SD3 critique says flow matching looks elegant but hides bad conditioning
kalomaze · x · 2026-07-23
A pointed critique of diffusion-style design choices, centered on SD3 and flow matching.
- The author argues SD3 relies on a “bullshit heuristic weightage” that hides the fact that vanilla flow matching has poor conditioning.
- They compare it to a hypothetical nanoGPT submission that would pass proxy evals even with no layer norm, while still being fundamentally bad.
- A follow-up reply says the field still hasn’t learned that poorly conditioned reparameterizations are not necessarily useful just because they come with elegant theory.
The thread’s core claim is that theoretical elegance does not make up for weak empirical conditioning or lack of simple wins.
More from Research
- New ASCIITermDraw benchmark says top VLMs still miss simple text diagrams — East-Muffin-6472 · 2026-07-23
- DocOps benchmark finds frontier agents still fail on long-horizon document tasks — Jiazhen Jiang · 2026-07-23
- Stanford’s vine-like soft robot grows from the tip to reach trapped people — lukas_m_ziegler · 2026-07-23
- First CAR-T Cell Therapy Approved for Solid Tumors in Gastric Cancer — Dr_Singularity · 2026-07-23
- A production multi-agent team says deterministic orchestration works better than deterministic LLMs — njanChe1 · 2026-07-23
- LxMLS 2026 shares a public video-lecture collection from Lisbon Machine Learning School — caglar_ee · 2026-07-23