Official Review · ICLR 2026

Diffusion2 is a generative diffusion model that predicts high-fidelity RF heatmaps from 3D point clouds with only 15 pre-measurements, supporting Wi-Fi and mmWave bands. It introduces an RF-3D Encoder for multi-modal fusion and an RF-3D Pairing Block for conditioned denoising, enabling accurate multipath modeling and dynamic scene generation. + The diffusion-based methodology provides a solid foundation for future multimodal unification. Its probabilistic denoising framework supports flexible conditioning on diverse inputs such as 3D geometry, 2D images, and RF signals, enabling seamless integration of additional modalities like audio, inertial, or environmental data to enhance holistic wireless scene understanding in 6G applications. - While Diffusion2 applies diffusion models to RF heatmap generation from sparse 3D points, its novelty is modest. The method mainly reuses standard DDPM/DDIM pipelines, latent diffusion encoding, and simple additive conditioning without introducing new RF-specific mechanisms. Its modules, such as the RF-3D Pairing Block and RF-3D Encoder, assemble existing components like MinkUNet, Swin Transformer, and Fourier embeddings, mirroring prior multi-modal diffusion works. - While the RF-3D Encoder presents an innovative fusion of multi-modal inputs, including 2D images, 3D geometry via MinkUNet, and RF signals through Fourier embeddings, its use of simple element-wise addition for feature fusion may fail to capture complex non-linear relationships among modalities. This simplification can cause misalignment or feature dilution in cluttered or heterogeneous environments. Although ablation results show that each modality improves accuracy, the lack of cross-modal attention or comparisons with more advanced fusion mechanisms leaves uncertainty about the model’s robustness and generalization in real-world scenarios such as lighting variations or occlusions. - The paper omits an essential baseline by not comparing with WRF-GS (Wen et al., 2024), which shares the same objective as Diffusion2 and NeRF2. If Diffusion2 can be evaluated against NeRF2, it should likewise be compared with WRF-GS, which performs comparable tasks at a significantly lower runtime. Without this comparison, it is difficult to assess whether Diffusion2 achieves comparable or superior efficiency and accuracy. - Its practicality is constrained by low output resolution and restrictive input requirements. The generated heatmaps with resolutions of 352×705 for images and 52×72 for videos are too coarse for dense receiver grids or detailed multipath analysis, which limits their value for high-precision 6G diagnostics. Data collection using smartphone-based LiDAR scans introduces variability caused by occlusions and lighting conditions. The fixed use of 15 pre-measured points also appears arbitrary, as performance is sensitive to this choice without adequate justification or analysis. Please see the points raised in the Weaknesses section.