AI photo rendering technology may be having problems with its use in the world of stock photography and photo art competitions, but AI’s certainly not lacking for practical uses in human-created images.
Palette offers an example of exactly this that many photographers and photo restoration professionals might find useful. The web-based tool, which is free to use, lets you upload any black and white shot and colorize it with different filters.
Palette’s creator is in fact calling it the “DALL-E of Color” because of his trust in its quality. DALL-E is a well-known AI image rendering platform that has recently been heavily reported on in the news for the artificial photos it can generate.
Palette’s creator is Emil Wallner, a French artist-in-residence who operates at Google and who shared Palette earlier in September through Reddit so that internet users could give it a spin.
According to Wallner, the technology under Palette’s hood uses “unreleased research” that works off text-to-image technology similar to that used in famous AI-image rendering programs. These include DALL-E, Stable Diffusion and Midjourney.
Wallner has already gained recognition with previous work he did on colorizing the famous 20th-century painter Gustav Klimt’s old faculty paintings with AI technology.
Palette’s release to Reddit quickly went viral and the original post garnered over a million views, thousands of upvotes and hundreds of comments. Over 100,000 users also tried it with their own black and white photo samples. You can too if you like, right here on the Palette website.
The artist/AI developer shared it with a simple description: “I made a new and free AI colorizer tool. Colorize black and white photos and with smart filters and words.”
He also shared a sampling of his own colorization work that readers could check out on his Twitter feed.
According to Wallner, “I’ve been creating AI colorization models for five years, both as a hobby and in a professional capacity. I find it fascinating for many reasons. I’m curious what black and white photos look like in color. You can often see new details and the general mood of the image changes,”
To be honest, the colorization tool does indeed seem to deliver decent results. A quick look through some of the samples submitted by both Wallner and other Reddit users shows more than a few black and white photos of all types and ages that have been colorized reasonably well.
Here’s one example from Palette’s own archive:
Also, here’s a sample of a shot from this writer’s own black and white photography colorized through Palette. Close, but just a bit off:
It’s worth noting that historians have pushed back against this practice for its supposed harm to historical accuracy, and yes, it’s true that the AI tool can’t always be certain of accurately placing historically correct colors into images.
Wallner however disagrees with it being a problem and claims his tool offers superior results:
“Most AI colorizations average the colorization resulting in mostly brown colors and purple and red tones for say clothes and cars. With the advancement of text-to-image AI technology, I saw an opportunity to both make more vibrant and dynamic colorizations, but also enable people to edit the colorization,”
With that said, Wallner also explained that he understands how professional human colorization is better and the research involved in making it accurate is something AI technology can’t yet replicate:
“While you can make educated guesses based on statistics, much of colorization is doing research and artistic choices. The research required to make historically accurate colorizations still requires context and manual research.”
Palette doesn’t just add color to images either. It also has enough “intelligence” to adapt color to changing light conditions. This result is something that even the AI’s creator didn’t train it for and didn’t know the AI would achieve it.
Anyone can try out Palette and it’s compatible with most browsers. Wallner also notes that most colorized images are deleted from the site’s servers soon after being submitted for treatment. He further states that users keep the copyright for their own colorized results.