Stability AI ebipụtala mbipụta nke abụọ nke sistemu mmụta igwe Stable Diffusion, nke nwere ike ịmekọrịta na gbanwee onyonyo dabere na ụkpụrụ atụpụtara ma ọ bụ nkọwa ederede asụsụ okike. Edere koodu ngwaọrụ maka ọzụzụ netwọkụ akwara ozi na ọgbọ onyonyo na Python site na iji usoro PyTorch wee bipụta ya n'okpuru ikike MIT. Ụdị a zụrụ azụ emeghewo n'okpuru ikikere ikike Creative ML OpenRAIL-M, nke na-enye ohere iji azụmahịa. Ọzọkwa, ihe ngosi ihe onyonyo ntanetị dị.
Nkwalite isi na mbipụta ọhụrụ Stable Diffusion:
- A ọhụrụ nlereanya maka oyiyi njikọ dabeere na ederede nkọwa - SD2.0-v - e kere, nke na-akwado ọgbọ nke oyiyi na a mkpebi nke 768×768. A zụrụ ụdị ọhụrụ a site na iji nchịkọta LAION-5B nke onyonyo ijeri 5.85 nwere nkọwa ederede. Ihe nlereanya ahụ na-eji otu usoro ihe atụ dị ka ihe atụ Stable Diffusion 1.5, mana ọ dị iche site na ntughari iji koodu OpenCLIP-ViT/H dị iche, nke mere ka o kwe omume imeziwanye ogo nke onyonyo apụta.
- A kwadebere ụdị SD2.0 dị mfe, zụọ ya na onyonyo 256 × 256 site na iji ụdị amụma amụma oge gboo yana ọgbọ onyonyo na-akwado ya na mkpebi 512 × 512.
- A na-enye ohere nke iji teknụzụ nke supersampling (Super Resolution) iji mee ka mkpebi nke ihe oyiyi mbụ dịkwuo elu na-enweghị ibelata àgwà, na-eji algọridim maka nhazi oghere na nhazigharị nkọwa. Ụdị nhazi ihe onyonyo enyere (SD20-upscaler) na-akwado 2048x upscaling, nke nwere ike iwepụta onyonyo nwere mkpebi 2048 × XNUMX.
- A na-atụpụta ụdị SD2.0-depth2img, nke na-eburu n'uche omimi na nhazi oghere nke ihe. A na-eji usoro MiDaS maka nleba anya omimi nke monocular. Ụdị ahụ na-enye gị ohere ịmepụta ihe oyiyi ọhụrụ site na iji ihe oyiyi ọzọ dị ka template, nke nwere ike ịdị iche na nke mbụ, ma na-ejigide ihe mejupụtara ya na omimi. Dịka ọmụmaatụ, ịnwere ike iji pose nke mmadụ na foto iji mepụta agwa ọzọ n'otu ọkwa ahụ.
- Emelitela ụdị maka ịgbanwe onyonyo - SD 2.0-inpainting, nke na-enye gị ohere iji ederede ederede dochie ma gbanwee akụkụ nke onyonyo.
- Edozila ụdịdị maka ojiji na sistemụ nwere otu GPU.
isi: opennet.ru