News for hackers. Stuff that matters.
Re:South Korea has jailed a man for using AI to create sexua... by dragonwriter on Saturday September 30, 2023 @03:27PM (#37718868)
> There are guides on the darknet that teach you how to fine-tune on CSAM and model weights can be found quite easily.

There are guides on the regular net that teach you how to fine tune on anything, and there is nothing at all that would need to be added or altered in those instructions for fine tuning on CSAM material. I'll take your word for the results of your searches of the darknet for CSAM model weights; I've never done that, as I have no need for CSAM model weights or, really, any other reason to explore the dark net.

> You need to fine tune generative models on actual CSAM to get the most realistic looking images.

I’m not going to spend time trying to create simulated CSAM images to test this, but my experience with using Stable Diffusion in oother contexts says that’s probably not true, though its probably (ignoring the issue of sourcing training imagery) the obvious and straightforward way. But the ability of models to generalize and combine concepts means you don’t generally don’t need to fine tune on the specific combination of concepts you are generating images of. A general purpose model can often produce photorealistic representations of concepts its not trained on when conditioned with a LoRa trained purely on, say, anime representations of the concept.

> The prosecutors noted how lifelike and realistic the images were

Nothing in those articles refers to the darknet, CSAM-specific models, or fine tuning on CSAM imagery. It just says the guy was arrested for using an AI generator to produce imagery that resembled CSAM imagery that was never distributed, under a legal framework which had also allowed that for sexually explicit animations of fantasy subjects that resemble adolescents, and that the prosecuting authorities claim that the imagery was particularly lifelike. As support for your claim of requiring actual CSAM to fine tune to produce realistic CSAM-resembling imagery, it is perfectly circular, resting entirely for its essential premise on the conclusion it is offered to support.

> You can’t do that without fine-tuning on actual images of children.

Not all actual images of children are CSAM, and most models are trained on actual (non-CSAM) images of children to start with. Heck, there are common recommendations for prompt patterns to steer clear of generation with SD1.x-derived models (which probably apply to other models, excluding the ones that are so aggressively self-censored as to make it difficult to even intentionally produce nude imagery of adults, like the SD2.x base models) to avoid producing conditions in which models may combine concepts to produce CSAM-resembling images without user intent. [0]

[0] common ones including avoid “young”, “girl”, or “boy” (the latter outside of the established pattern on which many models are trained of using 1girl/1boy/2girls/etc. to identify number and gender of subjects without regard to age) with prompt terms calling for or leaning toward nudity, sexualized clothing or appearance, poses or other compositional features which would take on an inappropriate character with children involved, and, a fortiori, explicit sexual concepts.

There are no comments on this comment yet.