AI-powered software that can automatically colourise old black and white photos exists, but it’s often far from perfect. In comparison, manually colourising an image in Photoshop yields stunning results, if you have a lot of time and impressive skills. But a new app, developed at the University of California at Berkeley, cleverly merges both approaches so it’s easy to accurately colourise a black and white pic.
Photos: Richard Zhang
[referenced url=”https://gizmodo.com.au/2016/04/this-software-creates-vivid-color-pictures-from-black-and-white-photos/” thumb=”https://i.kinja-img.com/gawker-media/image/upload/t_ku-large/ftoml2zd98aei2rnzcaa.jpg” title=”This Software Creates Vivid Colour Pictures From Black-and-White Photos” excerpt=”Want to inject some colour to your photographs in a hurry? Well, new software can take an alarmingly good guess at what a colour version of your black-and-white photographs may look like.”]
A year ago, Richard Zhang and his team at the University of California at Berkeley demonstrated a neural-network powered app that could automatically add colour to black and white photos. After studying millions of images for reference, the app had learned how various features in colour photos were tinted, and could apply what it learned to images it had never seen before. The results were often impressive, even if the app got all the colours wrong.
To deal with those mistakes, Zhang and his team have now developed a follow-up app that uses the same highly-trained convolutional neural network to automate the process of adding colour to a black and white photos, but with the addition of user-added cues and hints that help the software, and the underlying algorithms powering it, produce more realistic results.
The app starts by taking its best shot at colourising a black and white source image it’s been fed, and in the process it creates a small palette of suggested colours. The user can then refine the colourisation process by adding simple colour markers to the source image. For example, they can specify a plant is a certain shade of green, or choose a specific skintone colour for a person, just by placing tinted dots right on the source image in the app.
Photos: Richard Zhang
The neural network continues to work away in the background, updating the colourised results every time a new marker is added by a user. So instead of having to spend hours in Photoshop painstakingly painting in hundreds of different shades, the colourisation process happens in just minutes. It’s intelligent enough to limit where colours are applied, so a handful of dots are all that’s needed to bring an image to life.
Want to try it for yourself? There’s no word on when or if Adobe plans to include capabilities like this directly in Photoshop, but Zhang has made his Interactive Deep Colorization app available for download on GitHub if you want to take a stab at colourising some cherished photos of your great grandparents.