VOIR Real Mage Albedo v1

Full-resolution trainable albedo decoder using frozen internal states from microsoft/Mage-Flow-Edit-Turbo.

Training data

Cache dataset: ApacheOne/voir-mage-pilot48-cache-v1
Actual Mage cache images: 38 train / 10 validation
Projected hidden channels: 2240
Sampler channels: 2688
Auxiliary channels: 63

Best held-out validation

{
  "epoch": 26,
  "train_loss": 0.08465919854413521,
  "lr": 0.00040548211460896867,
  "seconds": 299.20014929771423,
  "validation": {
    "loss": 0.0830109141767025,
    "rgb": 0.038485938543453814,
    "gradient": 0.024017343949526548,
    "ssim": 0.15509827733039855,
    "chroma": 0.02399692116305232,
    "color": 0.024400425766361877,
    "multiscale": 0.03854453098028898,
    "total": 0.0830109141767025,
    "mae": 0.03898913008160889,
    "mse": 0.007759020489174873,
    "psnr": 23.279880106718778,
    "ssim_7x7": 0.8449017226696014,
    "global_ssim": 0.852774053812027
  }
}

The training set contains only twelve real aerial-image pairs. This checkpoint is a real Mage-conditioned model, but it is an initial narrow-domain model rather than a general-purpose albedo estimator. More paired scenes are required for broad generalization.

Downloads last month: -; Downloads are not tracked for this model. How to track