ImageNet should retire: DiT method rankings negatively correlate between ImageNet and text-to-image

YouJiacheng · x · 2026-10-16

• Halcyon (Radical Numerics): early-fusion text-to-image base models trained from scratch outperform strong late-fusion baselines with up to 275× less pretraining compute; models and recipes are open.
• DiffusionBench (arXiv 2606.24888): argues DiT research has over-converged on class-conditional ImageNet eval. Introduces NanoGen, a unified training/eval framework where switching from ImageNet to T2I takes 12 config lines at comparable compute. Across 21 latent diffusion models, method rankings on ImageNet vs T2I show negative correlation (Pearson -0.377 to -0.580)—ImageNet gains don't transfer to real T2I.
• The retweaker adds: ImageNet's 1.28M samples force >100-epoch training of billion-scale class-to-image models, a clearly overfit setup; diffusion may just be good at ImageNet, while AR models scale fine, and GPIC with FD (100M richly annotated samples) is a better T2I benchmark.

Original post →

More from Multimodal

• How meloform works: WASM pitch detection plus differentiable circuits learning 1,022 sine coefficients — henloitsjoyce · 2026-10-16
• Documentary filmmaker urges peers to watch 'You Can See Everything' — kyliebytes · 2026-10-16
• Redditor's first Cumfy/Seedance video nails a cinematic one-shot generation — Wave_N_Wind · 2026-10-16
• Kling 4 returns after 8 months: 10 hyper-realistic AI ads for under $5 — CurieuxExplorer · 2026-10-16
• Shanghai AI Lab unveils Intern Lumina U2, a diffusion LLM unifying image, video and 3D tasks — Shanghai-AI-Laboratory · 2026-10-16
• NAVER AI Lab proposes Speculative Forcing to speed up autoregressive video generation — naver-ai · 2026-10-16

Multimodal channel →