In this video from May 2024, I compare approaches to adding colour to black-and-white photographs. The examples include DDColor in ComfyUI, Photoshop Colorize, and experiments with Stable Diffusion and ControlNet.
This is a record of those tests. Software versions, custom nodes and model combinations change; the screenshots describe the workflows I used at the time.
What colourisation can tell us
A colourised photograph proposes colours that look plausible. It cannot establish the original colours of clothes, buildings or objects from a black-and-white image alone. Where other evidence exists, it should guide the editing and be recorded separately.
I keep the source photograph and distinguish cleaning and tonal correction from reconstruction and added colour. Examples of the different treatments are in my photograph-restoration project.
The video
Three workflow examples
Comparing DDColor models
The first workflow processes the same photograph through four DDColor model variants so that I can compare the results rather than accept the first output.
Processing a folder of images
The second workflow uses a file-path input to process a folder. In these tests the colours could vary between images, which limited its usefulness for a consistent sequence of video frames.
Recolouring with Stable Diffusion and ControlNet
The third screenshot records a separate recolouring workflow. Its checkpoint, VAE and ControlNet model need to be compatible; it is not interchangeable with the DDColor workflows.
Source and related work
The official DDColor repository provides the model documentation and code. The examples on this page describe my own workflow tests, not a current ranking of colourisation services.