Нейронные сети (картинки)

StableDifusion

Установка

1. Установить гит https://git-scm.com/download/win 2. Установить python 3.10.6 (https://www.python.org/downloads/windows/) c галкой add to PATH 3. Установить CUDA https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=10&target_type=exe_local 4. в git bush (или в VC в терминале) git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui Создаст папку там где запущен баш (C:\Users\Misha\stable-diffusion-webui) git pull — чтобы обновить (из папки stable-diffusion-webui) Или скачать https://github.com/AUTOMATIC1111/stable-diffusion-webui/archive/refs/tags/v1.6.0.zip 5. Запустить webui-user.bat (ждать 15 мин) Если ловим ошибку TypeError: AsyncConnectionPool.__init__() got an unexpected keyword argument ‘socket_options’

source venv/Scripts/activate
pip install httpx==0.24.1

Для автоматического обновления — дописать в код webui-user.bat

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --autolaunch --xformers --medvram --opt-sdp-attention
git pull

call webui.bat

Если xformers не поставился

set COMMANDLINE_ARGS= --reinstall-xformers --xformers

xformers нужен для снижения нагрузки на видеокарту, работает на архитектурах: turing, hopper, pascal, ampere, ada lovelace 6. Ставим свои модели в C:\StableDifusion\stable-diffusion-webui-1.6.0\models\Stable-diffusion 7. Чтобы поменять значения при запуске — редактировать ui-config.json

Автономная версия (xrist)

Не работают XL модели https://github.com/serpotapov/stable-diffusion-portable разархивировать — webui-user-first-run.cmd Модели ставить сюда C:\stable-diffusion-portable-main\models\Stable-diffusion\ Запускать тут webui-user.bat

Дополнительные модели

https://civitai.com/models https://huggingface.co/XpucT/Deliberate/tree/main mj — в начале промта чтобы эмулировать стиль миджорни cinematic — добавить кинематографичности cozy — добавить уюта https://civitai.com/models/125703/protovision-xl-high-fidelity-3d-photorealism-anime-hyperrealism-no-refiner-needed

Локальный сервер

Отредактировать webui.py def webui():

app, local_url, share_url = shared.demo.launch(
            #share=cmd_opts.share,
            share=True,
            server_name=server_name,
            server_port=cmd_opts.port,
            ssl_keyfile=cmd_opts.tls_keyfile,
            ssl_certfile=cmd_opts.tls_certfile,
            debug=cmd_opts.gradio_debug,
            auth=[tuple(cred.split(':')) for cred in gradio_auth_creds] if gradio_auth_creds else None,
            inbrowser=cmd_opts.autolaunch,
            prevent_thread_lock=True
        )

Обучение на LORA

https://github.com/serpotapov/Kohya_ss-GUI-LoRA-Portable Для каждой версии семплера — отдельная лора (1.5; XL) Extra networks — Add hypernetwork to promt — none (для обучений) Путь куда класть фотки (примерно 16) Тренировать на 2й вкладке — dreamboots Lora C:_StableDifusion\Train\100_Misha В настройках указать папку в которой лежит папка 100_Misha При начале трейна сразу удалять появившуюся папку smple

Дополнительные аддоны

На вкладке Extentions Ставятся в C:\_StableDifusion\stable-diffusion-portable-main\extensions\ dynamic thresholding — https://github.com/mcmonkeyprojects/sd-dynamic-thresholding ControlNet В Extension: sd-webui-controlnet manipulations Сами модели https://huggingface.co/lllyasviel/ControlNet/tree/main/models Ставить сюда C:_StableDifusion\stable-diffusion-portable-main\extensions\sd-webui-controlnet\models canny — контуры mlsd — интерьеры openpose — копирование поз scribble — раскраска скетча Scribble/Sketch — по скетчу генерирует изображения

