stable-diffusion

stable diffusion

图形界面

• stable-diffusion-webui

模型

• 国风
• waifu
• Stable diffusion v1.5
• F222系列
• Anything V3
• Open Journey
• 机器人扩散
• Robo Diffusion
• Inkpunk Diffusion
• Waifu-diffusion

学习

• sdupdates3
• stable-diffusion-art
• prompts

sd插件

• lora、hypernetwork list
• List of SD innovations
• prompts editing
• Prompt Generator
• Prompt Translator
• sd-tagging-helper
• AIDraw
• sd_dreambooth_extension with lora
• multidiffusion-upscaler
• 标记器扩展-tokenizer
• openpose-hand-editor

ps插件

• Auto-Photoshop-StableDiffusion-Plugin

视频

• Swap Face

训练

• 综述1
• 综述2
• hypernetwork
  • 网络结构最好不要超过三层，几十个素材就用默认1,2,1结构，几百张就用1,1.5,1.5,1结构

  • 激活函数

    • liner:通用性较好
    • gelu:质感比较强，结构学习稍差
• drambooth
  • dreambooth
• lora
  • LoRA-SD-training
  • lora
  • LoRA Training Guide
  • LoCon

小模型使用:

• lora:bbb:1, hypernet:ccc:1

prompts

• https://www.chichi-pui.com/
• https://worddrop2022.notion.site/worddrop2022/AIi-WordDrop-2a87f5fc397a4171b031a12a44ff6b76
• https://seesaawiki.jp/nai_ch/d/��R18�Ӿ���ʸ
• How does attention/emphasis in prompts actually work? #2905

links

• News: https://rentry.org/sdupdates#newsfeed
• Prompting: https://rentry.org/sdupdates#prompting
• Models, Embeddings, and Hypernetworks: https://rentry.org/sdupdates#models-embeddings-and-hypernetworks
• Embeddings: https://rentry.org/sdupdates#embeddings
• Hypernetworks: https://rentry.org/sdupdates#hypernetworks
• Aesthetic Gradients: https://rentry.org/sdupdates#aesthetic-gradients
• Models, embeds, and hypernetworks that might be sus: https://rentry.org/sdupdates#polar-resources
• DEAD/MISSING: https://rentry.org/sdupdates#deadmissing
• Training: https://rentry.org/sdupdates#training
• Datasets: https://rentry.org/sdupdates#datasets
• FAQ: https://rentry.org/sdupdates#common-questions-ctrlcmd-f
• Link Dump: https://rentry.org/sdupdates#rentrys-link-dump-will-sort
• Confirmed Drama: https://rentry.org/sdupdates#confirmed-drama
• Unconfirmed Drama: https://rentry.org/sdupdates#unconfirmed-drama
• Fed Bait Information: https://rentry.org/sdupdates#fed-bait-information
• Hall of Fame: https://rentry.org/sdupdates#hall-of-fame
• Miscellaneous: https://rentry.org/sdupdates#misc
• Github: https://github.com/questianon/sdupdates/

采样类型比较

• Euler a ：是个插画，tag利用率仅次与DPM2和DPM2 a，环境光效菜，构图有时很奇葩 ,适合光滑结构
• Euler：柔和，也适合插画，环境细节与渲染好，背景模糊较深。
• Heun：单次出图平均质量比Euler和Euler a高，但速度最慢，高step表现好。
• DDIM:适合宽画，速度偏低，高step表现好，负面tag不够时发挥随意，环境光线与水汽效果好，写实不佳。
• DPM2:该采样方法对tag的利用率最高，几乎占80％+
• DPM2 a：几乎与DPM2相同，对人物可能会有特写
• PLMS: 单次出图质量仅次于Heun。
• LMS: 质感OA,饱和度与对比度偏低，更倾向于动画的风格
• LMS Karras：会大改成油画的风格，写实不佳。
• DPM fast：此为上界开发者所遗留的测试工具，不适合魔术师使用

how to describe body ratio in stable diffusion

• https://photogpedia.com/stable-diffusion-tips-and-tricks/
• When describing body ratio in Stable Diffusion, it is important to consider the proportions of the body parts to each other. For example, the head should be proportional to the body, and the arms and legs should be in proportion to each other. Additionally, the body parts should be in proportion to the overall size of the figure. When using Stable Diffusion to generate images, it is also important to consider the aspect ratio of the generated image, as this can have a significant impact on the visual effect of the generated image.