Hacker News Reader

That could be the VAE? The "latent" part of latent diffusion models is surprisingly lossy. And however much the image is getting inpainted, the entire thing gets encoded and decoded.

Edit: I'll note some new models (SD3 and Flux) have a wider latent dim and seem to suffer from this problem less.

AI generated images are also biased strongly towards medium lightness. The photographer's adjusting of the tone curve may simply give it that "look".

View on HN

You absolutely do not need to encode and decode the whole image, even in ComfyUI. All you need to do is composite the changed areas back in the original photo. There are nodes for that and I’m sure that’s what Adobe does as well, if they even encode in the first place. These tools don’t really work quite like inpainting. There’s no denoise value - it’s all or nothing.

I’ve used Photoshop’s generative fill many times on singular images and there’s no loss on the ungenerated parts.

Reply on news.ycombinator.com