16bit-from-8bit Image Reconstruction Model
This model reconstructs 16-bit per channel images from standard 8-bit input images. It is trained on paired datasets and optimized to preserve color fidelity, structural consistency, and high-frequency detail.
- Median MAE: 410
- LPIPS (Alex): ~0.044 (60-image evaluation)
- Architecture Update: Added Leaky ReLU
- Training Resolution: 256×256(46k Patches from Raw HDR images with 8,580 48bit synthetic images.)
- Training Resolution: 512×512 (Hand Selected Dataset 2k)
- Training Resolution: 1024×1024 (Hand Selected Dataset 500)
Dataset
- Total images: 54,580
- RAW patch images: 46,000 @ 256x256 (~10 GB)
- 48-bit synthetic images: 8,580 (~2 GB)
Addtional 10GB in Hand Selected RAW images, for the 512px and 1024px High Frequency Training Passes
MAE Distribution (8-bit → 16-bit reconstruction)
| MAE Range | Accuracy Comment | Percent (%) |
|---|---|---|
| ≥1000 | Occasionally visible in uniform areas | 1.06 |
| 600–1000 | Almost never visible | 10.03 |
| 400–600 | Fully imperceptible | 27.39 |
| 200–400 | Near perfect | 59.95 |
| ≤200 | Near exact scientific | 1.57 |
Perceptual Metrics (60-image test set)
| Metric | Result | Interpretation |
|---|---|---|
| LPIPS (Alex) | 0.044 | Low perceptual distance / high similarity |
| Gradient Energy | 0.088 → 0.108 | Preserved fine detail, slight sharpening |
| FFT Structure Score | 1.07 → 1.23 | Improved high-frequency retention |
| Histogram Continuity | 11.2 → 11.3 | Stable tonal distribution |
Interpretation Summary
- LPIPS values (~0.03–0.07 range) indicate high perceptual similarity
- Structural metrics (FFT + gradients) show consistent detail reconstruction
- Histogram stability indicates no major tonal drift between bit-depth conversions
Intended Use
Primary Use Cases
- Reconstruction of 16-bit per channel images from 8-bit input
- JPG & GIF post-processing and enhancement
- Archival and art restoration workflows
Not Intended For
- Lossless scientific measurement or precision tasks
- Medical AI enhancement