DiT TPP baseline: 400 epochs on ImageNet hits ~1,009/286/194 TPP for B/L/XL
JiaweiYang118 · x · 2026-11-12
To give ImageNet-centric researchers a shared reference for how data-hungry diffusion models are, Jiawei Yang published a table of typical TPP values using ImageNet-1K and standard DiT sizes (B=130M, L=458M, XL=675M, excluding the frozen VAE):
- 100 epochs (32.8B tokens): 252/72/49 TPP for B/L/XL
- 200 epochs (65.6B): 505/143/97
- 400 epochs (131.2B): 1,009/286/194
- 800 epochs (262.4B): 2,018/573/389
- 1,400 epochs (459.2B): 3,532/1,003/680
The same data gets reused up to 1,400 times, underscoring how heavily diffusion training recycles data.
Related event: Researchers debate TPP as a training metric for diffusion models(3 posts)→
More from Infra
- AWS Emulator Was Written 100% by Thousands of Autonomous AI Agents — dfinke · 2026-11-13
- CoreWeave: qualified demand 10x capacity through 2027, 2028 demand already multiples of supply — Beth_Kindig · 2026-11-13
- Semafor: DOJ Opens Antitrust Probe Into NVIDIA — zephyr_z9 · 2026-11-13
- No succession plan, 60 direct reports: why Jensen Huang's NVIDIA exit looks messy — signulll · 2026-11-13
- Kubeflow graduates in CNCF as the open-source Kubernetes-native AI platform — PyTorch · 2026-11-13
- Google buys Crusoe's Wyoming data center site, commits $17B with 702 MW fuel-cell power — Beth_Kindig · 2026-11-13