CARA: Collision-Aware Resolution Adaptation for Multiresolution Hash Encoding Based Image Fitting

arXiv cs.GR details CARA, a new approach for multiresolution hash encoding that tries to fix a familiar efficiency problem: some hash levels get overloaded while others sit underused. The paper frames this as a capacity-allocation issue rather than a pure model-size problem, and adapts each level’s resolution to balance the effective information load.

For developers working on neural rendering, image fitting, or other implicit representation pipelines, the practical angle is parameter efficiency. CARA also adds an invertible pixel-shuffle transform that redistributes spatial information before hashing, reducing collision pressure without making the tables bigger.

The authors tested the method on Kodak images, gigapixel natural images, and raw whole-slide images, which is a useful spread from standard benchmarks to extreme-resolution data. They also curated what they describe as the first uncompressed whole-slide image dataset for academic research, which could matter for anyone benchmarking high-res reconstruction methods.

The headline result is strong: CARA reportedly matches state-of-the-art quality while using only 27.76% of the parameters, and reaches up to 6.11 dB PSNR improvement at comparable parameter counts. Code is included in the supplementary material, so teams interested in neural image fitting can likely try the approach without waiting for a separate release.