DDColor
github.com
- Free plan
- Linux
- API
- Self-hosted
What DDColor is
DDColor is a free model for colorizing historical black-and-white photos and recoloring anime-game landscapes. Its approach uses multi-scale visual features to tune learnable color tokens for automatic image colorization. The repository provides paper, ModelScope, artistic, and tiny pretrained variants. The maker recommends ModelScope for images outside ImageNet and describes tiny as the lightest option. Users can run inference locally through scripts or a Gradio demo; instructions also cover Hugging Face Hub and ModelScope, and the repository links to Replicate demos and an API. Training is supported with ImageNet or a custom dataset, and the repository documents ONNX export, including an example using the tiny model. The listed requirements are Python 3.7 or later and PyTorch 1.7 or later. The project includes an Apache License, Version 2.0. One limitation noted by the maker is that colorfulness loss can cause unpleasant color blocks, including red artifacts. The repository directs users to online demos and documentation but does not state a dedicated support service.
Who it suits
DDColor suits developers or image-processing users who want to run or train a colorization model locally or through the listed integrations. It may also suit users who need ONNX export or a lightweight pretrained variant.
Good to know, and what to watch
- Offers paper, ModelScope, artistic, and tiny pretrained variants.
- Includes local inference scripts and a runnable Gradio demo.
- Supports training with ImageNet or a custom dataset.
- Documents ONNX export.
- Licensed under Apache License, Version 2.0.
- Colorfulness loss can create unpleasant color blocks, including red artifacts.
- Requires Python 3.7 or later and PyTorch 1.7 or later.
- No dedicated support service is stated.
Plans
Every plan and priceDDColor open-source software
Free plan
The verdict
DDColor provides several ways to run, train, and export an image-colorization model. Users should account for its stated color-block artifact risk and listed software requirements.
What the maker publishes
- Export
- The repository documents ONNX export, with an example using the tiny pretrained model. source
- Integrations
- Inference instructions cover Hugging Face Hub and ModelScope, and the repository links to Replicate demos and an API. source
- License
- The repository includes an Apache License, Version 2.0. source
- Local use
- The README documents local inference scripts and a runnable Gradio demo. source
- Maker attribution
- The paper authors are listed as affiliated with DAMO Academy, Alibaba Group. source
- Method
- The model uses multi-scale visual features to optimize learnable color tokens for automatic image colorization. source
- Model guidance
- The maker recommends the ModelScope model by default for images outside ImageNet and describes the tiny model as the most lightweight version. source
- Model options
- The repository provides paper, ModelScope, artistic, and tiny pretrained model variants. source
- Notable limitation
- The maker cautions that colorfulness loss can produce unpleasant color blocks, including red artifacts. source
- Purpose
- DDColor colorizes historical black-and-white photos and can recolor anime-game landscapes. source
- Requirements
- The listed requirements are Python 3.7 or later and PyTorch 1.7 or later. source
- Security
- The repository page does not state security controls or compliance certifications. source
- Support
- The repository directs users to online demos and documentation but does not state a dedicated support service. source
- Training
- Users can train the model using ImageNet or a custom dataset, with training dependencies and pretrained ConvNeXt and InceptionV3 weights. source
Checked against the maker's pages on 4 Oct 2026.