Eesh, OK so my thoughts on this after watching the DigitalFoundary comparison video by Alex, who right now I consider to be one of the best people with an eye for detail and technical breakdown of visuals.
https://www.youtube.com/watch?v=xkct2HBpgNY
Right off the bat I think they do a better job than I was expecting, given the constraints of what they're dealing with. Single frame using essentially post processing upscaling without access to temporal frames or much other data. It's certainly a win over general purpose up-scaling techniques in terms of quality so that's very positive. It does seem like the focus is more on edges than something like texture detail and this make sense because edges of straight lines have patterns you can easily detect/predict, where as internal texture detail is complex and can vary so much that enhancing it with post processing alone is going to be impossible without some form of additional data (from DL or subsamples or whatever). So it really feels like more of an aid to anti-aliasing than anything else.
I do think that vs DLSS 2.x Nvidia has the edge most noticeably in the more aggressive modes, this is where the deep learning really helps clean up textures and not just edges. Where as the higher quality 1440p to 4k upscaling looks somewhat comparable between the 2 and seems to be the core strength of FSR right now.
This leads me into use cases, which is tricky...to me it was always obvious that DLSS was a sister technology to ray tracing. RT was the core goal of Nvidia and they knew that rendering at more than 1080p is an impossibility. Adoption of RT was only ever going to happen if gamers could maintain their lovely 1440p or 4k resolutions. So DLSS was really invented to ease RT adoption. That comes across in DLSS use cases, typically you're taking games with RT at an internal res of 1080p and getting them up to 1440p or 4k. It has good enough upscaling from 1080p at least with DLSS 2.x to do this. It seems to be a sensible trade off for I think most people because of the improvement RT brings. However I'm not convinced it'll be commonly used anywhere else outside of getting RT playable, I certainly don't.
This is where I think this'll be a problem for AMD and FSR. They did not push RT hard in the current gen instead going for rasterization wins, they wanted reviewers to basically avoid it and treat the cards like more traditional rasterization cards because the performance just wasn't there. But I do expect them to move towards more RT cores in the next generation to catch up with Nvidia and that'll mean a push to more wide spread RT adoption and tackling the same problem as Nvidia. But the weakness is FSR is that it's pretty bad at 1080p -> 4k in terms of final quality, barely above regular upscaling it seems. What I believe will become the most common use case and primary reason for FSR in the next gen will struggle to compete. I'd be really interested to see usage data for both DLSS and FSR (as it becomes adopted) and how many people are playing without RT enabled but using upscaling of some kind. What % of people and what use cases are most common. My gut feeling on this is that it's close to zero. But it would be cool to see some real data on this rather than just speculating.
I could have sworn that FSR was aimed at general support across all games? Maybe I'm not remembering that correctly? That would have been a much bigger win for AMD over DLSS. But that opens up the question of integration into games, as Alex showed in his video game engine specific upscaling can be significantly better for the same performance cost. So for example if you're making an unreal engine game why would you bother integrating FSR when native upscaling is better?
The hope is of course that they do what Nvidia did which is continue to increase the quality over time, we want competition in this space. However I have a sneaky feeling this wont happen...Nvidia got wins because they could keep training the ML model over time, but this kind of more basic post processing it seems harder to get decent wins. The same way we've not really seen something like FXAA or SMAA improve over the years.
