CaliSplat cuts the search cost of Gaussian image compression

CaliSplat tackles an expensive part of image compression: searching for a compact representation of each individual picture. In a preprint that has not been peer reviewed, Zhenbang Zhang, Zhengquan Luo, Kai Zhao and Zhiqiang Xu report reducing image-specific fitting time by 3.0–15.7 times compared with representative Gaussian codecs. The contribution is a more selective search, not a universal improvement over every compression method.

Gaussian codecs reconstruct pictures using mathematical components with adjustable positions, shapes, weights and colors. In CaliSplat, each pixel within a block becomes a weighted mixture of component colors. These explicit representations can be decoded without a heavy neural image-synthesis network. The difficulty lies in encoding: trying many possible representations can consume substantial optimization time, including work on candidates that are ultimately discarded.

The method starts with a particularly cheap question: how badly would one constant color describe each image block? With a single Gaussian, CaliSplat’s normalized weighting makes the whole block one color. Under squared error, the best color is simply the average of its pixels. The remaining error can therefore be calculated directly, without iterative fitting. A low error suggests a relatively simple block; a high error identifies a candidate for additional capacity.

CaliSplat uses this signal to decide which more elaborate representations to try. Its reported setup starts with 16-by-16-pixel root blocks and considers component counts from one to 16. Larger candidates are tested on progressively smaller fractions of the most difficult blocks. Each admitted candidate receives 40 optimization steps. This filtering determines where computation is spent; it does not yet determine the final allocation of components.

A second stage addresses a problem with such short trials: different candidate sizes improve at different speeds. A representation that looks worse after 40 steps might outperform another after longer fitting. For each candidate size above one, CaliSplat independently fits up to 128 sampled blocks for 160 steps, then uses those reference errors to calibrate the short-trial scores. Crucially, these longer fits adjust selection scores only. Their fitted parameters are discarded, and the final image uses selected short-fitted candidates. The calibrated scores are therefore estimates, not measurements of the final reconstruction’s error.

The allocator then chooses the block structure and component counts under an image-wide Gaussian budget. It adjusts a penalty for using additional components, favoring extra capacity where the predicted reduction in error justifies it. This is not a guarantee of the exact best solution for every budget: the authors acknowledge that discrete choices can prevent their search from recovering the constrained optimum. Nor is a component-count budget identical to a file-size budget. Actual bytes also depend on parameter coding.

For storage, CaliSplat encodes geometry as differences from shared deterministic templates, alongside weights, palette-coded colors and block information. Quantization reduces parameter precision, and a range coder compresses the resulting symbols. Decoding recovers those parameters and renders the blocks. In component tests, the full structured coding scheme reduced bitstream size by about 26% on DIV2K and 28% on ICB relative to absolute parameter coding, with the selected representation and quantization held fixed.

The main evaluation contains 75 standard-resolution images and nine large images. Against the Gaussian baselines, results favor CaliSplat particularly at higher rates. On standard images, Image-GS has slightly better low-rate PSNR, 27.95 versus 27.71 decibels; CaliSplat leads at the high-rate point with 32.95 decibels. PSNR measures pixel-level reconstruction accuracy. On large images, CaliSplat leads the compared methods in PSNR at both reported rates, although SGI scores better on the structural-similarity measure MS-SSIM at the low-rate point.

Reported fitting times are approximately 22–24 seconds for standard images and 95–103 seconds for large images. Decoding throughput ranges from 175.9 to 478.2 megapixels per second, although CaliSplat is not the fastest Gaussian decoder in every setting. A separate component test reduces candidate-fitting work to 65% of exhaustive search with essentially unchanged quality. That workload reduction is not equivalent to a 35% reduction in elapsed time.

The broader comparison limits the headline. The paper’s appendix reports clearly better reconstruction quality for three learned codecs: at the standard-image high-rate point, GLIC reaches 38.45 decibels against CaliSplat’s 32.95. CaliSplat instead offers much higher reported decoding throughput. The supplied manuscript does not identify the timing hardware, and its nine-image large-image evaluation provides limited evidence of generality. These author-run results support a promising efficiency trade-off within Gaussian compression, not independently verified superiority across image codecs.