GeneralisedImplicitNeuralRepresentations

One interesting application of INRs is to train them using the derivatives of the target signal as supervision. This idea, which was introduced by Sitzmann et al.[1], can also be applied to the generalisedcase. Specifically,inthis experiment weconsider thebunnyreaction-diffusion texture and train ageneralisedINRtominimise: L= Lf Lfθ, (1) where L is the graph Laplacian, f = [f(v1),...,f(vn)] Rn is the target graph signal, and fθ = [fθ(e1),...,fθ(en)] Rn isthesignalpredictedbytheINR. Wesee that the model is able to correctly reconstruct the signal, although with some oversmoothing. In the second experiment of Section 4.1, we train the models using the default setting.