My idea is that there should be an option for people who have a dataset of high quality images of a specific person, in a video, to feed that dataset to an ai within video ai 7 that works in a similar way to deep fakes in that the ai instead of drawing information from the pixels in the original video to make a face. It trains on the dataset and applies what it knows from that dataset to the video and I think it should work in tandem with starlight since starlight is amazing at enhancing backgrounds, clothing, etc. Basically starlight enhances the footage then presents an option for you to train a face model based off a pre trained model so the ai already knows what a face is then you feed it a dataset for a specific person in the video. I think this would help increase accuracy for finer details like freckles, scars, eyes, noses, pores, etc. Starlight is amazing but cannot always pick up on smaller details, but if you could train the ai on a specific face so it can recognize patterns and specific details on a persons face I think it would create a more accurate result.
This is what i wanted some time ago Is it possible to increase quality of video based on image(s) via Machine Learning?
1 Like
Oh really? I hope something will come of it. I really think it’s possible with the current state of ai. If nothing else I believe deepfacelab is open source so perhaps topaz could still incorporate deepfakes somehow
i think they used this a long time ago successfully in the intro part
I don’t think any company out there is going to want to deal with the legal ramifications and public backlash of something like that. Even in communities dedicated to diffusion-based training they specifically make it a rule to limit themselves to do training on fictional characters (at least, in public).
1 Like
