Human Gaussian Graph for Avatar Modeling

• Human Gaussian Graph is a dual-layer graph architecture that aggregates per-frame 3D Gaussian predictions via a shared SMPL mesh scaffold to create a canonical human avatar.
• It employs intra-node attention and inter-node message passing to integrate temporal evidence and refine local body structure for pose-driven animation.
• Empirical evaluations reveal improved novel view synthesis and pose animation performance, with PSNR gains of approximately 1.6 to 2.3 dB over prior methods.

Human Gaussian Graph (HGG) is a dual-layer graph representation for learning animatable human avatars from video by reorganizing independently predicted 3D Gaussians around a shared SMPL scaffold. In its precise usage, the term denotes the mechanism introduced in “Learning Efficient and Generalizable Human Representation with Human Gaussian Model,” where per-frame Gaussian predictions are no longer treated as isolated outputs, but are attached to canonical SMPL mesh vertices and processed by attention-based aggregation and message passing to produce a temporal-invariant, SMPL-aligned Gaussian avatar within inference time (Liu et al., 24 Jul 2025). In a broader interpretive sense, adjacent work on human Gaussian modeling has also used mesh-aligned, semantic, or graph-like organizations of Gaussian primitives, but HGG is distinguished by explicit cross-frame aggregation through a dual-level graph whose first layer is composed of Gaussian nodes and whose second layer is composed of mesh-vertex nodes.

1. Problem formulation and motivation

Modeling animatable human avatars from videos is described as a long-standing and challenging problem. Conventional methods require per-instance optimization, while more recent feed-forward methods generate 3D Gaussians with a learnable network. The limitation identified by HGG is that these feed-forward methods “predict Gaussians for each frame independently, without fully capturing the relations of Gaussians from different timestamps” (Liu et al., 24 Jul 2025). HGG is introduced specifically to address that limitation by aggregating cross-frame evidence, aligning the predicted Gaussian representation with a shared SMPL human prior, and producing an animatable, SMPL-aligned Gaussian avatar.

The formal input is a video $\{I^{t}\}_{t=1}^{T}$. For each frame, the method first predicts a Gaussian set

$$G^{t}=\{g_m^t\}_{m=1}^M=\Phi_{\text{G}}(I^{t}), \quad g_m^t=\{\mu_{m}^t, f_{m}^t\},$$

where $\mu_m^t\in\mathbb{R}^3$ is the Gaussian center and $f_m^t\in\mathbb{R}^C$ is its feature vector. The first-layer node set is

$\mathcal{G}=\{G^{t}\}_{t=1}^{T},$

and the second-layer node set is the SMPL mesh vertices

$\mathcal{V}=\{v_i\}_{i=1}^{N}.$

The SMPL vertices are described as “equivalent throughout the temporal axis,” so the same mesh vertices provide a stable scaffold across all frames.

This construction indicates that the core objective of HGG is not merely Gaussian prediction, but temporal integration. A plausible implication is that the graph functions as the missing correspondence mechanism between per-frame Gaussian inference and pose-driven avatar recovery. The paper’s framing is explicit on this point: the graph converts independently predicted per-frame Gaussians into a temporally integrated human representation that can be driven by pose (Liu et al., 24 Jul 2025).

2. Dual-layer graph construction

HGG is built from two edge types: Gaussian-mesh edges $\mathcal{E}_{vg}$ and mesh-level edges $\mathcal{E}_{vv}$ (Liu et al., 24 Jul 2025). The first reorganize per-frame Gaussians by shared SMPL anchors; the second encode local body topology.

For each frame $t$, the SMPL template vertices are posed by linear blend skinning:

$$\mathcal{V}^{t}=\{v_{i}^{t}\}_{i=1}^{N} = \psi_{\text{LBS}}(\mathcal{V},p^t),$$

where $$G^{t}=\{g_m^t\}_{m=1}^M=\Phi_{\text{G}}(I^{t}), \quad g_m^t=\{\mu_{m}^t, f_{m}^t\},$$0 denotes shape and pose parameters, and LBS is defined as

$$G^{t}=\{g_m^t\}_{m=1}^M=\Phi_{\text{G}}(I^{t}), \quad g_m^t=\{\mu_{m}^t, f_{m}^t\},$$1

Each Gaussian $$G^{t}=\{g_m^t\}_{m=1}^M=\Phi_{\text{G}}(I^{t}), \quad g_m^t=\{\mu_{m}^t, f_{m}^t\},$$2 is then assigned to the closest posed vertex:

$$G^{t}=\{g_m^t\}_{m=1}^M=\Phi_{\text{G}}(I^{t}), \quad g_m^t=\{\mu_{m}^t, f_{m}^t\},$$3

