AIGC Drawing/Image Creation

2-minute read
AI
Edit

I. Installation 🔗︎

1. Install Homebrew 🔗︎

2. Install Stable Diffusion Web UI 🔗︎

Steps: 🔗︎

  1. Install Dependencies
    Use Homebrew to install necessary tools and libraries:
    brew install cmake protobuf rust [email protected] git wget
  2. Clone the Repository
    Clone the AUTOMATIC1111 Stable Diffusion Web UI repository:
    git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui
  3. Download Models
    Download and place the model file into the stable-diffusion-webui/models/Stable-diffusion/ directory. For example:
  4. Start the Service
    Start the Web UI:
    ./webui.sh
    Access URL: http://127.0.0.1:7860/

Common Issues and Solutions: 🔗︎

  • Error Message: RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'
    • Modify the webui.sh file and add the following arguments:
      export COMMANDLINE_ARGS="--precision full --no-half --skip-torch-cuda-test"
    • Restart the service.

More References: 🔗︎

II. Theory 🔗︎

1. LoRA Models 🔗︎

  • Definition: LoRA (Low-Rank Adaptation) reduces computational load by freezing pre-trained model weights and injecting trainable layers into each Transformer block.
  • Usage: Generate images with specific styles, faces, actions, etc.
  • Resources: CivitAI LoRA Model Collection

2. Colab (Google’s Free GPU) 🔗︎

  • Platform: Google Colab
  • Launch Command:
    python launch.py --share --xformers --enable-insecure-extension-access

III. Practice 🔗︎

1. Checkpoint Models 🔗︎

2. Quick Image Generation Tips 🔗︎

2.1 Adjust Parameters to Find the Best Results 🔗︎

  • Basic Parameters:
    • Sampler: DPM++ 2M Karras
    • Seed: Using the same value keeps the character consistent; if Resize seed from width/height is checked, consistency can be maintained even with different dimensions.
    • Restore Faces: Face restoration feature.
    • HiRes.Fix: Enhances image quality.
    • CFG Scale: Controls how much the generated image conforms to the prompt. The default value is 7 or 8.
  • Reference: Stable Diffusion Art Parameter Guide

2.2 Efficient Image Generation 🔗︎

  • Batch Count: Generates multiple times in a loop with the same configuration.
  • Batch Size: Generates multiple images at once, making it easier to select the best ones.

2.3 Upscaling Images to HD 🔗︎

  • The built-in features are usually sufficient, but you can use the following tools to further improve clarity:

2.4 One-Click Image-to-Video 🔗︎

  • Functionality not yet detailed.

3. Changing Image Poses 🔗︎

  • Plugin: ControlNet
    ControlNet is a powerful extension for controlling the pose, layout, etc., of generated images.

4. Prompt Generators 🔗︎

5. Community Resources 🔗︎

IV. Additional Notes 🔗︎

  • For Windows Users: You can directly search for relevant tutorials; the installation process is relatively simple.
  • For Mac Users: It is recommended to follow the steps above and adjust the environment configuration based on any error messages.