DDColor-tiny — fp16 ONNX (512×512)

An ONNX export of DDColor-tiny (piddnad/DDColor, ICCV 2023) for black-and-white / grayscale image colorization, exported for in-browser inference with onnxruntime-web.

This is the model behind the Edge Tools Image Colorizer (/colorize-image), which runs it fully client-side — images are never uploaded anywhere.

Files

File	Size	What
ddcolor-tiny-fp16.onnx	~130 MB	DDColor-tiny, float16 weights, float32 I/O, fixed 512×512 input

Tensor contract

	Name	Shape	dtype
Input	(first input)	(1, 3, 512, 512)	float32
Output	(first output)	(1, 2, 512, 512)	float32

The weights are float16 but the I/O boundary is plain float32 (keep_io_types=True), so no fp16 tensor plumbing is needed and op support on the ORT wasm execution provider stays maximal.

Pre / post-processing

DDColor predicts chroma only; luminance comes from the source image. Using CIE Lab in OpenCV's float convention (L∈[0,100], a,b∈~[-127,127], D65 white, sRGB gamma — DDColor is trained against cv2.cvtColor float Lab):

1. Normalise input to [0,1] and compute the original-resolution L channel.
2. Resize to 512×512, rebuild a grayscale RGB from L (a=b=0, Lab→RGB) and feed it as (1,3,512,512) float32. No ImageNet normalisation (do_normalize=False).
3. The model returns ab as (1,2,512,512).
4. Resize ab back to the original size, concatenate with the original-resolution L, and convert Lab→RGB. Only chroma is low-res; full-res luminance is preserved.

Provenance

Exported from the upstream PyTorch checkpoint — reproducible from public sources:

• Architecture code: piddnad/DDColor pinned at 2adb63f2656ac41cbdf7b894cddd94121a3faf13 (basicsr.archs.ddcolor_arch, encoder_name="convnext-t", decoder_name="MultiScaleColorDecoder", num_output_channels=2, last_norm="Spectral", num_queries=100, num_scales=3, dec_layers=9).
• Weights: piddnad/ddcolor_paper_tiny pytorch_model.bin, SHA-256-verified (8a1277bc90a1bfbb6d2d83933a9a6bc821931879ca93e26e4fcec12165d41fce).
• Export: torch.onnx.export, opset 17, fixed 512×512 input, then weights cast to float16 via onnxconverter_common.float16.convert_float_to_float16( ..., keep_io_types=True). onnxsim / symbolic-shape-infer are skipped — the input is fixed-size, so shapes are already static.
• Validation: every published build is checked by an automated colorization test against this exact artifact before upload — a grayscale input must come back with real chroma while preserving the source luminance.

License

Apache-2.0, inherited from the upstream DDColor project (LICENSE).

Citation

@inproceedings{kang2023ddcolor,
  title={DDColor: Towards Photo-Realistic Image Colorization via Dual Decoders},
  author={Kang, Xiaoyang and Yang, Tao and Ouyang, Wenqi and Ren, Peiran and Li, Lingzhi and Xie, Xuansong},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2023}
}