Stability AI ebipụtala ụdị ndị emebere maka sistemụ mmụta igwe Stable Diffusion, nke nwere ike ịmekọrịta na gbanwee onyonyo dabere na nkọwa ederede n'asụsụ eke. Enyere ụdị ikike ikike n'okpuru ikikere Creative ML OpenRAIL-M maka ojiji azụmahịa. Iji zụọ usoro ahụ, a na-eji ụyọkọ 4000 NVIDIA A100 Ezra-1 GPUs na nchịkọta LAION-5B, gụnyere ihe oyiyi ijeri 5.85 nwere nkọwa ederede. Na mbụ, koodu maka ngwaọrụ maka ịzụ netwọkụ akwara ozi na imepụta onyonyo bụ nke mepere emepe n'okpuru ikike MIT.
Ịnweta ihe nlereanya akwadoro yana chọrọ usoro dị mma nke na-enye mmadụ ohere ịmalite nnwale na PC nwere GPU ọkọlọtọ emeela ka ọ pụtapụta ọtụtụ ọrụ metụtara ya:
- ederede-inversion (koodu) - mgbakwunye na-enye gị ohere iji agwa, ihe ma ọ bụ ụdị nyere gị mepụta ihe oyiyi. N'ime Stable Diffusion mbụ, ihe ndị dị na onyonyo agbakwunyere bụ enweghị usoro na enweghị njikwa. Ihe mgbakwunye a na-atụ aro na-enye gị ohere ịgbakwunye ihe a na-ahụ anya nke gị, jikọta ha na mkpụrụokwu ma jiri ha mee ihe na njikọ.
Dịka ọmụmaatụ, na Stable Diffusion mgbe niile ị nwere ike ịrịọ usoro ahụ ka ọ mepụta ihe oyiyi na "cat na ụgbọ mmiri". Na mgbakwunye, ị nwere ike dokwuo anya njirimara nke pusi na ụgbọ mmiri, ma ọ bụ ihe a na-atụghị anya ya nke a ga-ejikọta pusi na ụgbọ mmiri. Ntugharị ederede na-enye gị ohere ịzụ usoro ahụ na ihe oyiyi nke pusi ma ọ bụ ụgbọ mmiri gị ma jiri otu nwamba ma ọ bụ ụgbọ mmiri megharịa ihe oyiyi ahụ. N'otu aka ahụ, ọ nwekwara ike dochie ihe oyiyi na ihe ụfọdụ, setịpụ ihe atụ nke ụdị a na-ahụ anya maka njikọ, ma kọwaa echiche (dịka ọmụmaatụ, site na ndị dọkịta dị iche iche, ị nwere ike iji nhọrọ ziri ezi na nke dị elu). n'ụdị achọrọ).
- stable-diffusion-animation - imepụta ihe oyiyi animated (na-akpụ akpụ) dabere na ntinye n'etiti foto ewepụtara na Stable Diffusion.
- stable_diffusion.openvino (koodu) - ọdụ ụgbọ mmiri Stable Diffusion, nke na-eji naanị CPU maka mgbako, nke na-enye ohere nnwale na sistemụ na-enweghị GPU dị ike. Chọrọ ihe nhazi akwadoro n'ọbá akwụkwọ OpenVINO. N'ihu ọha, OpenVINO na-enye plugins maka ndị nrụpụta Intel nwere AVX2, AVX-512, AVX512_BF16 na mgbakwunye SSE, yana maka Raspberry Pi 4 Model B, Apple Mac mini na bọọdụ NVIDIA Jetson Nano. Na-akwadoghị, ọ ga-ekwe omume iji OpenVINO na ndị nrụpụta AMD Ryzen.
- sdamd bụ ọdụ ụgbọ mmiri maka AMD GPUs.
- Mmejuputa mmalite nke njikọ vidiyo.
- stable-diffusion-gui, stable-diffusion-ui, Artbreeder Collage, difffuse-the-rest - interfaces eserese maka imepụta onyonyo site na iji Stable Diffusion.
- beta.dreamstudio.ai, Mmakọ Ihu Oghere, hlky Stable Diffusion WebUI - webụsaịtị maka njikọ onyonyo site na iji Stable Diffusion.
- Ngwa mgbakwunye maka ijikọ Stable Diffusion na GIMP, Figma, Blender na Photoshop.
Tụkwasị na nke a, anyị nwere ike ịhụ akwụkwọ Google bipụtara nke koodu nke RawNeRF (RAW Neural Radiance Fields) usoro mmụta igwe, nke na-enye ohere, dabere na data sitere na ọtụtụ ihe oyiyi RAW, iji melite ogo nke ihe oyiyi na-eme mkpọtụ nke a na-ewere na ọchịchịrị na n'ime. ọkụ na-adịghị mma. Na mgbakwunye na ikpochapụ mkpọtụ, ngwá ọrụ ndị ọrụ ahụ mepụtara na-eme ka o kwe omume ịbawanye nkọwa, kpochapụ ọkụ, synthesize HDR ma gbanwee ọkụ zuru ezu na foto, yana ịmegharị ọnọdụ akụkụ atọ nke ihe site na iji ọtụtụ foto site n'akụkụ dị iche iche. gbanwee echiche ahụ, gbanwee nlebara anya ma mepụta foto na-akpali akpali.
isi: opennet.ru