Learning-based Gaussian image initializers, such as those used in Instant GaussianImage, typically need supervision generated by per-image optimization. That step is costly. A new arXiv preprint asks whether those Pseudo-PPM targets can be swapped for classical image gradients, which are cheaper to compute.
The authors frame this as a direct test of whether learned targets are necessary at all. If classical gradients work, they could simplify training pipelines for 2D Gaussian image representations. However, the abstract does not include experimental outcomes, so it is not yet clear whether the replacement matches the original in quality or speed.
Because the digest draws on a single source, there are no conflicting findings to reconcile. The paper is listed as arXiv:2610.25809v1.