ELSAC — Eliz's LoRA Samples Analyzer and Curator

By Gabriel Solis (Gabriel X)

ELSAC helps you organize, identify, and curate character images for training a consistent LoRA in diffusion models.

It uses CLIP embeddings + HDBSCAN to cluster visually similar images, then guides you through several scoring passes to find your strongest samples, and finally generates technical quality reports (blur, face ratio, head yaw).

📄 Documentation, User Guide & Dataset Recommendations are hosted on gabrielx.com/open-source-mit-ai-tools/.

Workflow

1. Project Setup — create a character project and upload raw images
2. Clustering — group images by visual similarity (CLIP + HDBSCAN)
3. Single-Reference Scoring — rank against one reference image
4. Face Likeness Scoring — rank against a reference face (face-cropped)
5. Multi-Reference Scoring — rank against 4–8 references (recommended)
6. Technical Analysis — blur / face ratio / head yaw → CSV reports

Requirements

• Python 3.11 (exactly 3.11.x)
• Windows, macOS, or Linux
• NVIDIA GPU with CUDA support (optional — CPU works)

Quick Start (Windows)

setup.bat
venv\Scripts\activate.bat
python app.py

Quick Start (macOS / Linux)

chmod +x setup.sh
./setup.sh
source venv/bin/activate
python app.py

The setup script asks which installation you want:

Option	For
1	CUDA 12.8 (RTX 4000 and 5000 series / Blackwell)
2	CUDA 11.8 (RTX 3000 series and older)
3	CPU only

ELSAC opens at http://127.0.0.1:7860.

Manual Installation

python -m venv venv
source venv/bin/activate          # Windows: venv\Scripts\activate.bat
pip install --upgrade pip setuptools wheel
pip install -r requirements/requirements-cpu.txt     # or a GPU file
python app.py

First Run

The first time you use face scoring/analysis, InsightFace downloads its model (buffalo_l) automatically to your home folder (~/.insightface). This requires an internet connection the first time.

Project Layout

Projects are created in the characters/ folder next to the application:

characters/
└── <character_name>/
    ├── raw_images/
    ├── clustered/
    │   ├── cluster_0/
    │   ├── outliers/
    │   └── cluster_log.txt
    ├── scored/
    │   ├── general_likeness/
    │   ├── face_likeness/
    │   └── multi_likeness/
    │       ├── reference_set/
    │       └── scoring_results.csv
    └── analyzed/
        └── quality_analysis_*.csv

Building a Standalone Windows App

pip install -r requirements/requirements-dev.txt
pip install pyinstaller
pyinstaller elsac.spec

The build appears in dist/ELSAC/.

License

MIT — see LICENSE.