Streaming 3D Gaussian Splatting requires highly scalable, progressive representations. Existing progressive methods rely on discrete layering, accumulating separate splat sets for each level of detail. This structural independence between layers inherently leads to error accumulation, severe splat redundancy, and uncontrolled quality transitions.

We propose EvoGS, the first continuous-layering representation. Organized as an Evolution Tree, EvoGS generates finer details via an explicit, wavelet-inspired parent–child refinement. This empowers child nodes to structurally correct ancestral errors and yields inherently sparse, highly compressible inter-layer signals.

Extensive experiments show EvoGS eliminates splat redundancy from over 65% to under 25%. Compared to state-of-the-art baselines, it reduces transmission payload and GPU VRAM footprint by up to 2.4× and 5.5×, respectively, and achieves smooth quality transitions optimal for real-time adaptive streaming.

Immersive XR streaming faces heterogeneous devices, interactive 6DoF navigation, and dynamic bandwidth/memory budgets. A streamable 3DGS representation must therefore offer quality scalability, rendering scalability, and real-time adaptation. Existing progressive 3DGS methods all share a common construction principle — each LOD is an independent splat set trained for its target resolution. We term this discrete layering.

Discrete layering's inter-layer relationship is compositional, not evolutionary: higher layers visually mask lower-layer errors via 2D blending rather than correcting them. This compositional construction creates a cascade of problems: error accumulation, severe splat redundancy (>65% transparent ghost splats at fine LODs), uncontrolled quality transitions, and unexploitable inter-layer correlation.

Let \(\boldsymbol{P}\in\mathbb{R}^D\) denote a parent splat's parameters (position, rotation, scale, opacity, SH). We evaluate four constructions for its two children \(\boldsymbol{C_1},\boldsymbol{C_2}\):

Options A and B place no structural constraint on parent–child proximity, so coarse renderings are unreliable and the refinement signal is unstructured. Option C enforces a symmetric collinear residual, \[ \boldsymbol{C_1}=\boldsymbol{P}+\boldsymbol{\psi},\qquad \boldsymbol{C_2}=\boldsymbol{P}-\boldsymbol{\psi}, \] which is exactly the Haar wavelet lifting scheme: \(\boldsymbol{P}\) is the approximation coefficient and \(\boldsymbol{\psi}\) the detail coefficient. This guarantees sparse refinements but the rigid equal-magnitude constraint wastes capacity on real, non-symmetric scene details.

We therefore adopt Option D, the asymmetric collinear residual: \[ \boldsymbol{C_1}=\boldsymbol{P}+\boldsymbol{\psi},\qquad \boldsymbol{C_2}=\boldsymbol{P}-\boldsymbol{\alpha}\odot\boldsymbol{\psi}, \] where \(\boldsymbol{\alpha}\in\mathbb{R}^5\) is a learned per-attribute asymmetry factor. Paralleling JPEG 2000's biorthogonal lifting over orthogonal Haar wavelets, \(\boldsymbol{\alpha}\) trades exact mean preservation for representational power while preserving wavelet sparsity (see Inter-Layer Analysis below).

The Evolution Tree consists of root nodes (baseline geometry \(\boldsymbol{P_{\text{root}}}\in\mathbb{R}^D\)), internal nodes (learned refinements \(\boldsymbol{\psi}, \boldsymbol{\alpha}\)), and leaf nodes (renderable splats). A leaf's final parameters are accumulated along its ancestral chain: \[ \boldsymbol{S} \;=\; \boldsymbol{P_{\text{root}}} \;+\; \sum_{k=1}^{\ell}\boldsymbol{s_k}\odot\boldsymbol{\psi_k},\qquad \boldsymbol{s_k}\in\{+\mathbf{1},\,-\boldsymbol{\alpha}_k\}. \]

Progressive training. The base model \(L_0\) is trained at the coarsest resolution \(\mathbf{D}_0\) and each splat becomes a root. At every subsequent level \(L_i\), gradient-driven densification splits selected leaves into two children whose \(\boldsymbol{\psi}\) and \(\boldsymbol{\alpha}\) are unfrozen, while all ancestors remain fixed. The resulting tree is unbalanced and spatially adaptive: complex regions densify deeply, smooth regions stay shallow.

Rendering. At inference, the client requests a quality level \(L_i\); the active leaf set \(\mathcal{F}_i\) is reconstructed via the equation above and passed to the standard 3DGS rasterizer. Because nodes role-swap (leaf at coarse, internal at fine), one stored tree serves all quality tiers, and spatial regions can even be rendered at mixed qualities simultaneously — structurally impossible in discrete layering.

We evaluate on Synthetic Blender, Mip-NeRF360, Tanks&Temples, and Deep Blending against Monolithic, Single, LapisGS, and L3GS. Two size metrics are reported: Storage (cumulative transmitted bytes) and Memory (rendering GPU footprint, strictly smaller than storage for EvoGS because internal-node \(\boldsymbol{\psi}, \boldsymbol{\alpha}\) are consumed during leaf reconstruction).