Zhang, Li, Kuznetsov, Jindal, Diolatzis, Chen, Sochenov, Kaplanyan and Sun (2024) — ARXIV-2407.01866, read as NOTE-256.
Standing
A fifth source for SOTA-205, and the first outside 3D. That practice
— replace a large coordinate network with a compact explicit structure and a
small decoder — is converged on four radiance-field papers. This carries the
same claim into single-image representation and supplies the decode-cost
number that makes it concrete.
The extends: on LIT-108 is earned rather than a family
resemblance. The paper says its work “is inspired by the recent success of
Gaussian Splatting”, defines its primitives “similar to the Gaussian
primitives in 3D Gaussian Splatting”, and states what it drops — no spherical
harmonics, “as an image essentially shows a single view”.
What it beats at matched size includes the coordinate MLPs. Its six neural baselines include SIREN (LIT-551) and a Fourier-feature network (LIT-550), run from their official implementations with only model size changed: at 2K×2K they are 161 KB and 154 KB against Image-GS’s 160 KB, with bitrate on the axis, and Image-GS outperforms both across the evaluated bitrate range. The comparison is of fitting pixel values; SIREN’s claim about fitting derivatives is not tested by it.
Held also as the constructive counterpart to LIT-494. GaussianToken reached this record two units earlier with no bit accounting: its comparisons were matched on token count while each token carried an index plus five continuous floats, and NOTE-243 recorded that the obvious rival explanation was never ruled out. Image-GS is the same Gaussian machinery in the same year, reporting bits per pixel throughout and matching model sizes in kilobytes. No relation is declared between them: nobody ran that comparison, and ADR-011 means what it says. The connection is this record’s reading and is stated in prose.