README.md

---
license: apache-2.0
pipeline_tag: image-to-image
tags:
- colorization
- unet
- pytorch
- safetensors
- onnx
datasets:
- johnowhitaker/imagenette2-320
- detection-datasets/coco
---
# Mini U-Net Colorizer — broader-data trained candidate

**Status: evaluated app-testing candidate.** Some broad color patches and
incorrect object hues remain. Predicted colors are not evidence of original
historical colors.

This checkpoint has 3,968,892 learned parameters and 236 fixed color bins.
It starts from the audited bin-mapping repair of main commit
`6c47ea40724d8fcd67d4f36ce837dc1cb5b1b2a8` and changes all 65 learned parameter
tensors. The color vocabulary is unchanged. The selected weights are update
748 of a completed 1,122-update BF16 L4 run using a 17,325-photo mixed training
pool (11,943 Imagenette plus 5,382 COCO), batch 32, initial learning rate 1e-5,
weighted classification loss and frozen BatchNorm running statistics.

Selection compared Imagenette50 and reserved COCO200 validation images.
The selected model retained color strength better than spatial-loss candidates.
It was then scored on separate Imagenette200 and COCO-val100 checks.

| Test sample | Previous repaired error | This release error | Fine excess-edge reduction |
|---|---:|---:|---:|
| Imagenette 200 | 13.318 | 12.847 | 85.8% |
| COCO-val 100 | 15.266 | 14.532 | 84.3% |

Error is mean Lab chroma distance. The release includes guided8 decoding;
raw learned weights alone improve error by 1.97% and 3.25%, respectively.
Excess-edge reductions are proxies, not counts of visible blotches removed.
COCO is a convenience sample; older upstream training exposure is unknown.

## Use the complete pipeline

```bash
python -m pip install -r requirements.txt
python inference.py --model . --output-dir colorized photo.jpg
```

```python
from PIL import Image
from model import load_model
from inference import colorize
model = load_model('.')
colorize(model, Image.open('photo.jpg')).save('colorized.png')
```

Defaults: temperature 0.38, guided radius 8, epsilon 0.001, one network pass.
The wrapper preserves aspect ratio and original luminance. Old app code that
only loads safetensors will not automatically gain guided filtering.

For ONNX without PyTorch:

```bash
python -m pip install -r requirements-onnx.txt
python colorize_onnx.py --model colorizer.onnx --output-dir colorized photo.jpg
```

The 15.9MB ONNX graph includes the model and guided decoder, with dynamic
batch/spatial sizes and verified PyTorch parity. See `DEPLOYMENT.md` for the
Lab input contract, CPU timings, publication commands and integration limits.
See `RESEARCH_ROUND2.md` and `reports/round2/` for complete measured evidence.

## Limitations

Smoothing removes fine color fluctuations but can suppress true small color
details, especially without luminance boundaries. Semantically wrong hues
remain. The coffee and rocket failure examples are retained. Browser/mobile
performance, video consistency and general production quality are unvalidated.
The separate experiment bundle contains all runs and reproduction code; GPU
optimizer state is not included. This is a weights-only continuation point.