Про CONTROLNET

https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main

Промты

Кот в шапке санты cat in santa hat, (((ink))), ((((grayscale)))), Концепт для магической книги рецептов Full page concept design how to craft life Poison, intricate details,infographic of alchemical, diagram of how to make potions, captions, directions, ingredients, drawing , magic,wuxia Для артов a highly detailed epic cinematic concept art an alien pyramid landscape , art station, landscape, concept art, illustration, highly detailed artwork cinematic, hyper realistic painting Декор environment living room interior, mid century modern, indoor garden with fountain, retro,m vintage, designer furniture made of wood and plastic, concrete table, wood walls, indoor potted tree, large window, outdoor forest landscape, beautiful sunset, cinematic, concept art, sunstainable architecture, octane render, utopia, ethereal, cinematic light, –ar 16:9 –stylize 45000 Открытка beautiful fashion elegant goddness of water, chic strapless dress, tropical sea background, character design, in the style of artgerm, and wlop, chanel jewelry, cinematic lighting, hyperdetailed, 8 k realistic, symmetrical, global illumination, radiant light, love and mercy, frostbite 3 engine, cryengine, dof, trending on artstation, digital art, crepuscular ray Готичный лев ultra wide angle, in focus, dark epic background, gorgeous lifelike, moody golden fur design, dynamic pose of a lion, the full and whole body, full length shot, hyper details, lighting art, cinematic, insane details, intricate details, hyperdetailed, goth, fractal, dark shot Статуя в музее inside museum, alabaster statue, (made of marble:1.2), antique statue, Greek Goddess of beauty, medium boobs, 1girl, wide hips, nipples, detailed hands, ilya kuvshinov, raymond swanland, james gilleard, very detailed, 8k, hd, full color Девушка на фоне города night,neon lights,cyberpunk,city, cityscape,The views of the buildings are stunning,detailed background,depth of field,boken,\\ (masterpiece), best quality, highest quality, extremely detailed, beautiful face, CG unity 8k wallpaper, detailed and intricate, original,highres,\\ nsfw,show nipples,A fashionably dressed girl wearing sunglasses, jacket,jeans,(Stand in a very high and dangerous place,Extreme selfie:1.5),a bird’s-eye view,(tattoo:1.2),(thrilled expressions,excited expressions),black hair,// , girls focus, solo, from above, standing,// realistic:1.3, (ambient light:1.3),(cinematic composition:1.3),(neon lights:0.1),(HDR:1),Accent Lighting, pureerosface_v1:1 Деревня sharp details, a medieval village in switzerland, mess jungle in background, ornate, beautiful, atmosphere, vibe, flowers, concept art illustration, greg rutowski, volumetric lighting, sunbeams, particles, colorful clothes, by Jean-Baptiste Monge, Gilles Beloeil, Tyler Edlin, Marek Okon, Pixar, album art, comic style, golden ratio, perfect composition, a masterpiece, trending on artstation, oversaturated, epic realistic, hdr, intricate details, rutkowski, intricate, cinematic, detailed Портрет под арт (близко к аниме 3д) highly detailed portrait of halo, blue eyes, tartan scarf, white hair by atey ghailan, by greg rutkowski, by greg tocchini, by james gilleard, by joe fenton, by kaethe butcher, gradient yellow, black, brown and magenta color scheme, grunge aesthetic!!! graffiti tag wall background Портрет под арт a 21 years old (european:1.3) devil girl with wild black hair with red braids| centered| key visual| intricate| highly detailed| breathtaking beauty| precise lineart| vibrant| comprehensive cinematic| Carne Griffiths| Conrad Roset Портрет под арт 2 Ultra-detailed woman, (looking thoughtfully into the distance:1.4), in the park, sitting on a bench, full body, outdoors, (not looking at the camera:1.7), (adjusting hair:1.5), (short open shirt:1.2), solo, beautiful detailed sky, (closed mouth), small breasts, beautiful detailed eyes, floating hair, NSFW, seminude bosom, saliva, style-empire, by Alex Huguet, Mike Hill, Ian Spriggs, JaeCheol Park, Marek Denko, picture painted in impressionist style, brushstroke painting technique, palette knife