NetDiff: Inria's graph diffusion model hits 95% of target network performance

Directional antennas promise high throughput in mobile ad hoc networks, but they require globally consistent link decisions under tight sector, interference, connectivity, and half-duplex constraints. Existing approaches struggle to balance local link choices with global topology goals. Enter NetDiff, a node-conditioned denoising diffusion model from researchers at Inria (Félix Marcoccia, Victor Fagoo, Gilles Monzat, Cédric Adjih, Thomas Watteyne, and Paul Mühlethaler). The model jointly generates directional link topologies and a two-slot transmit/receive parity, ensuring every node has a coherent schedule—a key challenge when half-duplex radios can't send and receive at the same time.

NetDiff's core innovation is Absolute Cross-Attentive Modulation (ACAM) tokens, permutation-invariant global signals that let the model match graph-level counts like density and sector usage. This improves global coherence compared to standard diffusion graph transformers. The authors also propose partial diffusion: instead of regenerating from noise, the model updates an existing topology with just a few denoising steps. That enables fast reconfiguration as nodes move, maintaining over 95% of target performance with constant inference time. In experiments, NetDiff outperformed heuristic and omnidirectional baselines and surpassed a strong diffusion graph-transformer baseline on key metrics.