- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper presents CaliSplat, a per-image compression framework based on 2D Gaussian representations. The method uses the analytic single-Gaussian reconstruction error to selectively fit higher-capacity candidates, calibrates short-horizon distortion scores using a small number of longer-horizon teacher fits, and performs global Gaussian-budget allocation followed by structured parameter coding. It aims to reduce fitting costs while maintaining reconstruction quality and efficient decoding.
Critical strengths
1)The paper addresses a clear and practically relevant problem. Separating computational effort for candidate fitting from final representation-capacity allocation provides a reasonable approach to reducing redundant per-image optimization. 2)The complete system shows promising performance. The reported fitting times and decoding throughput suggest an attractive trade-off between reconstruction quality and computational efficiency. 3)Several ablations are well motivated. Comparisons against exhaustive fitting, random candidate selection, and spatially shuffled allocations support the value of selective search and content-adaptive capacity allocation.
Critical weaknesses
1)The calibration target is misaligned with the final encoded representation, and a critical ablation is missing. Calibration predicts the distortion of 160-step teacher fits, whereas the final codec retains the 40-step candidate parameters without any reported post-selection refinement. The paper does not explain this mismatch or establish the practical benefit of calibration through a controlled comparison with uncalibrated allocation using the same candidate bank.
2)The rate-distortion comparison protocol is insufficiently transparent. The main tables and qualitative comparisons do not specify actual bitrates or how operating points are matched. The correspondence between the high-rate values and method rankings in Table 1 and Figure 4(a) also requires clarification.
3)The efficiency evaluation lacks essential details and a directly relevant baseline. Hardware configurations, timing boundaries, and baseline optimization settings are insufficiently specified, making the reported speedups and their sources difficult to assess. Instant GaussianImage, discussed in the introduction as a fitting-acceleration method, is omitted from the comparisons without explanation.
4)The existing ablations do not sufficiently isolate individual contributions. The allocation study mainly compares against uniform and random strategies under matched Gaussian counts, which does not establish the advantage of the proposed allocation algorithm at matched actual bitrates. The parameter-coding ablation reports bitstream reductions without verifying whether reconstruction quality remains unchanged.
5)valuation coverage and sensitivity analysis are limited. The large-image evaluation contains only nine images, with insufficient information about sample selection and image resolutions. End-to-end sensitivity analyses for fitting iterations, teacher sample counts, and candidate-retention ratios are also missing, limiting the evidence for stability and scalability.
Questions for authors
1)The calibration target is misaligned with the final encoded representation, and a critical ablation is missing. Calibration predicts the distortion of 160-step teacher fits, whereas the final codec retains the 40-step candidate parameters without any reported post-selection refinement. The paper does not explain this mismatch or establish the practical benefit of calibration through a controlled comparison with uncalibrated allocation using the same candidate bank.
2)The rate-distortion comparison protocol is insufficiently transparent. The main tables and qualitative comparisons do not specify actual bitrates or how operating points are matched. The correspondence between the high-rate values and method rankings in Table 1 and Figure 4(a) also requires clarification.
3)The efficiency evaluation lacks essential details and a directly relevant baseline. Hardware configurations, timing boundaries, and baseline optimization settings are insufficiently specified, making the reported speedups and their sources difficult to assess. Instant GaussianImage, discussed in the introduction as a fitting-acceleration method, is omitted from the comparisons without explanation.
4)The existing ablations do not sufficiently isolate individual contributions. The allocation study mainly compares against uniform and random strategies under matched Gaussian counts, which does not establish the advantage of the proposed allocation algorithm at matched actual bitrates. The parameter-coding ablation reports bitstream reductions without verifying whether reconstruction quality remains unchanged.
6)valuation coverage and sensitivity analysis are limited. The large-image evaluation contains only nine images, with insufficient information about sample selection and image resolutions. End-to-end sensitivity analyses for fitting iterations, teacher sample counts, and candidate-retention ratios are also missing, limiting the evidence for stability and scalability.
3 comments
W3&Q3
Thank you for pointing out that the timing comparison needs more detail. Below we give the hardware and optimization settings.
Hardware and software. Each job used one NVIDIA A100 PCIe GPU with 40 GB of memory. The host has four Intel Xeon Platinum 8260 CPUs at 2.40 GHz, with 96 physical cores and 192 logical threads in total.
Optimization settings.
| Method | Optimizer and learning rate | Fitting schedule |
|---|---|---|
| GaussianImage | Adan, 0.001 initially, halved every 20k steps | 50k fitting + 50k QAT steps |
| GaussianImage++ | Adam, 0.018 for attributes and 0.001 for the quantizer | 50k fitting + 50k QAT steps, including a 6k-step QAT warm-up |
| Image-GS | Adam, 0.0005 / 0.005 / 0.002 / 0.002 for position / color / scale / rotation | Up to 10k steps, with learning-rate decay and early stopping |
| LIG | Adam, 0.018 | 30k steps at each of two scales |
| SGI | Adam, with parameter-specific learning rates | 3k + 6k + 6k steps across three scales |
| CaliSplat | Adam, 0.035 | 40 steps per retained candidate and 160 steps per teacher fit, with up to 128 teacher samples per image and per count |
Image-GS starts at half the target Gaussian count and adds one eighth of the target every 500 steps, four times. Our GaussianImage++ runs use SLV initialization and pruning, with adaptive_add disabled to control the Gaussian count. SGI uses 50-dimensional features and 10 offsets per seed.
