ELSAC, the open-source way to build a consistent character dataset for LoRA training.
Eliz's LoRA Samples Analyzer & Curator
ELSAC takes your raw character images, organizes them by similarity, scores them against your best references, and tells you exactly which ones are worth training on — so you spend less time staring at folders and more time creating.
It works with 6 Steps:
Project Setup: Create your project or load and resume a previous one.
Clustering: CLIP + HDBSCAN group similar images into clusters and outliers.
Single Reference: Rank everything against one image that best represents your character.
Face Likeness: Faces are detected, cropped & aligned before comparing — real facial consistency.
Multiple Reference: Combine 4–8 varied references into one averaged identity, then rank your set.
Analysis: Analyze for blur, face-ratio & yaw metrics exported as CSVs for final filtering.
About this project and why it's free:
I began this project a long time ago, and by that, I mean: years. Every single step you see here was a single script I wrote for personal dataset training; then I decided this year to "vibe code" them all together and make a single nice app with them. I also added new features it didn't have.
So, I'm just giving it away to the community; you can use it, modify it, make it better- anything; it's your choice now.
Documentation and links:
Features and installation: https://gabrielx.com/elsac-elizs-lora-samples-analyzer-and-curator/
Docs & Recommendations: https://gabrielx.com/docs-elsac-documentation-user-guide-dataset-recommendations/
GitHub Repo: https://github.com/gabrielxcom/elsac
More free projects: https://gabrielx.com/open-source-mit-ai-tools/
¿Hablas español? La documentación también está disponible en español:
🔗 https://gabrielx.com/es/herramientas-ia-de-codigo-abierto-licencia-mit/
Support us!
You can support us by listening to our songs (yes, as silly as that). Visit: www.IamElizAi.com 🎶