VEAI models with different FP precision cause different outcomes

My computer used to do really good VEAI upscales, so I know it’s not incapable. But recently, the app has been defaulting to FP16 models (I check the models directory after checking the “As Needed” option in model preferences for downloading). This is causing a noticeably less quality render (using Proteus). I can confirm it’s the model, because I manually downloaded the FP32 version with the exact same flags, and it looks like my old renders.

The only problem with the manual download, besides having to do it for any other models, is that it takes roughly twice as long to render. So why could I have the best of both worlds before - an average 0.67 sec/frame render time AND the same results?

1 Like