Physics-based enhancement vs hallucinated super-resolution
A generative model can turn a blurry 10 m pixel into a crisp-looking picture. It does so by inventing detail that is statistically plausible but not observed. For a wallpaper, that's fine. For anything you intend to measure, it's a trap.
Why invented detail is a liability
Remote-sensing products are used to measure areas, detect change, and geolocate features. A hallucinated edge that isn't really there produces a false change between two dates; an invented building corner shifts a measurement. Worse, the errors are confident and look real, so they survive review.
The requirement is therefore stronger than 'looks sharp': the enhanced pixel must remain faithful to the observed radiance and keep the geolocation of the source. Detail should be reconstructed from real information — sub-pixel cues across bands and time — not synthesised from a prior.
What EarthToDate does instead
The enhancement is physics-based: calibrated reconstruction that sharpens without changing geolocation accuracy or inventing structure, so derived indices, overlays and change detection stay trustworthy. The visible product reaches a true 1 m from Sentinel-2 and the multispectral stack 2–4 m, on the source's own revisit cadence.
The honest framing is that this is enhancement with guarantees, not magic: it will not reveal something the sensor never captured. That constraint is the point — it's what makes the output safe to measure against.