Just read a paper on diffusion models generating hyper-realistic images without explicit ethical safeguards. How do current frameworks mitigate risks like deepfake misuse? Or are we relying too much on post-generation filters?
Can diffusion models cross ethical lines?
👁️ 8 views💬 3 replies❤️ 0 likes
3 Replies
Just starting out with my first JS calculator here, and you're already talking about diffusion models 😄 Crystal clear now, I'll know that "post-gen filters" isn't about putting quotes in the code 😎
And again with the deepfakes? 😅 I can't even take a decent photo with my phone and now they're saying I'm the future of misinformation.
I’ve been messing around with Stable Diffusion and DALL·E for a while now, mostly for dumb memes and AI avatars, but every time I try to generate a face from a prompt that even hints at a celebrity, the servers flag it instantly—even if the result’s obviously fake. The open-source side is way scarier; you can literally drop a LoRA trained on Elon Musk’s leaked photos and the thing just spits out whatever you ask for.
My take: post-gen filters are way too late. We need pre-prompt watermarking baked into the weights so the model itself refuses «generate nudes of [X]» in the first place. Fine-tune the checkpoints with synthetic toxicity datasets before they even hit the download page, not after some lawyer’s team complains. And yeah, Apple’s on-device diffusion probably does this already for the «Imagine» wallpapers—good luck extracting the model without a Class Action lawsuit.