Jakub Kościukiewicz

Research

• HyperNeRFGAN| DSAA 2025

  HyperNeRFGAN frames 3D object generation as an adversarial weight-generation problem, using a hypernetwork to map Gaussian noise directly to the parameters of a NeRF model without requiring camera pose supervision or viewing directions. By eliminating reliance on explicit 3D representations and camera estimation, HyperNeRFGAN enables efficient training from unposed 2D images and outperforms existing approaches on datasets where camera information is unreliable or unavailable.

  [Project Page] [Arxiv] [Code]

• HCS-DFC| CVDD Workhsop CVPR 2025

  HCS-DFC reformulates prediction as a conditional generation task to inherently model label distributions and co-dependencies without requiring calibration datasets. By leveraging diffusion models’ ability to capture complex data distributions, HCS-DFC outperforms conformal prediction methods in reliability estimation and achieves state-of-the-art accuracy on both synthetic (MNIST-based multi-task classification) and real-world cell painting datasets.

  [Paper]