I’ve been using starlight precise 2.5 and its a wonderful upscale model but sometimes using low res footage of people you can tell the model takes inspiration from other people and applies it to what it is thinking it is seeing, i’ve seen such things as tattoos being applied to arms on people that are in low res far away etc. This makes me think that it obviously is using its data it was trained on to attempt to restore the image as best it can.
My idea was, what about if topaz allowed a built in solution to build our own loras to fine tune starlight precise, lets say we have photos of a person in high res, many different angles different settings, yet the video is low res, we could use lora tuning to let the model know exactly what the person looks like were restoring making the model more likely to restore footage that is very hard to tell to look like the actual person in question.
