ApacheOne / voir-real-mage-albedo-v1

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-flow-edit-turbo-cache16
• Actual Mage cache images: 12 train / 4 validation
• Projected hidden channels: 2240
• Sampler channels: 2688
• Auxiliary channels: 63

Best held-out validation

{
  "epoch": 2,
  "train_loss": 0.18375273452450833,
  "lr": 0.0016333333333333334,
  "seconds": 51.5237991809845,
  "validation": {
    "loss": 0.19177169352769852,
    "rgb": 0.09192945715039968,
    "gradient": 0.056893778033554554,
    "ssim": 0.33402326703071594,
    "chroma": 0.04807095415890217,
    "color": 0.07322677690535784,
    "multiscale": 0.09202386252582073,
    "total": 0.19177169352769852,
    "mae": 0.09289983753114939,
    "mse": 0.019575452664867043,
    "psnr": 17.985229048658795,
    "ssim_7x7": 0.6659767329692841,
    "global_ssim": 0.683689147233963
  }
}

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.