This blog post sums up the most important improvements in Atomontage Engine done in the past two years.
The project made significant progress in data management, geometry processing, voxelization, VR and a lot more.
All you see in this blog post was voxelized/generated and then rendered in real-time on a XMG P722 notebook (assembled in 2012, specs).
You can find more technical info in the descriptions of the particular YouTube videos.
Also check out the related Gallery on Facebook.
The original 3MPoly model would be a bit difficult to render and modify in real-time. The voxel model can be rendered easily with hundreds of FPS (could be thousands on a modern gaming PC).
The time is near when artists won't have to care about polygon counts anymore!
Generators perform better thanks to faster code and improved multithreading.
Fast generation makes it possible to create an object or scene multiple times and iterate toward great results more quickly.
In a typical future scenario the scene will be generated dozens or hundreds of times automatically in order to find the best set of generation parameters for minimal cost.
The sand, hexahedron block, surface detail and texture- and noise-based damage compile-time generators have been improved significantly.
They are finally capable to create realistic geometry suitable for first-person view scenarios.
A ton of issues with ray-based voxelizer have been fixed. While this voxelizer can't be used with very flawed meshes,
it is now more robust in combination with a set of post-voxelization filters.
These can be used to automatically remove most of the artifacts caused by voxelization or mesh processing (a flawed mesh export or import).
The severe artifacts on the cabrio (magenta) cannot be fixed automatically.
Ray-based voxelization can fail completely usually due to open volumes, seams, overlapping polygons or missing whole parts (e.g. the underbody of a car).
The solution to this problem is projection-based voxelization.
This method provides good results with almost any polygon geometry that looks good when rendered using a polygon renderer.
The performance of the engine is good enough to render VR, but the frame rate fluctuates strongly in high-resolution (first-person view) scenes when camera moves quickly.
This will be fixed after I rewrite one crappy 6-years old piece of code in the renderer.
That fix will make it possible to maintain stable FPS (can be 90FPS or even 120FPS) independently of the geometry complexity or scene detail.
A major update to the compression scheme is just happening – soon a sparse definition of some voxel geometry will reduce the memory/medium footprint by 70-90%.
This combined with perception-driven optimization will often reduce data volume by more than 95%.
So one of the goals for 2015 is to encode 100GVx to 1TVx scenes to file sizes that would fit on cheap phones from 2014, or a whole game with 10-20 levels on a more recent phone.
This is exactly what we need for having mobile VR + first person view + fully destructible huge scenes in a single app.
And movies. True three-dimensional movies!
What Comes Next?
The current top goal is to create a demo that will make it clear to everybody that voxels already are a viable alternative to existing approaches.
The new scene should feature 50+ fully destructible buildings, large landscape (possibly a whole planet), a number of destructible voxel vehicles, etc.
The detail level should make it look good in first-person view VR.