painting, an expressive oil painting, Leonid Afremov Портрет залитый краской (level difference:1.8),(Paint colliding and splashing on the canvas),(depth of field),1girl’s side face blends into it,((side face)),open mouth,(liquid paint rainbow hair:1.1) made of paint and defies gravity,thick flowing,(paint splatter:1.3),Liquid state,stunningly beautiful, masterpiece, detailed background,ultra high quality model, ethereal background,abstract beauty, explosive volumetric, oil painting,heavy strokes,Romantic lighting,Sub-Surface Scatterring,lens 135mm,f1.8,glow,8k,high resolution, dreamy,ray tracing,hdr,god rays Портрет под векторный арт vector art, artwork by Ilya Kuvshinov, artwork by WLOP, detail vector art, loose brush effect, luminance, warm bright pastel colors, perfect hips, Riuko Matoi stretching, luminous, warm copper hair, pale, blue eyes, vector art, hazed out, artwork by Ilya Kuvshinov, bright orange-pink car show background Портрет под векторный арт 2 (+ clip skip = 2) masterpiece, best quality, ( art by acrylic pouring | warm bright pastel colors | 2d minialist vector illustration |), 1gril, beautiful woman, smile, portrait, bright background Портрет под векторный арт 3 (+ clip skip = 2) masterpiece, best quality, (2d minimalist vector illustration), 1girl, beautiful woman, closed eyes, colorful, (Gustav Klimt blends blue color long hair), (Wassily Kandinsky blends dress:1.2), (acrylic pour art blends background:1.4), (bust shot:1.4) Портрет мужчины под арт portrait of the man, stunning, highly detailed, 8k, ornate, intricate, cinematic, dehazed, atmospheric, (oil painting:0.75), (splash art:0.75),(teal:0.2),(orange:0.2), (by Jeremy Mann:0.5), (by John Constable:0.1),(by El Greco:0.5),(acrylic paint:0.75) Стикеры (+ ) robocop ((stickers)) retro (high quality) Чернила (+ ) The eagle stood on the branch, traditional chinese ink painting Карта тарро 70s style poster art, cosmic galactic style art, Tarot card: «III — The Empress» Gorgeous poster art textured and aesthetic Детский рисунок Jinx fanart, drawn by a 5-year-old child with colored cranyon Двойная экспозиция (пейзаж в персонаже) Double exposure, Midjourney style, merging, blending, overlay double exposure image, Double Exposure style, An exceptional masterpiece by Yukisakura revealing a fantastic double exposure composition of Naruto body silhouette harmoniously intertwined with the visually striking details of the hidden leaf village towards a sunset over the mountains. While the sunset and mountains details are outwardly echoing through the fabric of the figure, beautiful tension climbs as the contrasting use of monochrome in the background maintains razor-sharp focus on the remarkable double exposure image. Characterized by its vibrant full-color scheme and crisp lines that shadow the contours just as meticulously. (Detailed:1.45). (Detailed background:1.4). Экшн фигурка Product photography: Plastic full-body action figure of Tom Cruise as Major William Cage from Edge of Tomorrow. The figure stands upright in full exo-suit armor, battle-worn and detailed with grime, scratches, and bullet marks. Face sculpt features a determined expression mid-loop, with tousled helmet hair and a subtle five o’clock shadow. “Live. Die. Repeat. Collection” «Major William Cage Action Figure» – matte black packaging with electric blue typography and yellow hazard stripes. The box looks slightly scorched, like it’s survived one too many dropship landings. The figure and accessories are sealed in heavy-duty molded plastic. 3 accessories on the left, 1 plastic Cage action figure on the right. Accessories: Pulse Rifle Grenade Belt Battle Helmet (in a red spikey bubble below the accessories): «Now with Built-In Voice Module!» : “I think there’s something wrong with your suit. There’s a dead guy in it.” «Did you just take my battery pack?» «I wish I didn’t know you, but I do.» Captured using a Sony mirrorless camera with a 70-200mm f2.8 lens, cinematic lighting for a gritty sci-fi vibe, showing micro-detail in the suit’s hydraulic pistons and weathering. Background: steel-gray with smoky atmosphere, mimicking the battlefield of Verdun.

