DDColor

What DDColor is

DDColor is a free image-colorization model for restoring color to historical black-and-white photos and recoloring anime-game landscapes. Its approach combines multi-scale visual features with learnable color tokens to produce automatic colorization. The repository provides paper, ModelScope, artistic, and tiny pretrained variants; its maker recommends ModelScope for images outside ImageNet and describes tiny as the lightest option. Local inference scripts and a runnable Gradio demo are documented, alongside inference guidance for Hugging Face Hub and ModelScope. The repository also links to Replicate demos and an API. Users can train with ImageNet or a custom dataset, and the project documents ONNX export with an example using the tiny model. Requirements are Python 3.7 or later and PyTorch 1.7 or later. The repository includes an Apache License, Version 2.0. One limitation noted by the maker is that colorfulness loss can create unpleasant blocks of color, including red artifacts.

Who it suits

DDColor suits developers or image-workflow users looking for a self-hosted colorization model, with local inference, training, and export options. It requires Python 3.7 or later and PyTorch 1.7 or later.

Getting started

1. Open the DDColor repository at https://github.com/piddnad/DDColor.
2. Choose a pretrained variant; ModelScope is the maker’s default recommendation for images outside ImageNet.
3. Prepare an environment with Python 3.7 or later and PyTorch 1.7 or later.
4. Follow the repository’s local inference instructions or run its Gradio demo.
5. For hosted inference, use the documented Hugging Face Hub or ModelScope instructions, or explore the linked Replicate demos and API.

Verdict

DDColor offers multiple pretrained options and documented local inference, training, and export paths. The possibility of color artifacts is worth considering for image-colorization work.