FUTON: Fourier Tensor Network for Implicit Neural Representations
This new architecture could make complex 3D and image AI dramatically faster and cheaper.
Researchers introduced FUTON, a novel Fourier Tensor Network for Implicit Neural Representations (INRs). It replaces traditional slow, overfitting MLPs by modeling signals as generalized Fourier series with low-rank tensor coefficients. This combines Fourier smoothness with efficient spectral structure. The result: FUTON consistently outperforms state-of-the-art MLP-based INRs on image and volume representation while training 2-5 times faster. It also shows superior generalization and speed on inverse problems like denoising and super-resolution.
Why It Matters
It could drastically reduce the cost and time for training AI models in computer vision, graphics, and medical imaging.