Researcher argues TPP is an odd metric for generative image models vs AR LLMs
torchcompiled · x · 2026-11-12
In a thread on diffusion training compute baselines, torchcompiled argues that TPP (tokens per image) feels like an odd metric for generative image models: it makes sense computationally, but it's unclear how "one sample" maps in information content compared to the AR LLM task, so a baseline for image data is worth establishing.
Related event: Researchers debate TPP as a training metric for diffusion models(3 posts)→
More from Research
- UvA opens two PhD positions on LLM uncertainty and explainability in healthcare — CharlotteHase · 2026-11-12
- RealtimeVAD: Open-Source VAD Trades False Positives for ~110ms Faster End-of-Speech — Lonligrin · 2026-11-12
- Diffusion 101: building an image generator from scratch on two CPU cores — dima806_dima · 2026-11-12
- Apple's AMP1 hits #1 on NVIDIA RoboLab leaderboard by writing Python to control robots — chris_j_paxton · 2026-11-12
- Robot imitation learning plateaus with more data, pushing the field toward RL — chris_j_paxton · 2026-11-12
- Mathematician Daniel Spielman on the AI-driven "Big Reset" in mathematics — stevenstrogatz · 2026-11-12