README.md

metadata

license: apache-2.0
library_name: onnx
pipeline_tag: image-to-image
base_model: piddnad/ddcolor_paper_tiny
tags:
  - colorization
  - image-colorization
  - onnx
  - onnxruntime-web
  - int8

DDColor tiny, ONNX 512x512 (int8)

An ONNX export of DDColor paper-tiny for automatic image colorization, built for in-browser use with onnxruntime-web. It powers the Colorize neural filter in the ToolsMonk Photo Editor, where it runs entirely on the user's device.

Provenance

• Model and code: piddnad/DDColor (Apache-2.0), weights ddcolor_paper_tiny (piddnad/ddcolor_paper_tiny).
• Architecture settings match the upstream scripts/export_onnx.py for model_size=tiny: ConvNeXt-T encoder, MultiScaleColorDecoder, 100 queries, 3 scales, 9 decoder layers, input 512x512.
• Exported with torch.onnx.export, opset 17, fixed input shape [1, 3, 512, 512].
• Weights of MatMul and Gemm nodes quantized to uint8 with onnxruntime.quantization.quantize_dynamic (fp32 export 257.7 MiB, this file 107.5 MiB). Mean absolute difference against the fp32 export on a test image: about 1.1 ab units.

File

File	Bytes	SHA-256
ddcolor_tiny_512_int8.onnx	112,723,138	e983347272b8e218edc974a0af2bd66f53b73212abb800e103c64806e4c14a14

Input and output

• input: float32 [1, 3, 512, 512], RGB in 0..1. Feed a grey image: take the CIELAB lightness L of the photo, set a = b = 0 and convert back to RGB.
• output: float32 [1, 2, 512, 512], the predicted CIELAB a and b planes.

To colorize, resize the predicted a/b planes to the original size and combine them with the original photo's L channel, then convert Lab to RGB. Detail and brightness come from the original photo.

Licence

Apache License 2.0, the same licence as the upstream DDColor code and weights. See the upstream repository for the paper and citation:

Kang et al., "DDColor: Towards Photo-Realistic Image Colorization via Dual Decoders", ICCV 2023.