How to spot AI-colourised pictures.
It has come to my attention that some colourised pictures of Peter Cushing have been making the round on Tumblr. Like these two, for example:
I saw those on my feed a few times and, yes, they look quite nice to be honest.
However, you should be aware of one important thing: Those pictures were colourised with generative AI.
Note: Those pictures aren't mine. I don't know who actually created them, but they've been circulating on social media for a while now.
Let me explain.
For those who don't know, I've been posted colourised pictures for a while on my blog. I colourise them using Photoshop's neural filters. I manually pick the colours and apply them on an picture that I then touch-up with the mixer brush. Colourising a picture could take from 20 minutes to a couple of hours, depending how much work needs to be done.
The results isn't always perfect, but if I like how it came out, I'd post it on here.
This is a process that I find quite enjoyable, to be honest.
While neural filters may use AI, they don't work the same way as generative AI, in way of colourising a picture.
What's the difference, you may ask? While my methods use a lot of manual work, images colourised with generative AI are often either done in one click — or tap — or are entered into an AI model with a prompt asking the image to be colourised.
The thing is generative AI doesn't colourise pictures. It actually generates a colour version of the submitted image.
While the results may look good, it's not — so to say — the same image with added colours.
Also, generative AI will "enhance" the colourised images, and by that I mean inventing details that don't exist in the original images.
Many people I know to share those images are quite opinionated on the subject on AI, and are often not aware that what they share was modified with AI.
Therefore, I've decided to share with you all a few pointers to spot AI use in those images.
Let's start with example #1:
Here's the colourised version and the original image side-to-side.
Already, some of AI usage are quite evident to my eyes: - The lighting is the colourised version is different, more dramatic, than the original's. - The contrast is stronger in the colour version, to the point that we can no longer distinguish the lapel of Peter's jacket that was quite visible in the original's. - The colour version has this unnatural smooth texture while the original has high level of JPEG artefacts. - The colour version has sharper details: more wrinkles visible, individual hair, eyelashes, crisp eyes, the lips are even more defined. The list could go on.
Now, here's a split version of the two images:
The difference in details is quite noticeable. AI even interpreted JPEG artefacts as being wrinkles.
Just look at the fingers. See those lines going along them? Those are JPEG artefacts. Also, notice how the knuckle in the colour version has creases while none are visible in the original.
Now, here's example #2:
The colourised version looks better here because the original picture was of quite good quality.
However, here's a couple of signs of AI use:
Peter's left eye (on the right side of the image) in the original was in a shadow, making it look all black with no visible iris. The colourised version suddenly shows both eyes having visible irises. However the pupils are misshapen, more in the left eye than the right. Also, the lights reflected in Peter's right eye are now different in the colourised version.
Then, there's the texture of the jacket's fabric. In the original version, it is surprisingly well-defined. You can see the weaving clearly. However, the colourised version lose all definition and looks rather, well, messy. Seems that the AI couldn't interpret the original's weaving correctly, and created whatever this is.
Now, here's an example of AI-enhanced image:
While it's not a colourised picture, I thought it deserved a closer look.
Already, I can see how smooth the enhanced image on the left is, in addition to the colours being more vivid — those blue eyes now look more disarming than ever. The hair is also quite detailed and the wrinkles are more pronounced — AI is always strong on those.
Now, here's a split version:
We've gone for low-quality screen capture to something trying hard to look like 4K, and I thought the original image was already of decent quality. I must admit that the enhanced version does look good though.
Here's a close-up of one of Peter's pretty blue eyes. Suddenly, the enhanced version shows creases on the eyelid and we get more crow's feet — AI is also strong on those. AI even attempted to generate a caruncle — which I had to look up what it was called.
In conclusion.
I must re-iterate I am no graphic expert, and everything I said above should be considered an amateur opinion at best.
Depending of the quality of the original image, and the AI model used, some colourisations can look so good that it's hard to tell if it's a genuine colour image, or if it was AI.
Even experts could get fooled at times, and AI models are getting better and better at enhancing/colourising images, to the point that it will end up looking authentic.
I wrote this post just so anyone reading it may be more informed into spotting the most obvious signs that an image may have been colourised/enhanced using AI and whether they want to share it or not on their blog.
Hope it helped. Have a good day.
— Xennial Warrior Princess