What the timing includes. CaliSplat’s encoding time in Table 1 includes candidate fitting, sampled teacher fitting, and allocation and bitstream coding. On standard images at the low rate, these stages take 8.52, 10.71, and 2.80 s/image, respectively, for a total of 22.03 s/image. The revised appendix subsection Encoding time breakdown reports the same breakdown for both image sets and both rates.
Where the speedup comes from. Table 2 compares sparse screening with exhaustive candidate fitting within CaliSplat at the same Gaussian budget. Screening reduces the -weighted fitting work to 65% of the exhaustive run and brings encoding time down from 292.49 to 221.69 s. Decoded PSNR stays close, at 30.485 dB with exhaustive fitting and 30.509 dB with screening.
Comparison with Instant GaussianImage.
Thank you for pointing out this omission. Instant GaussianImage is a relevant baseline because its learned initialization directly targets per-image fitting time. We have added it to the standard-image benchmark.
Bold and italics indicate the best and second-best results at each rate.
| Method | bpp | PSNR (dB) ↑ | MS-SSIM ↑ | Training time (s) ↓ | Decode speed (MP/s) ↑ |
|---|---|---|---|---|---|
| GaussianImage | 0.271 / 0.852 | 26.21 / 30.47 | 0.8829 / 0.9596 | 332.12 / 345.42 | 213.58 / 129.41 |
| Instant GaussianImage | 0.269 / 0.849 | 26.29 / 30.56 | 0.8841 / 0.9601 | 55.35 / 57.57 | 219.47 / 136.82 |
| GaussianImage++ | 0.230 / 0.631 | 26.30 / 30.30 | 0.8853 / 0.9578 | 341.99 / 346.98 | 256.53 / 241.74 |
| Image-GS | 0.215 / 0.900 | 27.95 / 32.08 | 0.8994 / 0.9607 | 77.50 / 71.51 | 227.00 / 205.51 |
| CaliSplat | 0.208 / 0.903 | 27.71 / 32.95 | 0.8866 / 0.9655 | 22.03 / 23.84 | 215.82 / 175.90 |
At nearly identical bitrates, Instant GaussianImage achieves similar reconstruction quality in about one sixth of GaussianImage’s training time. CaliSplat takes about 40% of Instant GaussianImage’s training time at both operating points. At the low-rate points, CaliSplat achieves 1.42 dB higher PSNR and uses 0.061 bpp less. At the high-rate points, CaliSplat achieves 2.39 dB higher PSNR at a bitrate 0.054 bpp higher.
W4&Q4
Thank you for flagging the missing measurements in Fig. 6.
1. Allocation at similar bitrates. At the same Gaussian budget in Fig. 6(b), the measured bitrates span 0.0056 bpp. CaliSplat gives the highest PSNR and uses fewer bits than either baseline.
| Allocator | Measured bpp |
|---|---|
| Uniform | 0.5220 |
| Random shuffle | 0.5191 |
| CaliSplat (ours) | 0.5164 |
2. Reconstruction after parameter coding. All four runs in Fig. 6(a) use the same partition and fitted Gaussians. Eq. (4) blends Gaussian colors as .
- Template residual. Adding the exactly decoded residual to the known template recovers the quantized geometry, .
- Relative log-weight. Subtracting a common reference from a block’s log-weights shifts every logit by . The common factor cancels in softmax.
- Palette RGB. Lookup preserves color when . A color change contributes to the pixel difference. Exact geometry and colors, with the same mixing weights, give the same reconstruction.
Size (KiB) / decoded PSNR (dB) / decoded SSIM
| Stage | DIV2K | ICB |
|---|---|---|
| Absolute | 260/30.52/0.891 | 3423/29.625/0.883 |
| +Template residual | 249/30.52/0.891 | 3213/29.625/0.883 |
| +Relative weight | 246/30.52/0.891 | 3178/29.625/0.883 |
| +Palette RGB | 193/30.51/0.889 | 2462/29.620/0.881 |
On DIV2K, the first two steps preserve both metrics to the reported precision. The full scheme saves 67 KiB, with drops of 0.01 dB and 0.002 SSIM. On ICB, the palette alone saves 716 KiB, with drops of 0.005 dB and 0.002 SSIM from the preceding stage.
