DDColor¶

Development line: project:ddcolor · thread ddcolor-development
Last event: 2024-01-15 · 1 dated since 2024-01-15 · Researched: 2026-09-04 · confidence: medium

What it is¶

DDColor is a PyTorch model from ICCV 2023 for automatic photo colorization.

• Monochrome photos: colorizes them automatically.
• Grayscale channels: recolorizes existing color images from them.
• Deployment: runs via local scripts, Hugging Face, ModelScope, Gradio, or ONNX export.

Development line¶

• 2024-01-15 — DDColor source repository and Colab workflow were linked. Camenduru published Colab and Gradio-Colab notebooks pointing to canonical DDColor code. This was a community wrapper, not a core-model release.

What changed¶

The two January 2024 routes were wrappers around the same model, followed by official inference options.

• 2024-01-15: Camenduru published Colab and Gradio-Colab notebooks pointing to canonical DDColor code as a community wrapper, not a core-model release.
• 2024-01-22: ModelScope's Old Photo Restoration Space used iic/cv_ddcolor_image-colorization behind an upload-and-slider interface, confirming a DDColor integration rather than a separate restoration model.
• Official updates added Hugging Face integration on 2024-01-26, a fix on 2024-01-29, ONNX export support on 2024-10-25, a non-BasicSR local inference refactor on 2024-12-31, and streamlined loading and inference on 2026-01-17.

How to use this¶

We can work with DDColor from the source repository and its Colab workflow since 2024-01-15. Verify the supported workflow in those resources before running it.

1. Create the documented Python 3.9 environment, then install PyTorch 2.2.0 with CUDA 11.8 and requirements.txt. The local inference route does not require BasicSR. — https://github.com/piddnad/DDColor
2. Start with ddcolor_modelscope for ordinary images outside ImageNet. Use ddcolor_paper only for paper reproduction, ddcolor_artistic for an alternate result, or ddcolor_paper_tiny when footprint matters. — https://github.com/piddnad/DDColor/blob/master/MODEL_ZOO.md
3. Run python scripts/infer.py --model_name ddcolor_modelscope --input ./assets/test_images. Alternatively, download the ModelScope checkpoint and pass its pytorch_model.pt through --model_path. — https://github.com/piddnad/DDColor
4. For Python integration, create a ModelScope Tasks.image_colorization pipeline with iic/cv_ddcolor_image-colorization, pass an image, and save OutputKeys.OUTPUT_IMG. — https://www.modelscope.cn/models/iic/cv_ddcolor_image-colorization/summary
5. Run python demo/gradio_app.py locally when an upload UI is needed. — https://github.com/piddnad/DDColor

Best practices¶

• Use ddcolor_modelscope by default for images outside ImageNet. Reserve the paper model for reproduction rather than general use. — https://github.com/piddnad/DDColor/blob/master/MODEL_ZOO.md
• Compare ddcolor_artistic when color artifacts are a concern, since its training omitted colorfulness loss. Use ddcolor_paper_tiny when footprint matters. — https://github.com/piddnad/DDColor/blob/master/MODEL_ZOO.md
• Review comics, low-resolution inputs, and noisy images carefully, because the model was trained on natural photos. — https://www.modelscope.cn/models/iic/cv_ddcolor_image-colorization/summary
• Do not use the hosted Old Photo Restoration Space as the current entry point. It returns a runtime error caused by an incomplete model configuration download. — https://huggingface.co/spaces/modelscope/old_photo_restoration

Superseded by this¶

• 2024-01-15 — Colab-only setup guidance: official local scripts, Hugging Face, ModelScope, Gradio, and ONNX routes replace it.
• 2024-01-22 — Hosted Old Photo Restoration Space demo: obsolete because it returns a runtime error.
• 2024-12-31 — BasicSR-dependent local inference: the current local script route does not require BasicSR.

Still unknown¶

• The recorded link text was inaccessible, so its editorial claim and intended comparison could not be verified.
• The sources do not state whether historical ModelScope ID iic/cv_ddcolor_image-colorization and current repository ID damo/cv_ddcolor_image-colorization are aliases or a migration. Do not assume they are interchangeable.
• No current local inference run, hardware measurement, or benchmark against a newer colorizer was verified.

Sources¶

Agent brief¶

• Subject: project:ddcolor, thread ddcolor-development, 1 dated events 2024-01-15 → 2024-01-15.
• Practical note: We can begin DDColor work from the linked source repository and Colab workflow since 2024-01-15. Verify the exact supported workflow from those resources before use.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.