Негативные промты

1. lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name 2. bad-artist, bad-hands-5, ugly, duplicate, censored, morbid, mutiliated, extra digits, bad hands, mutiliated hands, poorly drawn face, poorly drawn hands, mutation, deformed, blurry, bad anatomy, extra limbs, cloned face, gross proportions, malformed, missing arms, missing legs, mutated hands, low quality, lowres, normal quality, worst quality, standard quality, watermark, jpeg artifacts, weird eyes, plain background 3. easynegative, (worst quality, low quality:1.4), watermark, logo, bad anatomy, extra fingers, extra hands, body hair, mosaic 4. (worst quality, low quality, normal quality, lowres, low details, oversaturated, undersaturated, overexposed, underexposed, grayscale, bw, bad photo, bad photography, bad art:1.4), (watermark, signature, text font, username, error, logo, words, letters, digits, autograph, trademark, name:1.2), (blur, blurry, grainy), morbid, ugly, asymmetrical, mutated malformed, mutilated, poorly lit, bad shadow, draft, cropped, out of frame, cut off, censored, jpeg artifacts, out of focus, glitch, duplicate, (airbrushed, cartoon, anime, semi-realistic, cgi, render, blender, digital art, manga, amateur:1.3), (3D ,3D Game, 3D Game Scene, 3D Character:1.1), (bad hands, bad anatomy, bad body, bad face, bad teeth, bad arms, bad legs, deformities:1.3)

Настройки

Семплер: DPM++ 2M Karras, Euler a (не естественная кожа у людей), UniPC (новый, потестить) Restore faces — снять галку GFG = 7 (4-12) (чем выше, тем больше деталей дорисовывается, большие значения — пережжённое и пластиковое изображение, на фотографиях ставить 4.5) Sampling steps = 28 Mimic CFG Scale — 6 (CFG Scale = 12, Clip skip = 1) Если без Mimic CFG — Clip skip = 2 User interface — Quicksettings list — sd_model_checkpoint,sd_vae,CLIP_stop_at_last_layers Face restoration (лучше отключать) Code former = 0.6 live preview = 10 Multiplier for extra networks = 0.7 Vae https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main

