MADWE: Multi-Agent Diffusion World Engine

A multi-agent generative runtime that routes work to specialist diffusion agents over a shared world-state graph with budget-aware scheduling, generating 120 game assets per minute and cutting manual design time 73%.

Real-time game worlds need coherent assets, layouts and styles under strict latency and memory budgets, and hand-authoring them does not scale.

Asset agent on Stable Diffusion 1.5 with rank-8 LoRA adapters per biome and latent-consistency few-step inference; Wave Function Collapse for layout constraints; evaluator agent scoring quality; biome-aware prefab mapping into Unity. Shreya led AI models and generation. Unity bridge with batched prefab transfers, caching and request prioritization.

1. 01

   Optimized GPU inference with ONNX and TensorRT kernel fusion: p95 latency 200 ms to 95 ms, FPS 60 to 120, binary size down 38%.

2. 02

   Designed a neural world-state graph for coordination, a predictive generation protocol that pre-warms likely branches, and memory pooling that cut allocation overhead about 73% and peak VRAM about 41%.

3. 03

   Mixed precision, INT8 quantization and CUDA graphs delivered 1.8x throughput and 24% less memory while holding FID 12.3 and CLIP similarity 0.89 under 2 GB VRAM.

4. 04

   Orchestration layer sustained 31 FPS at 6.2 GB VRAM with PSNR 30.1 dB and 94% tileability across large variant sets.