| --- |
| license: mit |
| tags: |
| - mlx |
| - image-to-image |
| - scratch-detection |
| - photo-restoration |
| library_name: mlx |
| --- |
| |
| # BOPBTL-scratch-detection-fp32-mlx |
|
|
| The scratch/defect **detector** from Microsoft's *Bringing Old Photos Back to Life* |
| ([microsoft/Bringing-Old-Photos-Back-to-Life](https://github.com/microsoft/Bringing-Old-Photos-Back-to-Life)), |
| converted for MLX: grayscale photo in → per-pixel damage probability out. |
|
|
| - `UNet(in=1, out=1, depth=4, conv_num=2, wf=6, batch_norm, up_mode="upsample", antialiasing=True)`, |
| 37,626,177 params, fp32 (150.6 MB — the released 1.90 GiB checkpoint minus its Adam state). |
| - Keys are the original torch `state_dict` names (NCHW conv layout), including BatchNorm running |
| stats and the BlurPool `filt` buffers — consumers transpose per their framework's convention. |
| - Reference behaviour: short-side-256 (/16-aligned) BICUBIC resize, grayscale, Normalize([.5],[.5]); |
| `sigmoid(out) ≥ 0.4` is upstream's mask threshold. |
|
|
| **Swift consumer:** [`xocialize/mlx-bopbtl-swift`](https://github.com/xocialize/mlx-bopbtl-swift) — |
| parity-locked at **119.8 dB worst** vs fp32 PyTorch (and the reason its forward runs on the CPU |
| stream is documented there: a measured Metal fp32 conv2d divergence window, not a preference). |
|
|
| ## License |
|
|
| MIT — the upstream LICENSE file, all source headers, and README §License (which extends MIT to |
| "the codes and the pretrained model" verbatim). The training dataset was never released; the model |
| cannot be retrained or audited. |
|
|
