Learned synthesis is becoming procedural.
InfiniteDiffusion targets seed consistency, unbounded extent and random access. Geomorphological transport supplies drainage-connected structure; controllable patterns add cheap near-field detail.
Research brief / 2 September 2026
Real-Time
Natural Worlds
I went looking for what is actually current in procedural terrain, procedural foliage and flora, and water rendering for real-time videogames.
What I found
The frontier is hybrid.
Placement and rendering are separate systems.
Terrain descriptors and ecology determine where species belong. Work graphs and mesh shaders can then expand compact plant programs into visible geometry with continuous LOD.
No single representation covers water.
Broad spectral surfaces, localized topology-capable waves, directed flow fields, foam and spray, and underwater optical transport each own a different scale and phenomenon.
The engine boundary I would draw
Generate visuals from stable world data.
World contractTerrain fieldsBiome + floraWater stateRender + stream
I would keep deterministic tiles, shared descriptor fields, stable organism IDs, a water-body graph, and sparse persistent gameplay deltas on the world side of that line. Camera state should select rendering detail, never persistent identity.
Evidence
Recent primary sources, clearly labeled.
I went first to SIGGRAPH 2026, SIGGRAPH Asia 2025, High-Performance Graphics 2025, Computer Graphics Forum, official production talks, and author project pages. I mark peer-reviewed papers, production R&D, preprints, and adjacent work apart from each other throughout.
Open the complete report