COMFYUI

https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file Обновить C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\update\update_comfyui.bat C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\update\update_comfyui_and_python_dependencies.bat Запуск C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\run_nvidia_gpu.bat Модели класть в: C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\models\checkpoints Чтобы ссылаться на модели: (Для скорости — хранить все модели на ссд) https://schinagl.priv.at/nt/hardlinkshellext/HardLinkShellExt_X64.exe ПКМ — Запомнить источник ссылки ПКМ — Поместить как… — Символическую ссылку Включить режим разработчика (чтобы был доступен save API): Шестерня — Enable Dev mode Options Чтобы «изменения» в нодах действовали сразу: Шестерня — Widget Value Control Mode — before Менеджер: https://github.com/ltdrdata/ComfyUI-Manager C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\custom_nodes в адресной строке — cmd git clone https://github.com/ltdrdata/ComfyUI-Manager.git Чтобы видеть какие ноды свои: Lite Graph - Node source badge mode Lite Graph - Node ID badge mode Мониторинг ресурсов: Crystools из менеджера или https://github.com/crystian/ComfyUI-Crystools Clone this repo into custom_modules: cd ComfyUI/custom_nodes git clone https://github.com/crystian/ComfyUI-Crystools.git cd ComfyUI-Crystools pip install -r requirements.txt Горячие клавиши: 2 ЛКМ — поиск по вставке ноды CTRL + B — деактивировать/активировать ноду CTRL + стрелка вверх\вниз — усилить\ослабить промт CTRL + SHIFT + V — вставить с добавленными связями Использовать 2 GPU: https://github.com/neuratech-ai/ComfyUI-MultiGPU C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\custom_nodes>git clone https://github.com/neuratech-ai/ComfyUI-MultiGPU.git Переключить на 2ю GPU: C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\cli_args.py parser.add_argument(«—cuda-device», type=int, default=1, metavar=»DEVICE_ID», help=»Set the id of the cuda device this instance will use.») Выражения лица онлайн: https://github.com/kijai/ComfyUI-LivePortraitKJ/tree/main/examples Controlnet: ComfyUI-Advanced-ControlNet ComfyUI’s ControlNet Auxiliary Preprocessors Скачать модели: https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main Поместить модели: ComfyUI\models\controlnet Для XL: lineart — https://huggingface.co/promeai/sdxl-controlnet-lineart-promeai/tree/main canny — https://huggingface.co/diffusers/controlnet-canny-sdxl-1.0/tree/main Скачать diffusion_pytorch_model.safetensors и переименовать в controlnet-canny-sdxl-1.0.safetensors Лоры: https://huggingface.co/stabilityai/control-lora/tree/main/control-LoRAs-rank256 IPadapter: ComfyUI_IPAdapter_plus Скачать модели: https://github.com/cubiq/ComfyUI_IPAdapter_plus/ Поместить модели: /ComfyUI/models/clip_vision /ComfyUI/models/ipadapter Blip: Скачать — https://huggingface.co/Rookiehan/BLIP/tree/main положить в ComfyUI\models\BLIP Обновить torch (если какой-то мод всё сломал): Из папки ComfyUI_windows_portable\python_embeded .\python.exe .\Scripts\pip.exe install —upgrade —force-reinstall torch torchvision torchaudio —index-url https://download.pytorch.org/whl/cu121 Florence (картинка в текст): https://huggingface.co/microsoft/Florence-2-base Ставить в C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\models\LLM git clone https://huggingface.co/microsoft/Florence-2-base Из https://huggingface.co/microsoft/Florence-2-base/tree/main скачиваем pytorch_model.bin и кладём в Florence-2-base Janus-Pro: В менеджере ComfyUI-Janus-Pro C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\custom_nodes\janus-pro -> запустить cmd -> pip install -r requirements.txt Скачать модель (7B) https://huggingface.co/deepseek-ai/Janus-Pro-7B/tree/main переместить всё в папку C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\models\Janus-Pro\Janus-Pro-7B

Flux

Для clip (декодирование промта) https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main clip в C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\models\clip Модели https://huggingface.co/Kijai/flux-fp8/tree/main Кладём модель в C:\ConfUI\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\models\unet Для schnell оптимально 4 шага https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main Для dev оптимально 30 шагов https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main Vae https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main/vae https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main/vae Сжатая модель https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main Энкодер (clip) для сжатой модели https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/tree/main

Полезные ссылки

Текст обрамлённой картинкой https://ideogram.ai/ Скрытый в картинке текст https://replicable.art/learn/generate-images-with-hidden-text-using-stable-diffusion-and-controlnet Аналог chatGPT (US and UK) https://claude.ai/ Растровую в векторную графику https://vectorizer.ai/ Генератор скайбокса https://skybox.blockadelabs.com/ Варианты одного изображения https://imagevariations.com/ Генератор музыки https://app.songr.ai/ Топ-новости отобранные ии https://www.newsminimalist.com/ Топовые модели для сравнения https://yupp.ai/

Нарисуй обложку для статей по "Postgresql" как скетч с цветными зарисовками в разрешении 1:1
Текст на обложку не добавляй. 
На белом фоне. Минимальные отступы от изображения по бокам.