Research
DiffusionBench shows near-zero correlation between ImageNet and text-to-image DiT rankings
Tags AI
arXiv·
Researchers introduced NanoGen and DiffusionBench, demonstrating that method rankings on ImageNet classification and text-to-image generation have near-zero correlation (-0.377 to -0.580), arguing that DiT research must evaluate both tasks.
Technical significance
The finding challenges assumptions that architectural improvements transfer across tasks. DiT researchers should benchmark on both classification and generation to avoid overfitting evaluation to a single domain.