Near-identical image scores, huge gaps: AI denoising must serve science, not looks
bravo_abad · x · 2026-10-05
A study by Kamijo and colleagues compares two AI denoising methods for scientific measurements: their image-similarity scores are nearly identical, yet they differ sharply in recovering diffraction spots that reveal a material's crystal structure. A missing spot hides structural information; an invented one suggests something that isn't there.
- The better approach uses a self-supervised neural network that sees nine diffraction patterns from neighbouring specimen positions, learning from shared information without clean training images
- The comparison method denoises each pattern separately, then smooths across neighbouring measurements
- At one simulated low-dose setting, spot-recovery score reaches 0.58 despite near-identical image scores
Takeaway: image-quality metrics can't replace downstream task evaluation; denoising should be judged by measurement reliability.
More from Models
- Dev claims mystery system hits 100% on agentic benchmarks, can't explain how — examachine · 2026-10-05
- Dev slams Opus for refusing to fix security holes that GLM5.3 found — lucasmeijer · 2026-10-05
- Whistle, an on-device speech recognition model, trends on Hugging Face — Cactus-Compute · 2026-10-05
- New "systems are thinking" guardrail message hints at OpenAI human review — Darpinian · 2026-10-05
- Decision Index: 70 open reproductions of Jev's Decision Model, one benchmark — bibryam · 2026-10-05
- Critic Concedes OpenAI's GPT Computer-Use Now Drives Safari Like a Human, Barely Errs — kimmonismus · 2026-10-05