Researchers introduce DiffusionBench for holistic DiT evaluation

The negative correlation between ImageNet FID improvements and text-to-image success is a significant wake-up call for the generative AI community, showing how narrow benchmarks can mislead research directions. By introducing NanoGen, the authors eliminate the common excuse that text-to-image evaluation is too costly, democratizing comprehensive model testing. DiffusionBench has strong potential to become the new standard in DiT research, steering the field toward architectural innovations that generalize rather than overfit to standard datasets.