Let me break down Krea 2’s filtering mechanism piece by piece. At the end of the day, it’s just a few numbers—this breakdown is pretty clear. The author starts by saying this LoRA was written with AI assistance, and he doesn’t care, as long as it works. Here’s the analogy: Krea 2 is a diffusion transformer. Think of it as a long assembly line, where each station uses your prompt to push the half-finished image closer to your goal.
Between the text and the assembly line, there’s a little mixing board, where each knob controls the strength of some text signal—like style sharpness, anatomical correctness, or whether the content is allowed.
According to his analysis, the safety filter is hidden in two of those knobs. When it triggers a rejection, those two push the result toward something generic or output a censored version. What’s surprising is that this filter isn’t some massive, standalone moderation model like you’d imagine—it’s just a tiny string of numbers crammed into the diffusion model itself.
From a product perspective, that’s pretty interesting—it shows how lightweight the implementation of a so-called safety mechanism can be. Of course, this kind of stuff walks in a gray area. Use it or not, that’s on you—just don’t go doing anything against the rules with it.