it would be great to have a filter where you can autodetect or use a mask to select a logo and it does some inpainting to remove it.
could also be used for removing object in a scene.
It looks great, but implementing it should be more complex than any of Topaz’s existing models.
I think this is a great idea, we are all very much waiting.
Fairly simple to do with free software for plain or simple backgrounds - such as the before/after clips shown by commercial software that claims to do this.
But as soon as you have a dynamic background involving say moving people or animals behind the watermark/logo, or a camera panning across a complex background, or low quality video, it becomes a horrendously difficult task because the human eye is sensitive to unnatural frame-to-frame image changes. So you get a kind of flickering effect when inpainting guesses wrong, which happens a lot. If you then try to stabilize the footage, it becomes even more noticeable.
Non-blurry watermark removal is fine for many photo images via say photoshop, but for video, it’s problematic on many kinds of material and unless it works on all kinds of material, many people would complain like mad.
The best I’ve seen is in Borisfx Silhouette but that costs over a thousand pounds/dollars. And even then, you need to choose your material/clips carefully. I’d say there’s no chance of Topaz being able to develop good quality watermark/logo removal over a wide range of video types any time soon - though in future years, AI may eventually come to the rescue.
For “moving background” watermark removal can be more effective task than the same task for still image. Just you have more information from footage and need no “guessing”. Watermarks usually just “pinned to place” so recognizing of such elements can be also “intellectualized”. Like in case of “Motion Deblur” scene changes detection required.
@falkor yes I agree that for a ‘fixed’ moving background such as a camera pan of an empty street scene (the parts under a logo say) or distant landscape you can do that already, Mocha Pro has done it for years of course.
But there has to be some good useable footage from areas outside the watermark and even then, anything dynamic such as people, animals, foliage blowing in the wind (especially if in the foreground) will break it because successive frames across the entire watermark are constantly changing, often in an unpredictable way.
So yes, a smallish logo can sometimes be invisibly erased for a short scene but the chances of doing that for an entire movie are slim to none, I’d say. The logo (or some blur or defect) would annoyingly come and go.
As for a big watermark such as the huge flickery Topaz one, your chances of success are even lower on most videos and all movies. Topaz know what they are doing watermark-wise, and I’d say AI is at present nowhere near being able to solve the problem other than in a few particular scenes, the kind usually shown by firms claiming to do this. Stable Diffusion when developed for video could probably be trained to replace logos believeably but like face restoration, it wouldn’t be what was originally filmed.