W5&Q5
Thank you for flagging the missing image list and resolutions. Our nine-image benchmark matches SGI’s published large-image evaluation in scale and composition (Pan et al., CVPR 2026, Sec. 4.1), the closest precedent at tens of megapixels. Both use 4 FGF2, 2 ICB, and 3 STimage images. LIG uses 4 FGF2 (4K), 15 STimage (9K), and 100 DIV-HR (2K) images.
Table 1 and Fig. 4 cover 84 images across six datasets, including 75 standard-resolution images (15 Kodak, 35 Tecnick, 25 CLIC2020). The standard set is comparable in size to baseline benchmarks. Image-GS uses 45 RGB images and separately evaluates the CLIC2020 professional validation split. GaussianImage and GaussianImage++ each use 24 Kodak and 100 DIV2K validation images.
The 4 FGF2 and 3 STimage inputs came from the SGI authors’ experimental package. ICB’s big_building and big_tree match SGI’s reported 39.1 and 27.7 MP images, respectively. The nine images cover satellite, natural, and histopathology content. Their native sizes, verified from PNG headers, span 27.70–92.08 MP and total about 497 million pixels.
| Dataset | Image(s) | Native W×H | MP/image |
|---|---|---|---|
| FGF2 | GF2_0.png, GF2_1.png, GF2_2.png, GF2_3.png | 7300×6908 | 50.43 |
| ICB | big_building.png | 7216×5412 | 39.05 |
| ICB | big_tree.png | 6088×4550 | 27.70 |
| STimage | ST_0.png | 6954×7435 | 51.70 |
| STimage | ST_1.png | 9830×8602 | 84.56 |
| STimage | ST_2.png | 9272×9931 | 92.08 |
End-to-end sensitivity
We tested five configurations on Table 1’s standard images at its low-rate Gaussian budget. Each variant changes one default setting.
| Configuration | PSNR (dB) ↑ | MS-SSIM ↑ | Enc. (s/image) ↓ |
|---|---|---|---|
| Default (40 steps, teacher cap 128, default retention) | 27.71 | 0.886 | 22.03 |
| 20 fitting steps | 27.42 | 0.8802 | 17.77 |
| 80 fitting steps | 27.86 | 0.891 | 50.55 |
| Teacher cap 64 | 27.66 | 0.8858 | 16.68 |
| Retention 0.75× default | 27.69 | 0.8862 | 20.07 |
The teacher cap is per image and per count , with 160 steps per teacher fit. The retention multiplier scales default quotas for .
Fitting steps show the clearest quality–time trade-off. Compared with 40 steps, 20 saves 4.26 s/image for a 0.29 dB PSNR drop. At 80 steps, PSNR gains 0.15 dB and MS-SSIM reaches 0.891, but encoding takes 28.52 s/image more. This supports the 40-step default.
Halving the teacher cap and lowering retention save 5.35 and 1.96 s/image with smaller PSNR drops of 0.05 and 0.02 dB, respectively. MS-SSIM stays near 0.886 in both cases.
W1&Q1
Thank you for raising this concern. Fig. 3(b) of the paper shows why we use a longer fitting horizon for scoring. For the same block, the K=12 candidate has lower distortion at 40 steps, and the K=16 candidate has lower distortion at 160 steps. This reversal shows that early rankings can depend on how quickly candidates converge. We use sampled 160-step fits to calibrate the scores that guide the choice of partition and Gaussian counts. After selection, the encoder directly quantizes and codes the stored 40-step parameters.
To test whether these scores improve the image we actually decode, we added a controlled ablation on the standard-image set. At each rate setting, all three variants use the same 40-step candidate bank, Gaussian-count budget, tree search, and coding settings. We vary only the allocation score. The 160-step reference measures long-fit distortion for every admitted candidate with K>1 and also encodes the selected 40-step fits.
Paired values are reported in low-rate / high-rate order. Changing the allocation score improves decoded PSNR by 3.39 dB at low rate and 4.50 dB at high rate and improves MS-SSIM at both settings, with about 9 s of additional encoding time per image. Agreement with the measured 160-step ranking also rises from 64.1% to 91.6% on held-out blocks. The decoded gains show that calibration helps select among the available 40-step fits at the tested Gaussian budgets.
We added the controlled comparison in More ablation studies → Teacher calibration in the revised appendix.
W2&Q2
Thank you for pointing this out. We have added the measured bpp for every operating point in Table 1. The Low-rate and High-rate columns report two evaluated configurations per method. Their achieved bitrates vary across methods, so the table reports the quality of each configuration at its measured rate.
High-rate comparison at similar measured bitrates. On standard images, CaliSplat and Image-GS operate at 0.903 and 0.900 bpp in Table 1, a difference of only 0.003 bpp. At these closely matched rates, CaliSplat achieves 0.87 dB higher PSNR, with MS-SSIM of 0.9655 versus 0.9607. On large images, CaliSplat reaches 31.86 dB at 0.856 bpp, compared with 30.55 dB at 0.917 bpp for SGI, the strongest baseline by PSNR.