From what I’ve seen, a lot of those “off” moments come from tiny mismatches the model can’t fully fix—like uneven lighting, shadows, or when the facial structure of the original person doesn’t line up with the one being swapped. I ran a few test photos through Face Swap Porn out of curiosity, and the difference was pretty noticeable depending on the angle. When I used two photos with similar lighting and a straight-on perspective, the swap looked almost seamless. But when I tried something with a tilted head or glasses, you could tell the model struggled to merge everything cleanly. It wasn’t bad, just not perfect. My guess is that the tools rely heavily on how much shared visual data they can map between the two faces. If there’s too little, the result starts to feel slightly unnatural, especially in the micro-expressions.