From what I understand (and from testing a few tools out of curiosity), these systems don’t really “remove” anything. They rely heavily on trained neural networks that predict what might be behind clothing based on patterns learned from huge datasets. It’s closer to probabilistic image synthesis than editing. The model analyzes body structure, lighting, pose, and context, then generates a new image that statistically fits those inputs.
I tried reading explanations on platforms like clothoff and similar sites, and they usually hint at diffusion or GAN-based approaches. What surprised me is how sensitive the output is to pose and image quality — even small shadows or folds in clothing can throw things off. Also, people often assume these tools are accurate, but in reality they’re guessing, not revealing anything real. That misunderstanding is probably where most controversy starts.