Research

The map is generated.
The measurement is not.

The goal is a realistic, plausible version of what the satellite saw — not a pixel-exact reconstruction, which the physics doesn't permit. Inventing plausible detail is the point; staying consistent with the source is the law.

System vision target architecture

  1. Multi-temporal LR acquisition — ten Sentinel-2 revisits per site
  2. Alignment — co-registration against the reference frame
  3. Modality set — spectral composites, OSM vectors, structure priors
  4. S — the positional representation of all modalities
  5. S + the date-matched LR condition two ControlNets in development
  6. Predicted HR → height estimation / 3D in development

ControlNet conditioning was tested and concluded negative in 2026-06 experiments; the redesign above is a target, not a shipped result — we say so here on purpose.

Shipping today validated

  1. Corpus: 730 co-registered S2↔aerial pairs, 7 granules, 10-frame temporal stacks
  2. S_struct — a 6-class structural map (roads, buildings, rail, water, vegetation, bare) + confidence
  3. Stage A — SEAN generator + OASIS discriminator GAN: 10 m → 2.5 m (×4), structure-conditioned
  4. Stage B — LoRA-adapted pixel diffusion texture decoder: 2.5 m → 0.5 m research-only · weights license
  5. Verification — round-trip to 10 m + texture statistics vs real-aerial banks, every output

Evaluation protocol

Pre-registered gates, anti-over-texture guards, an adversarial judge with its own published failure, and a human forced-choice test as the success criterion. Full yardstick with one-line explanations on the Evidence page; raw values in metrics.json.

Milestones

  1. 2026-06-16Slovenia S2↔DOF corpus v1 — grown to 730 co-registered pairs, ≈1,196 km²
  2. 2026-06-19ControlNet conditioning concluded NEGATIVE at both sites; pivot to the SEAN+OASIS GAN
  3. 2026-07-02Goal ratified: realistic-plausible over GT-exact; cascade beats the bicubic floor for the first time
  4. 2026-07-04Goal reorder: 2.5 m indistinguishability first; autonomous evaluation loop goes live
  5. 2026-07-05Goal-metric suite v2 frozen; 13 arms + 5 public external models scored on identical inputs
  6. 2026-07-06First human blind-test session logged (0.90 vs goal 0.50); conditioning-map rebuild becomes the active workstream

Negative results stay in the record — the ControlNet conclusion above is one of several we publish rather than bury.