Determine the cause of the FunDPS reproduction discrepancy

Determine which, if any, of the differences between the reproduced FunDPS configuration and the configuration used in the original report explains the discrepancy in accuracy.

Background

The paper reports that the reproduced FunDPS configuration does not attain the accuracy reported in the original FunDPS paper. Several methodological differences are identified, including the use of 5% rather than 3% observations, direct training at 128×128 resolution instead of a two-stage multiresolution training pipeline, and different hardware and distributed-training settings.

The authors explicitly leave unresolved which, if any, of these differences accounts for the observed performance discrepancy. This uncertainty limits the extent to which the reproduced comparison can be interpreted as a comparison with the originally reported FunDPS performance.