SAE Explorer

A tool for exploring AI-generated images by combining concepts with different stylistic features. This project allows you to generate and visualize how different artistic features affect various concepts.

Components

1. Generation CLI (main.py)

A command-line tool for generating images by combining concepts with features at different strengths.

# Generate 3 variations of a "mountain landscape" with "mystical" features
uv run main.py generate "mountain landscape" "mystical" --variations 3 --min-strength 0.2 --max-strength 0.8

# List all discovered features related to "minimalist"
uv run main.py list-features "minimalist"

Options:

• --variations/-n: Number of strength variations to generate (default: 1)
• --min-strength: Minimum feature strength (-1 to 1, default: -0.5)
• --max-strength: Maximum feature strength (-1 to 1, default: 0.5)
• --num-features/-f: Number of top features to apply (1-5, default: 1)
• --verbose/-v: Enable verbose output

2. Visual Explorer (app.py)

A Streamlit web application for exploring and visualizing the generated images. Features include:

• Browse concepts and their associated features
• View images at different feature strengths
• See generated prompts and metadata
• Interactive strength slider for real-time exploration

# Run the web interface
uv run streamlit run app.py

Setup

1. Clone the repository

git clone https://github.com/lakshyaag/sae-explore.git
cd sae-explore

2. Set up Supabase to persist generations and view them in Streamlit:

   1. Create a new project at Supabase
   2. Initialize the database schema:
      • Go to Database > SQL Editor
      • Copy the contents of db.sql from this repository
      • Run the SQL commands to create the required tables:
        • concepts: Stores the base concepts for image generation
        • features: Stores discovered features and their metadata
        • generations: Stores the generated images and their relationships

3. Create an account on Goodfire

4. Configure your image generation. By default, I'm using Fal.ai

5. Set up environment variables in .env:

GOODFIRE_API_KEY=your_goodfire_key
FAL_KEY=your_fal_key

SUPABASE_URL=your_supabase_url
SUPABASE_ANON_KEY=
SUPABASE_SERVICE_KEY=your_supabase_key

6. Install dependencies

uv sync

Architecture

The project uses:

• Goodfire: For feature discovery and prompt enhancement
• Supabase: For storing concepts, features, and generations
• Fal.ai: For image generation
• Streamlit: For the web interface
• uv: God's answer to poetry

License

MIT