There should be a new Iris model added that uses the same model that Photo AI uses for face recovery. The face recovery model in Photo AI is unparalleled across the Topaz Labs products and even the Iris 2 model in Video AI doesn’t come close to matching the quality from Photo AI. Futher adding customization of how much recovery to use in Video AI would be an added plus.
A few months ago a member of Topaz Labs said they are unable to do this.
According to the member at Topaz Labs, the recover face AI from Photo AI is either not temporally stable enough, or not fast enough, to make it accessible on currently generation hardware. As a result they resorted to a different method. Train their AI’s similar to how they normally would, but include a lot of face data so hopefully the AI will learn how to do face patterns better than their other AIs.
Note: I haven’t personally tested to see if the Recover Face AI is temporally stable or not. But it should be relatively easy to test.
Note 2: There are multiple ways to do face recovery. Topaz Labs seems to have taken two different approaches to this issue. But there are other avenues they could take that could result in large increases in quality.
I’m not a model developer so I wish it was as simple as this:
- Apply the static image AI model to a frame of video
- Before applying to the next frame, override the parameters using the ones from previous frame
- Create hysteresis by comparing results in both directions for the best result (but what is “best”?)
- Create hysteresis also for most similar results (hopefully that’s easier to define)
- Repeat for a “group of pictures”, i.e. a bag of frames
- Add trial & error on all of the above
- Worry about performance last, since it solves itself over years
But yeah it’s not just parameters; entire states exist per frame that exceed megabytes.
Turning a static image technology into a temporally-aware motion technology is insane!
