Products · ASI Prism · Detail tiers

Concept guide

Why there is more than one answer

A 10 m pixel can be filled in more than one way, and every way is a hypothesis. ASI Prism does not pick one on your behalf — it gives you a range of settings, tells you which parts of each output are measurement and which are reconstruction, and keeps the same guarantee at every point on the range.

Try the settings yourself → Back to ASI Prism

The trade-off that makes one setting impossible

Super-resolution has to invent what a coarser sensor never recorded. Push harder and edges sharpen, field boundaries separate, small structures appear — and the chance that some of that structure is invented rather than recovered rises with it. Hold back and almost everything you see is defensible, but you have gained less.

Neither end is correct in general. It depends entirely on what the image is for. Drawing a boundary you will legally rely on is not the same task as spotting where to walk a field, and the right amount of reconstruction differs between them. So the choice is exposed rather than hidden.

The range, described by behaviour

Settings are presented as a spectrum between two ends, with a continuous dial between them:

  • The conservative end. Adds the least. Used as the reference other settings are judged against — if something appears here, it is about as defensible as reconstruction gets.
  • The balanced setting. The evidence-supported default, and what we recommend unless you have a reason to move. Meaningfully more detail than the conservative end while still clearing the acceptance tests.
  • The detail end. The most visible reconstruction, and the highest measured hallucination risk. Useful for interpretation and orientation; the wrong place to take a measurement from.

The public demo happens to open on a detail-forward setting because it shows the capability best. That is a demonstration choice, not a recommendation — the supported default is the balanced one, and the demo labels every panel with its own maturity badge so the distinction stays visible.

What does not change, whatever you pick

Every setting is bound by the same constraint: average the reconstructed output back down to the original grid and you get the measurement back. That equality is structural — built into how the output is formed, not something the model was merely trained to approximate.

Measured on 681 real test tiles, the low-frequency residual is 7.3×10⁻⁹ before delivery formatting. We say before deliberately: writing the file applies clipping, nodata handling and integer quantisation, after which the residual is 2.4×10⁻⁷. Delivered files are therefore near-consistent, not literally exact, and we would rather publish both numbers than the flattering one.

What the error map is — and is not. Every reconstruction ships beside a per-pixel map showing where the output strains against the measurement it came from. It is a consistency check, not an accuracy check. A confidently invented texture that happens to average out correctly can score near zero. Read it as "where is this output arguing with the data", never as "where is this output wrong".

When the model is allowed to decline

Some settings can additionally mark pixels where the reconstruction is not trustworthy enough to use, and abstain there rather than answer. Two limits are worth stating plainly, because an abstention capability is easy to over-sell:

  • It is available only on settings that have been calibrated for it, and only at the scale that calibration was performed at. It does not transfer to other scales, and we do not let it pretend otherwise.
  • It supports triage. It is not a guarantee, and its status in our own evidence record is conditional — not supported.

Measured, and inferred

In any output, the average over each block of reconstructed pixels is the measurement. What varies inside that block is inference. That line runs through every setting on the range, and the product is built to keep it visible rather than blur it.

There is a consequence people are often surprised by: the metrics that score how much detail was correctly added need a higher-resolution reference image to compare against. Over a point you pick on a map, no such reference exists. So for your own imagery you get the consistency map and, where calibrated, the abstention layer — and an honest absence of a correctness score, rather than a number that cannot mean what it appears to.

What we don't claim. A ×4 output is 2.5 m grid spacing. It is not a 2.5 m measurement, and we will not describe it as one. This is frozen R&D V1, optical only, and reconstructed detail is an audited hypothesis throughout — never ground truth. Worst-case hallucination on the detail end is high by our own measurement, which is exactly why the setting carries a badge instead of a reassurance.