DiffusionBench: Towards Holistic Evaluation of Generative Diffusion Transformers

DiffusionBench is a unified codebase for training and evaluating generative diffusion transformers across multiple tasks including image generation and text-to-image synthesis. The project has released new 512×512 resolution text-to-image checkpoints trained on multiple VAE variants and published a preliminary technical report (v0.1) with plans for community-driven improvements. The framework supports standardized evaluation metrics (FID, IS, GenEval) and aims to move beyond ImageNet-only benchmarking for comprehensive assessment of diffusion models.