Well, considering the reactions on the E3 trailer, it looks like it was totally worth it, no? (sorry, too tempting
)HDR != floating point render targets - even your eyes have lower dynamic range than the world around and therefore do the exposure control to tone it down. To do it in game, you don't need to have floating point textures, you just need to adapt your rendering based on amount of brightness of the visible scene Therefore any higher range source of scene luminance is enough to do this - why would you do expensive stuff, when you can have the same result cheaper and save memory and bandwidth.
Of course doing real-time HDR reflections is cool and probably needs floating point textures to have nice filtering, but Killzone is not a racing game, no
Anyway - been on that talk too and talked to GG guys a bit.
Trade-offs to go for single lighting model and no specular reflections seems to be totally ok in their context - why would you have all that when 99% of your scene objects have Phong and do not use specular color - I cannot imagine developing custom in-house engine without considering every single trade-off artists have to make for tech and vice versa. They also still seem to have custom artist created shaders for everything (albedo, normals, material parameters) but direct light model - guy even talked about custom skin shader with scattering during Q/A.
They were later talking about more than 100 on screen real-time lights and around 10 on-screen shadow casting. I would not like to build forward render engine for such requirements. I think most of the trade-offs they made seem reasonable with these numbers and I think 4xMSAA and directional lightmap might be too much extra memory, but who knows why they don't have it.
Didn't he talk about 1024x1024 shadow map splitted into 4 regions for main directional light? That is much more than just 512x512. Also doing 12 HW filtered (I think nVidia HW does bilinear since G3, no?) taps per pixel and light is 3x more than many other games do
Anyway, their IBL technique is really cool, gotta love those SPUs.
I'm still not decided about those particles - if you have IBLs for particles and these are mostly clouds and dust anyway, you get pretty good results, particle is just flat so per-vertex should be perfectly fine. But I think he mentioned that per-vertex is optional anyway and it seems that the destruction particles have normal-maps in the trailer.
To me it looks like they have quite a lot of places where artists make decision what quality they want for speed and there are more code paths based on this.
Also I heard from some other guy that they will to release the presentation for download but I think all presentations will be online after develop - so we can read it and analyze.
