[AndrewThrowaway]

Out of topic question.

Would it be possible to combine machine learning with something like 23andMe and match 100 000 person photos with their DNA. Would we have an AI model which basically can tell how the persons looks like from a strand of hair?

Imagine finding some DNA on a crime scene. Witnesses can say that this person was a bit overweight so you slide some slider to the right a bit and get a pretty good AI generated portrait of a criminal.

How cool and scary that would be.

[bookofjoe]

>Recent advances in Forensic DNA Phenotyping of appearance, ancestry and age [July 2023]

https://www.fsigenetics.com/article/S1872-4973(23)00045-5/fu...

........................

>DNA Facial Mapping

https://www.policechiefmagazine.org/dna-facial-mapping/

........................

>How Australian police will use DNA sequencing to predict what suspects look like [2021]

https://www.theguardian.com/australia-news/2021/dec/08/how-a...

........................

>The Role of Hair in DNA Testing

https://www.hairphysician.com/the-role-of-hair-in-dna/

[starburst]

When playing with a custom LoRA (dreambooth / SDXL) trained on my brother’s photos of himself something interesting happened, when we used the prompt “as a women” instead of “a man”, the result looked incredibly similar to one of our cousin, so much that my parents asked how come she was now in the pictures generated.

[Gelomodns]

There are already features you can extract from DNA. 23andMe shows you your traits and confidence level depending on the research.

But you know just bring it to the real thing: With DNA, Video and ML you will be able to track and find that person through face recognition and probability automatically.

Hail skynet

[InSteady]

I'd definitely have to err on the side of scary.

Seems like more fodder to abuse minorities behind the guise of fancy tech and supposedly "objective" algorithms. "All [minority] look the same" isn't just something racists say, it is pretty well documented that many people struggle to discern distinguishing facial features of other ethnicities and consistently perform more poorly on facial recognition tasks involving races other than their own.

Now you have an AI generated photo to go by that further does some fudging and blurring of distinctive features... Can you imagine which races in America, for instance, would be subject to more false positives to receive police/prosecutor attention and ultimately be more likely to wand up with charges drawn against them unjustly?

If you are poor, having charges drawn up by the DA means a plea deal and time in jail or prison 9 times out of 10. If you have money for a good lawyer your odds are better, but overall are still significantly worse than if you don't have charges drawn in the first place, despite the fact that there is no evidence of a crime beyond circumstantial crap, maybe some testilying by the arresting officer (their own term, not mine), and a bogus facial match.

[largbae]

I think it's fundamentally possible but the training set needed might be so large that you're better off just actually cataloging the DNA of most humans and pulling up the picture of the closest match.

[lionkor]

And oh so biased