Distilling from weak labels: 1M images trained for $3.24

vanstriendaniel · x · 2026-08-18

The author presents a project extracting embellishments from the British Library book images dataset. The model was distilled entirely from weak labels with zero human annotations. The full 1M-image corpus training run cost only $3.24 in GPU compute. Links to the model, training data, and mask configuration are provided.

Original post →

More from Infra

Infra channel →