Researchers debate TPP as a training metric for diffusion models

Researchers debate whether TPP (tokens per image) is a sound training-volume metric for diffusion models, with torchcompiled arguing extra gradient noise from sampled t and noise slows training; Jiawei Yang offers DiT baselines, e.g. XL at 194 TPP over 400 epochs on ImageNet-1K.

2026-11-12 ~ 2026-11-12 · 3 related posts

• DiT TPP baseline: 400 epochs on ImageNet hits ~1,009/286/194 TPP for B/L/XL — JiaweiYang118 · 2026-11-12
• Researcher argues TPP is an odd metric for generative image models vs AR LLMs — torchcompiled · 2026-11-12
• torchcompiled: slower diffusion training likely stems from extra gradient noise — torchcompiled · 2026-11-12