Are you saying depth buffer based algorithms are not good enough quality for you?
That was a negative that was fixed with SAO. Their algorithm exploits the high precision z-prepass in modern GPU's (GTX 500 series and later) so they can rely only on the depth buffer to get what they want and they can do it with high quality.
They reconstruct the position and normals at high precision because of this using just the depth buffer.
Boris, will this ambient occlusion you are working on work in regular Skyrim Special Edition, or only for the VR version? From what I understand it will have better performance for the same quality? I currently turned SSAO down a bit in my preset because it is quite demanding even on a GTX 1070, but if there is an optimized SSAO coming for SSE that would be epic.
Hi Boris, found a (new?) bug. Time of day is broken in certain interiors in Solitude (only ones I know are Blue Palace and Bards College), where it reports that the time is 10.00, even though it is at night. Other interiors such as Winking Skeever have slow TOD which reports about ~2 hours slower. Mods used are Obsidian Weathers, ELFX (no enhancer) and Distinct Interiors.
Pritster5
It's not serious, i can turn my ao to same low quality, but it's a crap. Per pixel normals are must, depth reconstructed sucks.
drift123
I dont know yet, this depends from result. Right now the only fast way is to sacrifice small details (and i dont want to ignore normals like i said above), but without small details something like nose will not be properly occluded (adding another pass for small things is too high overhead).
Flamerapter
Time detected from sun direction or sky colors, if there is none, then time is ignored, so it's not a bug, but limitation.
Pritster5 Position can be reconstructed from depth and it's common practice to do so, but normals as partial derivatives from depth buffer are too low quality. Here is Boris' comparison video of what you propose vs. what he uses right now. It should also be noted that his AO also looks better now than it did when this video was recorded, so the difference is even stronger.
Actually the video you linked is simply showing off Forward AO vs Deferred AO.
SAO (Scalable AO) is also deferred. It computes normals with accuracy up to 0.2 degrees better than RGB8 G-buffer normals in the worst case.
You can still use this with normals and get decent quality. The picture i posted simply doesn't have much going on in the normal maps. This is how it looks with normal maps:
Well the worst case of 0.2 degrees is not valid at the singularities, which I believe they mention somewhere. Also, depth buffer ignores smoothing groups so even though the normal might be perfectly smooth in G buffer, it'll be faceted when reconstructed by depth. But what do I know, I have little knowledge of AO.
Pritster5
Something is wrong with normals on your picture, they look like not used in computation at all, just applied after and wrongly. Filtering result of normals based ao is much more complex task which cost sometime half of performance as ao itself or even more, because you can't just blindly use bilateral filter for bllurring everything and lose details.
I think it's worth taking a look at least what Activision did for Ground Truth Ambient Occlusion, it takes depth buffer and surface normals and runs 0.5ms on a PS4 @ 1080p.
