Twinning: A Simple Jailbreak That Bypasses AI Image Protections

Author here, thanks to whomever posted this originally. So I think there is a couple things at play here. Between the first images and the last images I had restarted my browser and didn't catch that the output size of the model was switched back to 1x from 4x, so some fidelity was lost, and second I did resize all of the images down before uploading them and the CMS I use also does some dynamic resizing.

But with that said you are right not all of them are great representations, I actually am surprised you didn't call out the turn 4 and 5 "Marc" images because those somehow really got off the rails, and I was surprised how well it reconverged on what Mark looked like once I switched the output 4x again.

Some of the weirdness with Elona comes from the gender change, if I did "twin brother" instead it would have likely had been even closer, but it wouldn't have been as outrageous to get peoples attention on the impact.

With all of this said, They could both have been more refined with additional prompting. My first attempt through actually used some well known US political figures, but I figured using those for a blog post in this political climate would be very polarizing so I thought this was a safer option. Those ones I did push through in 4x mode all the way through, and gave more Crescendo prompts to refine them further, and you are correct in some cases I was blocked, but by using the restartability that I described in the posts, with some restarts and prompt adjustments they were shockingly close.

Thanks for reading, and Cheers!