AI Level of Detail: Distance-Aware ML Model Precision Selection for Real-Time Human Motion Prediction in Games
AI Level of Detail (AI LOD) extends the classic geometry level-of-detail idea to machine learning inference. Just as a game engine swaps a distant mesh for a cheaper one when the difference is imperceptible, AI LOD swaps a full-precision motion model for a cheaper quantized one based on how far an NPC is from the player camera. The novelty is the concept itself: inference-time quantization precision as the LOD axis for AI-driven character animation.
Using the ConvSeq2Seq human-motion predictor of Li et al. as a fixed backbone, the trained checkpoint is exported into three ONNX Runtime tiers — FP32 (closest NPCs), FP16 (mid-range), and INT8 per-tensor (farthest) — routed at runtime by a distance-based selector. The tiers are evaluated on the CMU Mocap dataset for latency, model size, and motion accuracy, alongside a small perceptual study.
Key results (CPU, vs FP32 baseline):
- FP16 — 1.53× faster, 49% smaller, near-zero accuracy loss (relative L2 7.8e-4).
- INT8 per-tensor — 9.79× faster, 73% smaller, with degradation that stays below the perceptual threshold at far distances.
A 15-participant perceptual study found no detectable difference when each tier was viewed within its intended distance range, supporting the core premise that distance-aware precision selection is a viable LOD strategy for AI-based animation.