The script uses the Image Colorization API from DeepAi.
Here's the repo for the colorizing model itself: https://github.com/jantic/DeOldify
I don't think the DeepAi API allows you to adjust the model parameters, but you definitely can if you run it yourself from the repo above or a colab instance (e.g. https://colab.research.google.com/github/jantic/DeOldify/blo...). MyHeritage also offers this API as a service with tunable parameters.
If you want some quick results on your own BW images, just set your own DeepAi API key as an env var and run this script: https://github.com/skzv/colorize-photos
I can differentiate color mostly fine and have never mis interpreted grass as green, but clothing and fabric always pose problems, just as with your AI.
I’ve learned to be ok with things not looking real. Black and white is just as unreal to me as false color toning, but feels more honest and deliberate. The luma of the Armalite wielding guerrilla is amazing — it’s hard to beat black and white as a descriptor format.
An example: this same model colorized the Golden Gate white: https://camo.githubusercontent.com/3487a03626fd86d4f8c814636...
In my case I wasn't too concerned, since I was sharing the photos with my family, who understood they were artificially colorized. My father even asked me if it was possible to choose specific colours in articles of clothing, since it got those wrong. Nonetheless, to me and my family, it made our old photographs seem so close and alive, which made it worth it.
A good compromise may be to mark these retouched photos with a small watermark indicating that they have been altered. By the way, I think the colab instances running this model have the option to add watermarks for this reason (https://colab.research.google.com/github/jantic/DeOldify/blo...).
Did you figure out a way to achieve this other than repeat attempts or recoloring in post?
Is there a setup that retrains with user input such as by inputting an image with an object manually colored the model will then recognize it in later input and color similarly?
It seems like this could also help with persistence of colors in videos, training with the previously completed frame step before producing the next.
Panchromatic vs orthochromatic was still a choice even in the 70s. Spotmatics shipped with an ISO marker dial that included color, ortho, and panchro, 30 years after the Golden Gate Bridge was built.
Ortho bw — cheap 1930s snapshot — would show a red bridge as white, but if it was fancy panchro then I’d be more inclined to agree with the AI, given the luminosity.
Cambodia has called on US media group Vice to withdraw an article that featured newly-colourised photographs of victims of the Khmer Rouge, saying the images are an insult to the dead because some had been altered to add smiles.
“We urge researchers, artists and the public not to manipulate any historical source to respect the victims,” [Cambodia’s Ministry of Culture] said.
and what about those space photos?
It’s even a common trope in modern films to use black and white when showing flashbacks to time periods around the beginning of the 20th century. But that’s not how the world looked! It’s a cheep and deceptive tool imho.
That limit materially constrained the information that could be captured, but it didn't invent false information. It was a dimensionality reduction from color to simple brightness intensity. Sure, the information captured was through that lens/filter, but there wasn't some AI inventing/guessing at what it was seeing. The photons directly exposed the chemicals in the film. They are not equivalent at all, and your reduction based upon the fact that humans add an additional interpretive perspective to each is a stretch.
I say all of this while being fine with colorized photos, but they should be accompanied by the original photo and the disclaimer that they were colorized. I think colorized photos can add a lot of immersion and trigger emotion, but so does generative, interpretive art.
We shouldn't treat Monet's water lilies as an accurate portrayal of his flower garden at his home in Giverny. Colorized photos, while not as extreme as impressionist paintings, are still a generated piece of art. Perhaps they are more akin to portraits, which have been historically shown bias towards an unflattering view of the subject -- as when they are unflattering, they tend to not to survive
This makes my eyes roll like I'm a professional golfer. I love B&W, and even without using B&W film, I will still turn certain color images I take into B&W. There are lots of artistic reasons, but most of the time it just feels right. Sometimes, it's just an aesthetic reason of matching the decor of where the print is going to hang.
>I mean too many people literally picture the world at the time to be black and white.
So what? There are also people that believe the world is flat. There are plenty of museums full of paintings from well before photography was invented that clearly show color. Anyone that believes that the world was in B&W before color film just need to be walked away from as there's nothing but frustration there.
It is a machine learning model, so not really.
Instead, the model makes an educated guess of the most likely color, based on vast amounts of pictures sampled un-uniformly from very different areas of the world.
But they've alway wanted colors in photographs. Going as far as actually hand coloring them. This is kinda like that but automated.
I wish there was a little symbol you could put in the corner to indicate "colorized" or "altered" for composite photos done with photoshop.