LoRA Model Evaluator

Semi-automatic evaluation pipeline for LoRA fine-tuned diffusion models with interactive scoring and automated reports.

Slides

A semi-automatic system for evaluating LoRA fine-tuned Stable Diffusion models. Built as a team of 3 for Brown CSCI 2952Y.

Pipeline:

  • Accept LoRA weights, metrics, base model, prompt lists, and seed lists
  • Generate images across all (weight x prompt x seed) combinations
  • Interactive web-based scoring UI with pairwise image rating
  • Automated report with optimal weight analysis and representative best/worst examples

Highlights:

  • LLM-assisted metric generation for defining evaluation criteria
  • Weight-quality curve plotting across parameter ranges
  • User study with N=100 images validated the system’s effectiveness

Technologies: Python, Stable Diffusion, LoRA, Web UI, LLM