I’ve been ignoring this for a while - primarily because this is a bug that I liked - but now it is negatively affecting my video processing.
Iris’ facial recognition is still very active even when processing animation. For the most part this has been amazing - I’ve had to enhance scenes where the characters have very small, low-detail faces, and Iris was the only model capable of capturing the details on the faces.
Iris adds a surreal level of detail to animated faces that are already very human-like. Here is an example:
Iris perfectly captured every detail in the character’s… well… iris.
But, many cartoons are not so anatomically realistic. Here is an example of a character whose eye has no pupil and consists entirely of an iris - which is common in Western animation.
The eye is a single-hue oval, but Iris tried to add details that aren’t there - and it looks terrible. In other cases, Iris has actually tried to draw pupils inside of the iris.
The video AI struggles tremendously with animation, and this is just one example. I typically use a different AI application for animation, but I would like to bring this to the team’s attention.
That post doesn’t say that Iris is exclusively for live-action. However, it does say this:
Iris is designed as a flexible model for all input videos, so it works well on both interlaced and progressive footage. In fact, we’ve found that it will often produce better results than the specialized Dione deinterlacing model, even for interlaced videos without faces.
Even Topaz has acknowledged that Iris has more uses than just facial recovery.
Regardless, the fact that the facial recovery operates in exactly the same way on animation as it does on recordings shows that the model is capable of handling faces in cartoons. Generally, it handles them perfectly. But occasionally, it produces issues like the one described above, as it does in recordings (see the many complaints about Iris producing scary faces).
This is just another opportunity for Topaz to improve their product and make it even better for consumers.
The majority of the models available in Topaz Video AI have not been trained for this use case, however, we do have one model that was created to handle computer generated content. Gaia CG which will be found under the Enhancement filter
