa ntchito mawu olimbikitsa.
    Stable Diffusion 2.0 Image Synthesis System Adayambitsidwa
  • Mitunduyo idakonzedwa kuti igwiritsidwe ntchito pamakina wamba okhala ndi GPU imodzi.

Stable Diffusion 2.0 Image Synthesis System Adayambitsidwa


Source: opennet.ru

Kuwonjezera ndemanga