And again about Dion 2x models and Armitage source:
Original frame:
Processed:
As you can see bottom field (right one) is just sharpened and top field is different (artifact lines appears near steps).
So looks like top field is reconstructed by network. But it’s done wrong.
And strange thing is that next frame in output is same.
Top field with artifacts and bottom is just sharpened.
As I understand Dione should take original frame and in output make 2:
- Top field from original and bottom reconstructed.
- Bottom field from original and top field reconstructed.
But it is not so.
And my opinion is that Deinterlace networks should not denoise/sharpen/enlarge. Just deinterlace.
And we need ability to chain its processing with enlargement networks like artemis.
