QVFR with VD SGSR

[Qcumber]

Has anyone tried using QVFR with VD SGSR? I set VD to "High" instead of my normal "Godlike" then used QVFR to increase the centre resolution by 1.5; peripheral to 0.5; vertical 0.35; horizontal 0.25.

Using DLAA the quality looks about the same as when using Godlike (QVFR centre x1.1; periphery x0.5) but with a significant performance boost. Overall this works very well and gives me a steady 72 FPS with QP. 

Here is my reasoning. As I understand it SGSR uses the headset to SS from High to Ultra. I presume that QVFR applies the multiplier to the "High" setting and then the headset SS the image immediately before it is displayed. Would this then mean that SGSR increases peripheral resolution to "Ultra" or would it be "Ultra x0.5" and what happens to the centre resolution; is this now "Ultra x1.5"? I would appreciate feedback from someone with a better understanding. 

[mrsylvestre]

Just tried that out of curiosity. As I still have a Pico 4 (no eye tracking), not a Quest Pro, 0.25 x 0.35 for the (fixed) centre is too restricted. I increased that to 0.45 x 0.45 which is acceptable with fixed foveated rendering and I could still get 72 fps with my setup. However, I get increased shimmering, even in the centre with resolution x 1.5, with High + SGSR. Because my GPU is AMD, I use MSAA 2x as antialiasing (no DLAA or DLSS). Switched back to QVFR defaults.

My thoughts about the whole process: downgrading from godlike (3072x3216 on the QPro) to high (2496x2592) but increasing centre resolution from 1.1 to 1.5 results in DCS rendering that part of the image at an about 20% higher resolution BUT... after rendering the periphery at lower resolution, it combines both at 2496x2592 (VD's "high" setting). This is then sent, at that resolution, to the encoder and streamed to the HMD where it is upscaled. Hence the 1.5 resolution increase for the centre is akin to supersampling (which is good) but in the end, less information is sent to the HMD, which has to make up for it by a relatively basic, low-computational cost upscaling algorithm (SGSR). With Godlike, the image passed to the encoder is 3072x3216 and inherently contains more info. To which extent this will translate into improved final image quality is affected by the streaming (and thus encoder) bitrate. If the bitrate is capped too low, the lossy compression will reduce (but not negate) the advantage of passing an higher resolution image to the encoder, but I doubt this would be the case in practice, with bitrates >= 150 Mbps.

Hence, assuming the above is correct and encoding/streaming shenanigans notwithstanding, my understanding is that SGSR upscales the picture from "high" to "ultra" resolution but starting from a picture that has been produced with 1.5 supersampling at the centre (kind of SSAA, really). For the periphery it boils down to something like high with 0.5 downsampling then upscaled a bit from high to ultra.

 

Edited March 8, 2024 by mrsylvestre

[Qcumber]

6 minutes ago, mrsylvestre said:

Just tried that out of curiosity. As I still have a Pico 4 (no eye tracking), not a Quest Pro, 0.25 x 0.35 for the (fixed) centre is too restricted. I increased that to 0.45 x 0.45 which is acceptable with fixed foveated rendering and I could still get 72 fps with my setup. However, I get increased shimmering, even in the centre with resolution x 1.5, with High + SGSR. Because my GPU is AMD, I use MSAA 2x as antialiasing (no DLAA or DLSS). Switched back to QVFR defaults.

My thoughts about the whole process: downgrading from godlike (3072x3216 on the QPro) to high (2496x2592) but increasing centre resolution from 1.1 to 1.5 results in DCS rendering that part of the image at an about 20% higher resolution BUT... after rendering the periphery at lower resolution, it combines both at 2496x2592 (VD's "high" setting). This is then sent, at that resolution, to the encoder and streamed to the HMD where it is upscaled. Hence the 1.5 resolution increase for the centre is akin to supersampling (which is good) but in the end, less information is sent to the HMD, which has to make up for it by a relatively basic, low-computational cost upscaling algorithm (SGSR). With Godlike, the image passed to the encoder is 3072x3216 and inherently contains more info. To which extent this will translate into improved final image quality is affected by the streaming (and thus encoder) bitrate. If the bitrate is capped too low, the lossy compression will reduce (but not negate) the advantage of passing an higher resolution image to the encoder, but I doubt this would be the case in practice, with bitrates >= 150 Mbps.

Hence, assuming the above is correct and encoding/streaming shenanigans notwithstanding, my understanding is that SGSR just upscales the central part of the picture from "high" to "ultra" resolution but starting from a picture with 1.5 supersampling. Idem for the periphery, but it was downsampled to begin with, so it something like high with 0.5 downsampling then upscaled a bit from high to ultra.

 

Thanks for trying this out. I can't quite get my head around it. I need to look into it a bit more. 

[mrsylvestre]

Perhaps we should not overthink it. In the end, what matters is that with your settings and your QPro, you get stable native 72 fps with no (perceived) loss of image quality. The fluidity alone is probably worth it.