Cross-resolution dual-branch Vision Transformer for 1 km PM₂.₅ forecasting across Europe. 29M pixels in 1.8 s; zero-shot transfer to USA, Canada and India (arXiv 2026).

Physics-guided ViT for multi-pollutant forecasting over China (PM₂.₅, PM₁₀, SO₂, NO₂, CO, O₃). Wind-guided patch reordering + topography-aware attention; 71–80% improvement over operational systems (arXiv:2602.16821, under review).

Efficient point cloud upsampling via flow matching (AAAI 2026). Colab demo available.

Spatio-temporal air quality reanalysis model with LSTM, multi-head attention and neural kNN for PM₂.₅ prediction (SCIA 2025).

Adaptive blend pyramid network for ultra high-resolution photo retouching (CVPR 2021)

Deep blind image super-resolution via coupled noise estimation

Arbitrary style transfer via multi-adaptation network (ACM-MM 2020)

Audiovisual learning of funny moments in videos (ACCV 2022, Best Student Award HM)

Deep relighting networks for image light source manipulation (ECCV 2020)

Deep generative inpainting network for extreme image inpainting (ECCV 2020)

Hierarchical back projection network — top-ranked on Urban100, BSD100, Manga109