Stability AI yasindikiza zitsanzo zokonzeka za makina ophunzirira makina a Stable Diffusion, omwe amatha kupanga ndi kusintha zithunzi pogwiritsa ntchito kufotokozera malemba m'chinenero chachibadwa. Ma Model ali ndi chilolezo pansi pa chilolezo cha Creative ML OpenRAIL-M chololeza kuti agwiritse ntchito malonda. Kuti aphunzitse dongosololi, gulu la 4000 NVIDIA A100 Ezra-1 GPUs ndi gulu la LAION-5B, kuphatikizapo zithunzi za 5.85 biliyoni zokhala ndi malemba, zinagwiritsidwa ntchito. M'mbuyomu, zida zophunzitsira ma neural network ndikupanga zithunzi zinali zotseguka pansi pa layisensi ya MIT.
Kupezeka kwa mtundu wopangidwa kale komanso zofunikira zamakina ocheperako zomwe zimalola kuti munthu ayambe kuyesa pa PC yokhala ndi ma GPU okhazikika apangitsa kuti mapulojekiti angapo okhudzana nawo awonekere:
- textual-inversion (code) - chowonjezera chomwe chimakupatsani mwayi wopanga zithunzi ndi mawonekedwe, chinthu kapena mawonekedwe. Mu Stable Diffusion yoyambirira, zinthu zomwe zili muzithunzi zopangidwa mwachisawawa komanso zosalamulirika. Zowonjezera zomwe zaperekedwa zimakupatsani mwayi wowonjezera zinthu zanu zowoneka, kuzimanga ndi mawu osakira ndikuzigwiritsa ntchito pophatikiza.
Mwachitsanzo, nthawi zonse Stable Diffusion mukhoza kufunsa dongosolo kuti apange fano ndi "mphaka mu boti". Kuphatikiza apo, mutha kufotokozera bwino za mphaka ndi bwato, koma sizikudziwika kuti mphaka ndi boti zidzapangidwa. Textual-inversion imakupatsani mwayi wophunzitsa dongosolo pa chithunzi cha mphaka kapena bwato lanu ndikuphatikiza chithunzicho ndi mphaka kapena bwato linalake. Momwemonso, imathanso kusintha mawonekedwe azithunzi ndi zinthu zina, kukhala chitsanzo cha kalembedwe kawonekedwe kaphatikizidwe, ndikufotokozera malingaliro (mwachitsanzo, kuchokera kwa madokotala osiyanasiyana osiyanasiyana, mutha kugwiritsa ntchito kusankha kolondola komanso kwapamwamba kwambiri. m'njira yomwe mukufuna).
- stable-diffusion-animation - kupanga zithunzi zamakanema (zosuntha) kutengera kutanthauzira pakati pa zithunzi zopangidwa mu Stable Diffusion.
- stable_diffusion.openvino (code) - doko la Stable Diffusion, lomwe limagwiritsa ntchito CPU yokha yowerengera, yomwe imalola kuyesera pa machitidwe opanda ma GPU amphamvu. Imafunika purosesa yothandizidwa mulaibulale ya OpenVINO. Mwalamulo, OpenVINO imapereka mapulagini a Intel processors okhala ndi AVX2, AVX-512, AVX512_BF16 ndi SSE extensions, komanso Raspberry Pi 4 Model B, Apple Mac mini ndi NVIDIA Jetson Nano board. Mosavomerezeka, ndizotheka kugwiritsa ntchito OpenVINO pa mapurosesa a AMD Ryzen.
- sdamd ndi doko la AMD GPUs.
- Kukhazikitsa koyamba kwa kaphatikizidwe kakanema.
- stable-diffusion-gui, stable-diffusion-ui, Artbreeder Collage, diffuse-the-rest - zojambulajambula zopangira zithunzi pogwiritsa ntchito Stable Diffusion.
- beta.dreamstudio.ai, Mipata Yankhope Yokumbatira, hlky Stable Diffusion WebUI - malo olumikizirana ndi mawebusayiti opangira zithunzi pogwiritsa ntchito Stable Diffusion.
- Mapulagini ophatikizira Stable Diffusion ndi GIMP, Figma, Blender ndi Photoshop.
Kuonjezera apo, tikhoza kuzindikira kufalitsidwa ndi Google kwa code ya makina ophunzirira makina a RawNeRF (RAW Neural Radiance Fields), omwe amalola, kutengera deta kuchokera ku zithunzi zingapo za RAW, kupititsa patsogolo zithunzi zaphokoso kwambiri zomwe zimatengedwa mumdima komanso mumdima. kuwala koyipa. Kuphatikiza pa kuthetsa phokoso, zida zomwe polojekitiyi imapanga zimapangitsa kuti ziwonjezere tsatanetsatane, kuthetsa kunyezimira, kupanga HDR ndikusintha kuunikira kwazithunzi zonse, komanso kukonzanso malo atatu azinthu pogwiritsa ntchito zithunzi zingapo kuchokera kumbali zosiyanasiyana, sinthani malingaliro, sinthani malingaliro ndikupanga zithunzi zoyenda.
Source: opennet.ru