Hey Everyone, i would have loved Topaz to do an AI-based 3D conversion, as they certainly know how to process frames and their AI Models not only enhance within a single frame but use the information of adjacent frames to extract structure information/moving objects etc…
HOWEVER… take a look at https://www.owl3d.com/
One or two years ago these guys came up with some really rudimentary AI-stereo converter, which evolved a lot during the last months. It seems to be based on models for monocular depth estimation (like Midas / Leres) - which flicker a lot when processed frame by frame - an add a post process for temporal consitency to that so that eventually you get a smooth and really good depthmap and left/rigth views. It wokrs perfect for movies wit a lot of close up shots and a lot of movement (like any action movie) but struggles with more detailed scenes. And of course, a depthmap cannot represent transparencys or blur - its just like projectiong the image on a deformed rubber surface - but good enough to get a decent 3D effect. I dont want to advertise for OWL3D, but as Topaz guys, i would definitely say “Challenge accepted” as there is a lot potential to extract structural information/image composition form a motion picture and get much better sharper depthmap and object separation than just “temporal smoothing of a series of monucular depth estimates”.
