Decoupling Geometry from Appearance: How Textured Gaussians Break the 4K Barrier in Feed-Forward Splatting

Feed-forward 3D reconstruction has matured remarkably over the past two years. Methods that once required minutes or hours of per-scene optimization can now predict a complete radiance field from a handful of images in a single network forward pass. Yet a quiet ceiling has constrained nearly all of this work: resolution. Most feed-forward Gaussian Splatting systems operate at input and output resolutions measured in the hundreds of pixels, far below the 4K imagery that modern cameras produce by default. The paper "Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting" (LGTM, ICLR 2026) identifies the structural reason for this ceiling and proposes an architecture that removes it. The core argument is one I find compelling precisely because it questions an assumption almost everyone has accepted without examination.

The Quadratic Trap of Pixel-Aligned Primitives

The dominant recipe for feed-forward Gaussian Splatting, established by pixelSplat, MVSplat, and later NoPoSplat, is to predict one Gaussian primitive per input pixel. This pixel-aligned formulation is elegant. The network outputs a depth, an offset, a covariance, an opacity, and a color for every pixel, and the resulting Gaussians are naturally distributed across the visible surfaces. It also plays well with standard convolutional and transformer backbones that already operate on pixel grids.

The problem is that this design ties primitive count directly to input resolution, and resolution scales quadratically with area. Moving from 512 to 4K multiplies the pixel count, and therefore the Gaussian count, by roughly 64. The LGTM authors quantify the consequence directly. NoPoSplat, when pushed beyond roughly 1024 by 576, runs out of memory entirely. At 2K and 4K it simply fails to train. Their Table 1 shows NoPoSplat consuming 61.85 GB at 1024 by 576 with batch size 1, while going out of memory at any higher target resolution.

There is a second, subtler inefficiency. Standard 3D Gaussian Splatting couples appearance and geometry within each primitive: each Gaussian carries a single view-dependent color. To represent a richly textured but geometrically flat surface, such as a brick wall or a printed poster, you need many Gaussians purely to paint the texture, even though the underlying geometry is a single plane. This is wasteful. The geometry is trivial; the appearance is complex; and the pixel-aligned formulation forces you to spend geometric capacity to capture appearance detail.

LGTM's Dual-Network Decoupling

The central idea in LGTM is to stop conflating these two quantities. Geometry and appearance have fundamentally different spatial complexity, so the paper assigns them to two separate networks operating at two different resolutions.

• A primitive network processes low-resolution inputs (for example 512 by 288 or even 256 by 144) and predicts a compact set of geometric Gaussian primitives. This captures shape, which is comparatively smooth and low-frequency.
• A texture network processes high-resolution inputs and predicts a per-primitive texture map for each Gaussian. Instead of a single color per Gaussian, each primitive carries a small texture (2 by 2, 4 by 4, or 8 by 8) that modulates appearance across its footprint during rendering.

The texture network extracts high-resolution features through image patchification and projective mapping, then fuses those features with the geometric features from the primitive network. The effective rendering resolution becomes the product of the primitive resolution and the texture size. A 512 by 288 primitive grid with an 8 by 8 texture per primitive yields a 4096 by 2304 output, which is genuine 4K, while the number of geometric primitives stays at the low-resolution count.

The training follows a staged strategy. The primitive network is pre-trained first to establish a robust geometric foundation, then jointly trained with the texture network so the appearance branch can layer in high-frequency detail without destabilizing the geometry. This ordering matters: if you train both branches jointly from scratch, the texture network can mask geometric errors by overfitting appearance, producing primitives that look correct from training views but break under novel viewpoints.

The numbers support the design. At a 4096 by 2304 output, LGTM achieves LPIPS 0.200, SSIM 0.803, PSNR 24.489 using only 28.23 GB of training memory, a regime where NoPoSplat cannot run at all. Even at the resolution where NoPoSplat does function, 1024 by 576, LGTM with a 512 by 288 primitive grid and 2 by 2 textures improves every metric while cutting memory from 61.85 GB to 20.16 GB. That is roughly a third of the memory for better LPIPS, SSIM, and PSNR simultaneously.

Why This Is More Than an Engineering Trick

It would be easy to read LGTM as a clever way to save memory. I think the more important contribution is conceptual. The paper is the first feed-forward network to predict textured Gaussians, and in doing so it demonstrates that the per-scene optimization requirement of prior textured Gaussian methods (such as those by Xu, Chao, Rong, and others) was not intrinsic to textured primitives. Those earlier methods could not generalize across scenes because they optimized textures per scene; LGTM shows the texture maps themselves can be predicted by a network that learns cross-scene priors.

There is also a genuine architectural insight about matching network resolution to signal complexity. Geometry in most scenes is dominated by smooth surfaces, so spending a high-resolution network on geometry is largely wasted compute. Appearance carries the high-frequency information, so that is where resolution should be allocated. LGTM essentially performs a frequency-domain division of labor across two networks. This echoes ideas from classical graphics, where geometry meshes are coarse and texture maps are fine, but it is one of the first times this separation has been brought cleanly into the feed-forward neural reconstruction setting.

I would also point out a practical consequence the paper hints at through its versatility claims. Because the texture network is a bolt-on module that fuses with geometric features, LGTM improves existing baselines (Flash3D, DepthSplat, NoPoSplat) across monocular, two-view, and multi-view settings, with or without known camera poses. This suggests textured splatting is a general augmentation rather than a single-model design, which is the more durable kind of contribution.

Limitations and Open Questions

The decoupling is not free of caveats. A few stand out to me.

First, per-primitive textures assume that appearance within a primitive's footprint can be represented as a small fixed-resolution grid. For surfaces with view-dependent specular highlights or fine geometric detail mistaken for texture, this assumption could fail. The texture map can fake high-frequency appearance, but it cannot recover true parallax that depends on actual geometry. There is a real risk of geometry-appearance ambiguity at silhouettes and thin structures, where what looks like texture detail is actually missing geometry.

Second, the reported PSNR values, hovering around 24 to 25, are respectable but not dramatically high for 4K targets. This suggests the textured representation captures perceptual quality (LPIPS and SSIM are strong) more than pixel-exact fidelity. For applications requiring precise reconstruction rather than perceptually convincing renders, this distinction matters.

Third, the staged training and dual-network design add complexity. The relationship between primitive resolution and texture size introduces a hyperparameter trade-off: coarser primitives with larger textures save geometric capacity but risk appearance bleeding across geometric boundaries, visible in the slightly worse metrics for the 256 by 144 variants.

Closing Thoughts

LGTM's contribution is to identify that pixel-aligned Gaussians, the unexamined default of feed-forward splatting, are the precise mechanism that makes 4K intractable, and then to dissolve that constraint by decoupling geometry from appearance. The lesson generalizes beyond Gaussian Splatting: whenever a representation conflates two signals of different complexity, scaling one inevitably wastes capacity on the other.

The open questions I find most interesting concern how far this decoupling can go. Could the texture maps become learned neural fields rather than fixed grids, recovering view-dependent and high-frequency effects more faithfully? Could the same principle extend to dynamic scenes, where geometry and appearance evolve at different temporal frequencies? If the answer is yes, the phrase "fewer Gaussians, more texture" may describe a broader direction for efficient neural rendering, not just a single architecture.