@Guillaume.Abadie
Since we’ve last spoken I’ve done a lot more research on temporal solutions/designs have come to a rather fond of unexplored territory regarding a couple of requisites TSR doesn’t follow.
The problem in the industry is TAA/Temporal upscling is either too expensive or to poor visually. I’m not sure if you have ever heard of the term “Circus Method”. It’s derived from a joke among gamers regarding the hoops in order to achieve good image quality but is now one of the major methods. An example of this is say 1080p DLAA is going to be a lot more poor in motion vs 4k dlss performance(1080p) becuase of a 200% buffer. You can do the same with XESS 3.1 and you’ll find the same results.
The reason why is becuase a 200% buffer can hold pixels in a more appropriate place before downscaling, this mitigates OLPF(blurred) edges on moving objects. TSR uses too many frames becuase it’s reconstruction derives from getting that information from past frames via multiply jitter sequences(and now history resurrection). More jitter=More past frames=More chance to screw up and more to computation to keep that long array of information in check/from ghosting
Recently, DLSS and XESS have become much better in motion becuase they throw out past frame information much quicker and researching their output, their models have basically reverted to spatial upscalers but since they are not hard coded spatial upscalers and just AI based, they suck up too much power.
A better approach that would hurt performance less and give better results in motion(more important than stills) would to jittering conservatively to only sample thin undersampled objects and specular. You can do this with one frame using the Decima Jitter and use SMAA for stair aliasing edge detection. Then spatially upscale, on the current frame checkerboard the spatial upscaling influence from the last frames upscaled frame.
Unreals TAA/U, FSR2/3, also has the problem, it allows to many frames and we have no variable that can control how exponentially(currently linear) fast past frames are translucent over time. And that’s becuase the jittering is abused for 4 different issues(thin objects, specular, stair edges, and pixel crawl).
Also we understand not every project or user have the same needs as the engine default. Even Fortnite have some tweaks due to its competitive nature.
Yes, I and I find that no game using TSR or unreals TAA proving these are okay software designs. They have serious flaws that have been neglected for years and other rendering aspects have been abusing these flaws for years as well. This creating a false illusion of optimization and I and thousands(not on this site) of others are tired of current gen games produced with unreal being built around broken AA/upscaling.
DLSS is not an viable option for industry even though at the moment circus method combined with the newest dll is arguably the best temporal solution in motion, it’s a brand feature and a overrated one.
I will not being using Unreal as Epic provides it, there are too many problematic shaders and designs and no one on the Epic Developer teams seems to even care about this giant AA issues(unless your willing to review the topic with me).
TSR as of now, brute forces through issues of TAAU. Maybe TSR is viable on console due to the extremely lower cost, but like I said it’s still going against requisites for motion clarity making it a poor choice to what is possible.
EDIT: Again fix the documentation on Nanite Epic, I can have a landscape of small triangles grass and nanite will blow it up, nanite only helps scene with insane overdraw or billion meshed scenes: