Sunday, November 16, 2008

Smooth transitions

The T-rex night attack scene from Jurassic Park was a major milestone for CG. It showed a realistic and convincing CG character in a movie. As the T-rex came after a girl Dr. Grant tells her, "Don't move. If we don't move he won't see us". This is something graphics programmers should remember as well because it holds true for humans as well. Our eyes are very good at seeing sharp changes but not very good at seeing smooth changes. If there is a way to smooth a hard transition you can get away with a lot more than if you didn't. I think this is the prime thing missing in most LOD implementations. The LOD change gets pushed far enough in the distance that you can't tell when it pops from one LOD level to the next. If the pop was replaced with a smooth transition the LOD distance could be pushed significantly closer and still not be noticeable.

I only played the demo but I really liked their cascaded shadow map implementation. It looks like there is just 2 levels to it. After that it goes to no shadows. What is really nice about it is there's a smooth transition between the levels. As you walk out to an object it will fade from no shadow to low res shadow to high res shadow. So many cascaded shadow maps in games look strange or jarring because there is a line on the ground where it goes from one res to another in the shadows. This moves with your view direction and movement.

Gears of War 2
In my opinion any serious game graphics programmer or artist is obligated to play Gears of War 2. It is the bar now for graphics on a console. For being blown away by visuals it gives Crysis a run for its money too. As far as tech that seems absurd but it's just the combination of art and tech with enough things I'd never seen before that this takes the prize for me. It isn't this through and through so it really takes playing the whole game to get what I mean.

As far as tech I was a bit surprised when Tim Sweeney showed at GDC the new things they added to the Unreal engine for the next Gears of War. I was surprised because it wasn't very much. In the time between UT2k4 and GoW they built a whole new engine. Sure it was an evolution but it was a large one. The renderer was rewritten, they created a whole set of high end tools and changed the internal framework drastically. For GoW to GoW2 they added SSAO, hordes of distant guys, water, changed character lighting, and destructible geometry (which wasn't really in the game). This doesn't sound like very much considering they have 18 engine programmers listed in the credits.

The change that impressed me the most is fading in textures when they stream in. Some UE3 games have gotten some flak for streaming images popping in. Instead of not pushing the memory as much they added fading in of mip levels. For smooth transitions this is brilliant! I'm guessing most will never know textures aren't streamed in on time because they will never see the pop again. I also noticed they may have pushed the texture streamer further because they didn't have to worry about subtle pops for new mip levels. I couldn't tell if this happens when an image downsizes because I never noticed any image downsize.

Fading new mip levels once they stream in is something I've wanted to do but I just don't know how they are doing it. If anyone knows or is doing something similar themselves I'd love to hear how it works. The problem I see is there are min and max mip levels settable for a texture sampler on a 360. These unfortunately are dwords. Lod bias is settable but this happens before the clamp to min and max. The only way I could see this working is if they calculate the gradient themselves in the shader and clamp it to the fade value as they lerp from the old clamp to the new full mip level. This seems to me like it would create a shader explosion if this needs to be turned on and off for every texture for every shader. The alternative is always manually calculating the texture gradient for all uv's used and then clamping it individually for each texture which I believe would be quite a bit slower.

Next up is the screen space ambient occlusion (SSAO). This helped with their low res lightmaps and showed off their very high poly environments. Personally I think it was over done in many cases but I guess overall it was an improvement. I was surprised by the implementation. They are using frame recirculation to reduce the per frame cost. You can tell because obscuring objects will wipe the AO away for a moment before it grows back in. It seems to be at a pretty high res, possibly screen res. Previous results are found either using the velocity buffer or just the depth buffer and camera transformation matrix. Using this position they can sample from the previous calculated results without smearing things as you look around.

Much of the visual splendor comes from clever material effects. They have scrolling distortion maps to distort the uv's. There's pulsing parallax maps that look so good for a moment I thought the whole area was deforming geometry. There was inner glow from using an additive reverse fresnel like the cave ant lions in Half Life ep2. There was a window pane with rain drops coming down it that I had to study for like 2 mins to figure out what was going on. My guess, 1 droplet map, 2 identical drop stream maps independently masked by scrolling textures. The final normal map was used to distort and look up into an environment map. Their artists really went to town with some of this stuff.

The lighting is still mostly directional lightmaps. Shadows are character based modulate shadows, this time higher res than before. It seems they are only on characters this time leaving the SSAO to handle the rest of the dynamic objects.

No comments: