Additionally, Elite offers both its “HMD Quality” setting, which multiplies the rendertarget size, as described, and hands this big picture over to your VR software stack, for prewarping for the lens down the line, etc; and a “Supersampling” one, which also renders as much larger an image, but downsamples it on its own, to the originally recommended size, before handing it over.
This gives us the opportunity to try and compare both approaches, and see how much sharper and dynamic a result is produced by the former, than the latter.
As for the question of diminishing returns; That depends on several factors, both what the physical resolution can ultimately represent, and software matters, including something as simple as how far away any given object is in the game, but one important one concerns how many samples the final warping algorithm (downsampling should be inherently included in this) will take from its large input image, per output pixel. -At some point it will be too costly to read and average every hires input pixel that fills the same space as the lowres output one, and whole pixels will be skipped (this is why mipmaps exists). Depending on the filtering method used, this should begin to occur somewhere not far above x2.0 (…or 400% SteamVR ss), so going far past this point does probably not give very much.
…that said: a bit of that aliasing (skipped samples) is not necessarily a bad thing when using a motion-tracked head-mounted display, because it can allow you to perceive a bit extra detail, as it comes into view from minute head movements. At one point there was even a situation where Valve changed their filtering to a highly optimised method which takes more samples, for a smoother (better antialiased) output, but there was a backlash from mainly Elite Dangerous players, who immediately noticed the softer and less dynamic image, and demanded a switch for reverting to the old algorithm, with its higher aliasing. (This switch should incidently do nothing for the 8k/5k, which has their own lens compensation warp stage, and do not use OpenVR’s one.)
EDIT: For my own, by the way: Normally, I’d say as close to the metal as possible; So for Vive: SteamVR ss; For Rift: Oculus “pixels per pixel display pixel”; For Pimax: Pitool quality; But given OpenVR’s current 4096 ceiling, I’m for the time being sticking with PiTool 1.0, SteamVR 100%, and using any rendertarget size modifier the game has (equivalent to ED “HMD Quality”, not its “supersampling”), or if it doesn’t have any: SteamVR ss up until just below the ceiling.
Very excited for the next build.