自動提示LLM-Text與LLM-Vision的應用指南

Quick Links

• Auto prompt by LLM and LLM-Vision (Trigger more details out inside model)
• • SD-WEB-UI: https://github.com/xlinx/sd-webui-decadetw-auto-prompt-llm
  • ComfyUI: https://github.com/xlinx/ComfyUI-decadetw-auto-prompt-llm
• Auto msg to ur mobile (LINE | Telegram | Discord)
• • SD-WEB-UI :https://github.com/xlinx/sd-webui-decadetw-auto-messaging-realtime
  • ComfyUI: https://github.com/xlinx/ComfyUI-decadetw-auto-messaging-realtime
• I'm SD-VJ. (share SD-generating-process in realtime by gpu)
• • SD-WEB-UI: https://github.com/xlinx/sd-webui-decadetw-spout-syphon-im-vj
  • ComfyUI: https://github.com/xlinx/ComfyUI-decadetw-spout-syphon-im-vj
• CivitAI Info|discuss:
• • https://civitai.com/articles/6988/extornode-using-llm-trigger-more-detail-that-u-never-thought
  • https://civitai.com/articles/6989/extornode-sd-image-auto-msg-to-u-mobile-realtime
  • https://civitai.com/articles/7090/share-sd-img-to-3rd-software-gpu-share-memory-realtime-spout-or-syphon

SD-WEB-UI | ComfyUI | decadetw-Auto-Prompt-LLM-Vision

Update Log

• [add|20240730] | 🟢 LLM Recursive Prompt
• [add|20240730] | 🟢 Keep ur prompt ahead each request
• [add|20240731] | 🟢 LLM Vision
• [add|20240803] | 🟢 translateFunction
• • When LLM answered, use LLM translate result to your favorite language.ex: Chinese. It's just for your reference, which won't affect SD.
• [add|20240808] | 🟠 Before and After script | exe-command
• [add|20240808] | 🟠 release LLM VRAM everytimes

Motivation💡

• Call LLM : auto prompt for batch generate images
• Call LLM-Vision: auto prompt for batch generate images
• Image will get more details that u never though before.
• prompt detail is important

Usage

LLM-Text

• batch image generate with LLM
• • a story
• Using Recursive prompt say a story with image generate
• Using LLM
• • when generate forever modeexample as follows figure Red-box.just tell LLM who, when or whatLLM will take care details.
  • when a story-board mode (You can generate serial image follow a story by LLM context.)its like comic booka superstar on stageshe is singingpeople give her flowera fashion men is walking.

LLM-Vision 👀

• batch image generate with LLM-Vision
• • let LLM-Vision see a magazine
  • see series of image
  • see last-one-img for next-image
  • make a serious of image like comic

Before and After script

• support load script or exe-command Before-LLM and After-LLM
• javascript fetch POST method (install Yourself )
• • security issue, but u can consider as follows
  • https://github.com/pmcculler/sd-dynamic-javascript
  • https://github.com/ThereforeGames/unprompted
  • https://github.com/adieyal/sd-dynamic-prompts
  • https://en.wikipedia.org/wiki/Server-side_request_forgery
  • and Command Line Arg --allow-code

[🟢] stable-diffusion-webui-AUTOMATIC1111[🟢] stable-diffusion-webui-forge[🟢] ComfyUI1. SD-Prompt ✦1girl2.1 LLM-Text ✦2.2 LLM-Vision ✦a super star on stage.Who is she in image?2.3 LLM-Text-sys-prompt ✦2.4 LLM-Vision-sys-prompt ✦You are an AI prompt word engineer. Use the provided keywords to create a beautiful composition. Only the prompt words are needed, not your feelings. Customize the style, scene, decoration, etc., and be as detailed as possible without endings.You are an AI prompt word engineer. Use the provided image to create a beautiful composition. Only the prompt words are needed, not your feelings. Customize the style, scene, decoration, etc., and be as detailed as possible without endings.3. LLM will answer other detail ✦The superstar, with their hair flowing in the wind, stands on the stage. The lights dance around them, creating a magical moment that fills everyone present with awe. Their eyes shine bright, as if they are ready to take on the world.The superstar stands tall in their sparkling costume, surrounded by fans who chant and cheer their name. The lights shine down on them, making their hair shine like silver. The crowd is electric, every muscle tense, waiting for the superstar to perform4. Main Interface | sd-web-ui | ComfyUIComfyUI Manager | search keyword: auto

