Sury wrote:Yeah, those machines always look frightening to me, because they confuse the brain of the observer and the brain fears the unknown + it knows that something is wrong, but cannot figure what exactly and that forces it switch to "caution mode".
Yep, the dilemma if something is very close to look real, we tend to accept it less likely than we would accept a cartoon.
May become a problem - but not more than other things related to caharacters. For now my ragdoll looks pretty natural, sometimes it makes a few steps by accident, but always looks either silly, funny or natural - never as creepy as those real robots.
For the GI, i've done a lot work here and i think i'm done with 90% of research, but i do on CPU and need to learn GPU somewhen soon to proof it's fast enough.
The main problem i see with current algos is the question: "Is it really worth to spend so mutch cycles for single bounce reflection?".
I'd say no, even if you get two bounces. Unlimited bounces is a must for me, and you can get those for free if you use a technique that stores results and reuses them in the next step.
I think Enlighten is the only current game approach that may use this(?), but there's no reflection from dynamic objects - they're still only receivers?
Crassins method is promising - mainly because it supports speccular reflections - but i heared that Unreal engine gave up on it - too slow.
Cryteks method is fast, but as long as they sample the world only from screenspace it's the same fake as SSAO.
The VPL method you use misses shadows - fixing that led to very interesting ideas:
"Imperfect Shadow Maps" + "Micro Rendering" (Tobias Ritchel, Carsten Dachsbacher...)
"ManyLods" (Mathias Hollaender, T.Ritchel)
Very inspiring... be sure to read this stuff.
Georgios Papaioannou did something similar than Crytek, but used world space voxelization like Crassin, and tried some other ideas.
Cyril Crassin for sure.
Also the 10 years old stuff is still valid.
After reading this and a lot more, i came up with a mix influenced by most of those techniques with those properties:
* does not use rasterization or raytracing to compute diffuse interreflection
* supports unlimited bounces at the cost of temporal lag
* needs to store spherical harmonics samples either over surface using LOD tree structure (fast), or in a cascaded volume grid (GPU friendly)
* can support speccular reflections from samples and volumetric light for particles / fog.
* dynamic world and lights (for sure)
* may allow "force < X ms" at the cost of more temporal lag (the missing 10 %)
Sounds good on paper, but let's see first... Maybe real GI still have to wait for PS5?
Looking at your screenshots, if architecture is non destructable, SH radiance grid would suffice?