diffusionbench-evals
What it is
A unified repository containing four independent evaluation packages for image generation models: DPG-Bench, GenEval, VQAScore, and FID/FDR6/MIND. It consolidates these metrics into a single codebase, standardizing on modern dependencies like Hugging Face Transformers 5.3.0 and uv for installation.
The problem it solves
The fragmentation of image generation evaluation metrics across multiple upstream repositories with inconsistent dependencies and outdated codebases, making it difficult to run a comprehensive suite of benchmarks in a single environment.
What you could use it for
- Evaluating the compositional quality of text-to-image models using DPG-Bench
- Measuring object presence, count, and spatial relationships in generated images with GenEval
- Assessing image-text alignment using VQAScore with Qwen backends
- Calculating distributional metrics like FID and MIND across different representation spaces
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.