How can we improve Lavalite?

I’ve been experimenting with this quite a bit over the last few months, especially with portraits from the 1950s where the paper started to crack and the shadows get muddy. What I noticed when using undress ai is that the results depend heavily on how the original scan looks. When the image has visible grain but the facial structure is still somewhat recognizable, the enhancement tends to keep the person’s unique traits—like the slight asymmetry in their smile or the shape of the eyes. For example, I restored a picture of my aunt when she was around 20, and the tool brought back subtle cheek contours without over-smoothing. But when I tried the same on a severely faded childhood photo of my dad, the AI guessed a bit more than I expected, so the restored version looked younger but not exactly like him. So I’d say it’s realistic as long as the input has enough cues for the model to read; otherwise, it starts filling gaps on its own.

Jumping in because I’ve had a similar mix of outcomes. Sometimes the enhancement feels spot-on, especially with photos that still have a defined silhouette, and other times it adds a slightly modern look that doesn’t quite match the era. But for casual family archiving, it’s still surprisingly helpful. I usually run a few versions and keep the one that feels closest to the original mood rather than the “sharpest” one.