Usage

InputOutputLLM-Text: a superstar on stage.LLM-Vision: What's pose in this image?.(okay, its cool.)LLM-Text: a superstar on stage.LLM-Vision: with a zebra image(okie, cooool show dress. At least we don't have half zebra half human.)LLM-Text: a superstar on stage.(okay, its cool.)LLM: a superstar on stage.(Wow... the describe of light is great.)LLM: a superstar on stage.(hnn... funny, it does make sense.)CHALLENGELLM-vision:A Snow White girl walk in forest.(detect ur LLM-Vision Model IQ; if u didnt get white dress and lot of snow.... plz let me know model name)SD model: Flux.1 DLLM model: llava-llama-3.1-8bLLM model: Eris_PrimeV4-Vision-32k-7B-IQ3_XXS FLUX modelhnn...NSFW show. I'm not mean that, but not a wrong answer.(Trigger more details; that u never thought about it.)SD model: Flux.1 DLLM model: llava-llama-3.1-8bLLM model: Eris_PrimeV4-Vision-32k-7B-IQ3_XXSadvanced use | before-after-actionin fact, u can run any u want script | (storyboard) | random read line from txt send into LLMSpecial LLM LoopConnect 1st LLM-Text output to 2nd LLM-Text Input Special LLM Loop - keep each feature assign to different obj not mix it on one.LLM-Text output ask looply : here     [new tool 20240915] Civitai Prompt Grabberquick prompt from civitai. u can pick some prompt from another area model(ex indoor design or building model) with ur 1girl, ex: 1girl(up figure)) + in-door design model-prompt. then u will get full detail in background(bottom figure) : https://civitai.com/models/85691this is good present in FLUX model. trigger more detail in background. Make the photo getting more realistic feelingoption1. just quick append prompt from other model from civitai oroption2. of course u can send it into LLM too. [update] LLM-ask-LLM🌀[support] Cloud Service: Gemini Procloud service: https://generativelanguage.googleapis.com/v1model: gemini-1.5-flash (vision)support text and visionit will get more🌀 and more🌀 and more🌀 like....(bottom to top)

Usage Tips

• tips1:
• • leave only 1 or fewer keyword(deep inside CLIP encode) for SD-Prompt, others just fitting into LLM
  • SD-Prompt: 1girl, [xxx,]<--(the keyword u use usually, u got usually image)
  • LLM-Prompt: xxx, yyy, zzz, <--(move it to here; trigger more detail that u never though.)
• tips2:
• • leave only 1 or fewer keyword(deep inside CLIP encode) for SD-Prompt, others just fit into LLM
  • SD-Prompt: 1girl,
  • LLM-Prompt: a superstar on stage. <--(say a story)
• tips3:
• • action script - Beforerandom/series pick prompt txt file random line fit into LLM-Text [read_random_line.bat]random/series pick image path file fit into LLM-Vision
  • action script - Afteru can call what u want commandex: release LLM VRAM each call: "curl http://localhost:11434/api/generate -d '{"model": "llama2", "keep_alive": 0}'" @Pdonorex: bra bra. Interactive anything.
• tipsX: Enjoy it, inspire ur idea, and tell everybody how u use this.

