Scaling handbook for text-to-image on 100M pairs: retire ImageNet, drop CFG and EMA
DimitrisPapail · x · 2026-11-17
Wenhao Chai published "How to Scale Text-to-image Models: a Handbook," training pixel-space JiT models from 3.6e18 to 2.3e20 FLOPs on GPIC (100M permissive text-image pairs), bringing LLM pretraining science to visual generation.
Key findings:
- ImageNet should retire: only 1.28M images with class labels, causing overfitting over tens to hundreds of epochs and CFG abuse; prior work shows these tricks don't transfer to real text-to-image.
- Compute-optimal TPP sits around 50 image tokens per image parameter, with held-out loss as the main metric.
- EMA is unnecessary: a proper LR schedule beats EMA on held-out loss at every budget by 0.4–1.0%.
- Drop CFG: all results use only the conditional branch (CFG=1), saving compute and removing metric-hacking room.
All code, configs, and sweep data are open-sourced.
More from Multimodal
- YC-backed o11 launches Videos in Wallet: prompt-to-AI-video with no subscription — ycombinator · 2026-11-17
- Hooking Fable 5.2 to Streetview API spawns a hilariously accurate Amsterdam scene — IgorCarron · 2026-11-17
- VOOPA Render Check: An MCP Tool for Reviewing AI Video Takes Before Retrying — Decent_Diamond9197 · 2026-11-17
- 'Vintage rock video?' AI-generated music video fools Reddit on realism — Double_Language_6390 · 2026-11-17
- Hala MCP + Claude workflow generates pro-grade motion graphics videos from prompts — aziz4ai · 2026-11-17
- 50 sequential edits stress test: Seedream drifts least at 36.6, Nano Banana at 74.7 — aziz4ai · 2026-11-17