>>511181262
I've not exhaustively tested this, but I tried a local LLM (on my computer) to test image classification, and I was not impressed with the results. Like all LLM the result is suggestive, and often right. But also often wrong.
That might be okay for tagging or writing image descriptions for personal use, but it's not suitable for moderating a large scale forum where your ass is on the line.
That's probably one reason why serious players use a cloud service because they can have larger and more capable models.
However, this may change as local LLM improve, or if someone releases a finely tuned CP detection LLM.

I'll also add there's some confusion about the technology on offer. First there are image hashes, which just computer the image hash and compare it to a list of known files. The problem with this is that images can be cropped or slightly altered to pass through the hash filter.
So MS created PhotoDNA, which uses some math to detect altered images. But these only work for known images, not new images.
So AI is the next step, because it can classify an image as
> child
> sex
Which solves the problem of identification.
The only remaining problem is reliability.

If someone has better information I'd love to hear it.