Installtion

• You need install LM Studio or ollama first.
• • LM Studio: Start the LLM service on port 1234. (suggest use this one)
  • ollama: Start service on port 11434 .
• Pick one language model from under list
• • text base(small ~2G)
  • text&vision base(a little big ~8G)
• Start web-ui or ComfyUI install extensions or node
• • stable-diffusion-webui | stable-diffusion-webui-forge:go Extensions->Available [official] or Install from URLhttps://github.com/xlinx/sd-webui-decadetw-auto-prompt-llm
  • ComfyUI: using Manager install nodeManager -> Customer Node Manager -> Search keyword: autohttps://github.com/ltdrdata/ComfyUI-Managerhttps://registry.comfy.org/https://ltdrdata.github.io/
• Open ur favorite UI
• • Lets inactive with LLM. go~
  • trigger more detail by LLM

Suggestion software info list

• https://lmstudio.ai/ (win, mac, linux)
• https://ollama.com/ (win[beta], mac, linux)
• https://github.com/openai/openai-python
• https://github.com/LostRuins/koboldcpp (all os)

Suggestion LLM Model

• LLM-text (normal, chat, assistant)
• • 4B VRAM<2GCHE-72/Qwen1.5-4B-Chat-Q2_K-GGUF/qwen1.5-4b-chat-q2_k.ggufhttps://huggingface.co/CHE-72/Qwen1.5-4B-Chat-Q2_K-GGUF
  • 7B VRAM<8Gccpl17/Llama-3-Taiwan-8B-Instruct-GGUF/Llama-3-Taiwan-8B-Instruct.Q2_K.ggufLewdiculous/L3-8B-Stheno-v3.2-GGUF-IQ-Imatrix/L3-8B-Stheno-v3.2-IQ3_XXS-imat.gguf
  • Google-Gemmahttps://huggingface.co/bartowski/gemma-2-9b-it-GGUFbartowski/gemma-2-9b-it-GGUF/gemma-2-9b-it-IQ2_M.ggufsmall and good for SD-Prompt
• LLM-vision 👀 (work with SDXL, VRAM >=8G is better )
• • https://huggingface.co/xtuner/llava-phi-3-mini-ggufllava-phi-3-mini-mmproj-f16.gguf (600MB,vision adapter)⭐⭐⭐llava-phi-3-mini-f16.gguf (7G, main model)
  • https://huggingface.co/FiditeNemini/Llama-3.1-Unhinged-Vision-8B-GGUFllava-llama-3.1-8b-mmproj-f16.gguf⭐⭐⭐Llama-3.1-Unhinged-Vision-8B-Q8.0.gguf
  • https://huggingface.co/Lewdiculous/Eris_PrimeV4-Vision-32k-7B-GGUF-IQ-Imatrix#quantization-informationquantization_options = ["Q4_K_M", "Q4_K_S", "IQ4_XS", "Q5_K_M", "Q5_K_S","Q6_K", "Q8_0", "IQ3_M", "IQ3_S", "IQ3_XXS"]⭐⭐⭐⭐⭐for low VRAM super small: IQ3_XXS (2.83G)in fact, it's enough uses.

Using Online LLM Service Setup example

OpenAI ChatGPT

• In Auto-LLM Setup tab
• • LLM-URL=https://api.openai.com/v1
• get ur api key from openAI : https://platform.openai.com/api-keys
• • LLM-API-KEY = xxxxxxxxxxxxxxxxxxxxxxx
  • LLM-Model-Name = gpt-3.5-turbo

Google Gemini

• LLM-URL= https://generativelanguage.googleapis.com/v1
• get the api key: https://ai.google.dev/gemini-api/docs/api-key?hl=zh-tw

X Grok

• register here: https://x.ai/geo-block
• Its not open for my region. i cant test for u guys.

claude.ai

• seems not for api call: https://claude.ai/upgrade

Hugging face space

• https://huggingface.co/spaces

Javascript!

security issue, but u can consider as follows.

• https://github.com/pmcculler/sd-dynamic-javascript
• https://github.com/ThereforeGames/unprompted
• https://github.com/adieyal/sd-dynamic-prompts
• https://en.wikipedia.org/wiki/Server-side_request_forgery
• and Command Line Arg --allow-code

Buy me a Coca cola ☕

https://buymeacoffee.com/xxoooxx

Colophon

Made for fun. I hope if brings you great joy, and perfect hair forever. Contact me with questions and comments, but not threats, please. And feel free to contribute! Pull requests and ideas in Discussions or Issues will be taken quite seriously! --- https://decade.tw