Ua hoʻopuka ʻo Stability AI i nā hiʻohiʻona mākaukau no ka ʻōnaehana aʻo mīkini Stable Diffusion, hiki ke hoʻohui a hoʻololi i nā kiʻi e pili ana i kahi wehewehe kikokikona ma ka ʻōlelo kūlohelohe. Ua laikini ʻia nā kumu hoʻohālike ma lalo o ka laikini Creative ML OpenRAIL-M no ka hoʻohana ʻana i ka ʻoihana. No ke aʻo ʻana i ka ʻōnaehana, ua hoʻohana ʻia kahi hui o 4000 NVIDIA A100 Ezra-1 GPU a me kahi hōʻiliʻili LAION-5B, me nā kiʻi 5.85 biliona me nā wehewehe kikokikona. Ma mua, ua wehe ʻia ke code no nā mea hana no ka hoʻomaʻamaʻa ʻana i kahi pūnaewele neural a me ka hana ʻana i nā kiʻi ma lalo o ka laikini MIT.
ʻO ka loaʻa ʻana o kahi hiʻohiʻona mākaukau a me nā koi ʻōnaehana kūpono e hiki ai i kekahi ke hoʻomaka i nā hoʻokolohua ma kahi PC me nā GPU maʻamau ua alakaʻi i ka puka ʻana o kekahi mau papahana pili:
- textual-inversion (code) - he mea hoʻohui e hiki ai iā ʻoe ke hoʻohui i nā kiʻi me kahi ʻano, mea a i ʻole kaila. I loko o ka Stable Diffusion mua, ʻo nā mea i loko o nā kiʻi synthesized he maʻamau a hiki ʻole ke kāohi. Hiki iā ʻoe ke hoʻohui i kāu mau mea ʻike ponoʻī, hoʻopaʻa iā lākou i nā huaʻōlelo a hoʻohana iā lākou i ka synthesis.
No ka laʻana, i ka Stable Diffusion maʻamau hiki iā ʻoe ke noi i ka ʻōnaehana e hana i kahi kiʻi me kahi "pōkole i loko o ka moku". Eia hou, hiki iā ʻoe ke wehewehe i nā hiʻohiʻona o ka pōpoki a me ka waʻa, akā ʻaʻole hiki ke ʻike ʻia ka pōpoki a me ka moku e hoʻohui ʻia. Textual-inversion hiki iā ʻoe ke hoʻomaʻamaʻa i ka ʻōnaehana ma kahi kiʻi o kāu pōpoki a i ʻole ka moku a hoʻohui i ke kiʻi me kahi pōpoki a moku paha. Ma keʻano like, hiki iā ia ke hoʻololi i nā mea kiʻi me kekahi mau mea, hoʻonoho i kahi hiʻohiʻona o ke ʻano hiʻohiʻona no ka synthesis, a kuhikuhi i nā manaʻo (no ka laʻana, mai nā ʻano kauka holoʻokoʻa, hiki iā ʻoe ke hoʻohana i kahi koho ʻoi aku ka pololei a me ke kiʻekiʻe. i ke ʻano makemake).
- stable-diffusion-animation - ka hana ʻana i nā kiʻi animated (neʻe) ma muli o ka interpolation ma waena o nā kiʻi i hana ʻia ma Stable Diffusion.
- stable_diffusion.openvino (code) - he awa o Stable Diffusion, e hoʻohana wale ana i ka CPU no ka helu ʻana, e ʻae ai i ka hoʻokolohua ma nā ʻōnaehana me ka ʻole o nā GPU ikaika. Pono i ka mea hana i kākoʻo ʻia ma ka waihona OpenVINO. Hāʻawi ʻo OpenVINO i nā plugins no nā kaʻina hana Intel me AVX2, AVX-512, AVX512_BF16 a me SSE hoʻonui, a me Raspberry Pi 4 Model B, Apple Mac mini a me nā papa NVIDIA Jetson Nano. Me ka manaʻo ʻole, hiki ke hoʻohana iā OpenVINO ma nā kaʻina hana AMD Ryzen.
- ʻO sdamd kahi awa no nā AMD GPU.
- ʻO kahi hoʻokō mua o ka synthesis wikiō.
- stable-diffusion-gui, stable-diffusion-ui, Artbreeder Collage, diffuse-the-rest - nā loulou kiʻi no ka hana ʻana i nā kiʻi me ka Stable Diffusion.
- beta.dreamstudio.ai, Hugging Face Spaces, hlky Stable Diffusion WebUI - nā pilina pūnaewele no ka hoʻopili kiʻi me ka Stable Diffusion.
- Nā Plugins no ka hoʻohui ʻana i ka Stable Diffusion me GIMP, Figma, Blender a me Photoshop.
Eia hou, hiki iā mākou ke hoʻomaopopo i ka paʻi ʻia ʻana e Google o ke code o ka ʻōnaehana aʻo mīkini RawNeRF (RAW Neural Radiance Fields), e hiki ai, ma muli o ka ʻikepili mai kekahi mau kiʻi RAW, e hoʻomaikaʻi i ka maikaʻi o nā kiʻi walaʻau nui i lawe ʻia i ka pōʻeleʻele a i loko. ʻino ka mālamalama. Ma kahi o ka hoʻopau ʻana i ka walaʻau, nā mea hana i kūkulu ʻia e ka papahana e hiki ai ke hoʻonui i ka kikoʻī, hoʻopau i ka glare, synthesize HDR a hoʻololi i ke kukui holoʻokoʻa i nā kiʻi, a me ka hana hou ʻana i ke kūlana ʻekolu-dimensional o nā mea me ka hoʻohana ʻana i nā kiʻi mai nā kihi like ʻole. hoʻololi i ka manaʻo, hoʻopololei i ka manaʻo a hana i nā kiʻi neʻe.
Source: opennet.ru