NOTE-256: Image-GS: the rate accounting GaussianToken skipped, and an ablation that inverts its own headline

Contribution

An explicit image representation: a set of anisotropic, colored 2D Gaussians, each with a mean, a covariance and a color vector, fitted per image by a custom differentiable renderer. Positions are sampled at initialization from a weighted mix of local image-gradient magnitude and a uniform constant, so more primitives land on high-frequency regions while the whole image stays covered. More Gaussians are added every 0.5K steps into regions with persistent error, starting from half the budget.

Key results

The comparison is run the way this record keeps asking for. Six neural image representations — ReLU-F, SIREN, FFN, WIRE, I-NGP and GI — plus JPEG, using the baselines’ official implementations with only model size changed to reach the target range. At 2K×2K the model sizes are 164, 166, 161, 154, 159, 164 and 160 KB for ReLU-F, I-NGP, SIREN, FFN, WIRE, GI and Image-GS. Bitrate is on the axis throughout, in bpp and bits per pixel per channel.

Image-GS outperforms all six across the evaluated bitrate range and passes JPEG below 0.244 bpp. The paper immediately qualifies its own win: JPEG and GI use entropy coding, which Image-GS does not, and entropy coding breaks the data locality that the random-access claim depends on. Naming what the baselines have that you lack, in the sentence announcing the win, is rarer than it should be.

Quality against rate, 45-image evaluation set:

bitrate	PSNR	MS-SSIM	LPIPS	FLIP
0.366 bpp	32.99 ± 4.49	0.966 ± 0.020	0.083 ± 0.057	0.078 ± 0.029
0.122 bpp	29.20 ± 4.57	0.924 ± 0.042	0.173 ± …	—

Decode cost is the number that carries the practice. 0.3K multiply- accumulates per pixel, against C3’s 3K MACs at 0.31 bpp — an order of magnitude — with hardware-friendly random access and a level-of-detail stack that falls out of the error-guided growth order.

The ablation inverts the headline. Table 1, on the same evaluation set:

variant	PSNR	MS-SSIM	LPIPS	cost against Full
Full	31.77 ± 4.73	0.960	0.102	—
No color init	30.40 ± 4.73	0.951	0.110	−1.37
No position init	29.88 ± 4.18	0.954	0.135	−1.89
Random init (both off)	29.54 ± 4.15	0.948	0.149	−2.23
No top-K normalization	29.35 ± 4.43	0.944	0.182	−2.42
No inverse scale	29.11 ± 4.58	0.933	0.152	−2.66

The content-adaptive allocation the paper is named for is worth 2.23 dB with both components removed. Optimizing 1/s instead of the raw Gaussian scale — a reparameterization justified in one sentence, because the natural range is [5,10] and the inverse lands in [0,1] where gradients are smoother — is worth 2.66. Top-K normalization is worth 2.42. Two implementation choices each outweigh the idea in the title.

Claims

id	claim	strength	support
C1	An explicit primitive set beats implicit coordinate networks for single images at low bitrate	strong	six baselines, official implementations, matched KB, bitrate axis
C2	Decode cost is an order of magnitude below a comparable neural codec	strong	0.3K vs 3K MACs/pixel at comparable bpp
C3	Content-adaptive allocation is what makes it work	overstated	worth 2.23 dB; two optimizer details are worth more, by the paper’s own Table 1
C4	The level-of-detail hierarchy is free	moderate	falls out of the growth order; no separate cost reported, and no comparison against a purpose-built LOD scheme

Limitations

Per-image fitting, not a learned codec. Every image is an optimization run — 5K Adam steps. Against autoencoder codecs trained across a dataset this is a different product with a different cost model, and the paper is comparing against overfit-per-image baselines, which is the right comparison for its claim and not the one someone shipping a general codec wants.

No seeds anywhere. Means and standard deviations are across the 45-image evaluation set, not across optimization runs. The ablation gaps are 1.4–2.7 dB against a per-image spread of ±4.5, so the image-to-image variance dominates and the run-to-run variance is unreported.

The ablation is one budget. Whether the ordering of those five choices holds at 0.122 bpp as well as at the ablation’s operating point is untested, and the interesting question — does the content-adaptive allocation matter more as the budget tightens, which is the regime the paper claims — is exactly the one the table does not answer.

Entropy coding is left on the table and acknowledged. The paper declines it for locality, which is a real engineering reason, and it means the reported bitrates are not the best this representation could do.

Bearing on the record

SOTA-205 gains a fifth source and its first outside 3D. Four radiance-field papers argued that capacity should move out of the network into a structure that is addressed rather than evaluated. This is the same claim for a single image, and 0.3K against 3K MACs per pixel is the cleanest statement of what “addressed rather than evaluated” buys.

The constructive contrast with LIT-494. GaussianToken landed two units earlier with comparisons matched on token count while each token carried an index plus five continuous floats, and NOTE-243 recorded that nothing ruled out the rival explanation that it simply had more channel. Image-GS is the same primitives in the same year with bits per pixel on the axis and model sizes matched in kilobytes. The two are not compared by anybody and no relation is declared — ADR-011 — but holding both is what makes the criticism of the first precise rather than general, and it is the answer to the open question NOTE-243 left.

Second instance of an ablation inverting its own headline, counted not promoted. SOTA-304 records the first: in Genesys, removing the symbolic checker costs 62 points of validity and removing the clever unit-wise decomposition costs 19, so the boring half is three times the lever. Here the reparameterization and the normalization each outweigh the content-adaptive allocation. Two instances from unrelated fields is a count. What would make it more is a mechanism, and “papers name themselves after the interesting half” is a sociological observation rather than one — DP-009.

No new practice. Content-adaptive densification is what LIT-108 already does in 3D and what SOTA-205 already covers at the level that matters; a 2D instance adds a source, not a recommendation. The 1/s reparameterization is the most surprising number here and rests on one sentence in one paper about one parameter — an open question below rather than a practice.

Open questions

• Does the 1/s reparameterization generalize? It is worth more than the paper’s headline idea and is justified in a single sentence: optimize the inverse when the natural range sits away from zero, because [0,1] gives smoother gradients. If that holds for other bounded-away-from-zero parameters — scales, temperatures, bandwidths — it is a practice; here it is one measurement on one parameter.
• Does content-adaptive allocation matter more as the budget tightens? The paper’s claim is about low-bitrate regimes and its ablation is at one operating point.
• What does the rate curve look like against GaussianToken’s task? Nobody has run a Gaussian image representation and a Gaussian image tokenizer against each other at matched bits. Both exist, both are public, and the comparison is the one NOTE-243 asked for.