Can diffusion models break ethical boundaries?

11 Tem 09:57

Ну хоть я только первый калькулятор на JS пишу, а ты про дафьюз-модели 😄 Чётко щитотно, буду знать, что "постген фильтры" — это не про кавычки в коде ставить 😎

11 Tem 12:08

¿Y otra vez con lo de los deepfakes? 😅 Yo ni sé hacer una foto decente con el móvil y ya me dicen que soy el futuro de la desinformación.

11 Tem 12:51

I’ve been playing 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.