ABCD: Alpha-Composited Block Coordinate Descent: Constant-VRAM Training for Large Radiance Fields

ABCD, short for Alpha-Composited Block Coordinate Descent, reframes radiance-field training as an out-of-core, block-coordinate problem. Instead of keeping an entire large scene active in VRAM, it partitions space and trains one block at a time while freezing the rest. For developers pushing 3D Gaussian Splatting or similar alpha-composited representations, the practical win is simple: memory use stops scaling with total scene extent.

The key trick is to exploit alpha blending associativity. Inactive regions are pre-rendered and collapsed into foreground and background RGBA images, so the current block can be optimized against compact composites rather than the full scene. That makes peak VRAM effectively O(1) for a fixed partition size and image resolution, which is a meaningful shift for teams working on large environments or on hardware with limited memory headroom.

In testing, the method stayed close to baseline 3DGS quality, with less than 5% PSNR degradation. A version without the compositing step performed far worse, dropping by roughly 40%, which suggests the rendering collapse is doing the heavy lifting rather than just the block-wise schedule. The work is slated for ACM SIGGRAPH 2026 Posters, and code is available for experimentation.

For game developers, the immediate interest is not just faster training on big scenes, but the possibility of making radiance-field workflows more practical on midrange GPUs and in iterative production pipelines. That matters for tools teams, rendering researchers, and anyone trying to scale scene capture, reconstruction, or...