This assignment is the decisive reorganization step. Instead of preserving a frame-indexed Gaussian set, the method groups Gaussians spatially by shared SMPL vertices, so Gaussians from different timestamps that lie near the same body region are routed to the same vertex node. The second edge type, mesh-level edges, follows the SMPL topology:

$$G^{t}=\{g_m^t\}_{m=1}^M=\Phi_{\text{G}}(I^{t}), \quad g_m^t=\{\mu_{m}^t, f_{m}^t\},$$4

These edges connect neighboring SMPL vertices according to graph distance on the mesh and thereby encode local human-body structure.

The graph is therefore “dual-layer” in a literal sense. The Gaussian layer contains observations collected across time, while the mesh layer provides canonical body anchors. This suggests that HGG treats temporal redundancy as a structured signal rather than as a set of independent reconstruction attempts.

3. Intra-node aggregation and inter-node message passing

Graph processing in HGG is performed by two complementary transformer-like operations: the intra-node operation and the inter-node operation (Liu et al., 24 Jul 2025). The intra-node stage aggregates temporal information from multiple Gaussian nodes into a single SMPL vertex node; the inter-node stage propagates the resulting information across neighboring mesh vertices.

For a vertex $$G^{t}=\{g_m^t\}_{m=1}^M=\Phi_{\text{G}}(I^{t}), \quad g_m^t=\{\mu_{m}^t, f_{m}^t\},$$5, the Gaussian neighborhood is defined as

$$G^{t}=\{g_m^t\}_{m=1}^M=\Phi_{\text{G}}(I^{t}), \quad g_m^t=\{\mu_{m}^t, f_{m}^t\},$$6

A learnable query $$G^{t}=\{g_m^t\}_{m=1}^M=\Phi_{\text{G}}(I^{t}), \quad g_m^t=\{\mu_{m}^t, f_{m}^t\},$$7 is attached to each mesh vertex and updated using cross-attention:

$$G^{t}=\{g_m^t\}_{m=1}^M=\Phi_{\text{G}}(I^{t}), \quad g_m^t=\{\mu_{m}^t, f_{m}^t\},$$8

with projections

$$G^{t}=\{g_m^t\}_{m=1}^M=\Phi_{\text{G}}(I^{t}), \quad g_m^t=\{\mu_{m}^t, f_{m}^t\},$$9

followed by

$\mu_m^t\in\mathbb{R}^3$0

Conceptually, each query acts as a learned collector that summarizes all Gaussians assigned to one vertex, regardless of frame. This is the explicit cross-frame aggregation mechanism.

The inter-node operation uses the mesh neighborhood

$\mu_m^t\in\mathbb{R}^3$1

and updates the same query by attention over neighboring vertex features:

$\mu_m^t\in\mathbb{R}^3$2

then

$\mu_m^t\in\mathbb{R}^3$3

This stage performs local message passing across body-part neighbors. The paper states that it helps smooth noise, reinforce structure, and allow information from one body region to benefit nearby regions. Multiple such blocks can be stacked, increasing the receptive field of each vertex, analogously to deeper graph convolution or message passing on a mesh.

The supplementary analysis quantifies the efficiency rationale: because each vertex only attends to its associated Gaussians and to a small neighborhood of mesh vertices, the computation is far cheaper than naïve attention over all Gaussians across all frames. The graph is therefore both a structural prior and a computational sparsification strategy.

4. Canonical SMPL-aligned Gaussians and pose-driven animation

After vertex-level features are refined, HGG uses them to refine one chosen frame’s Gaussian representation into the final SMPL-aligned Gaussians (Liu et al., 24 Jul 2025). Given a reference frame $\mu_m^t\in\mathbb{R}^3$4, each Gaussian is updated by attending to the corresponding vertex query:

$\mu_m^t\in\mathbb{R}^3$5

where

$\mu_m^t\in\mathbb{R}^3$6

The resulting set

$\mu_m^t\in\mathbb{R}^3$7

is described as temporal-invariant.

This temporal-invariant status is central to the method’s notion of animation. The graph has distilled multi-frame evidence into a canonical Gaussian avatar that is no longer tied to any single input frame. At inference time, this avatar can be re-posed using the target SMPL pose $\mu_m^t\in\mathbb{R}^3$8 and then rendered with standard 3D Gaussian rasterization.

The training objective for the graph-produced SMPL-aligned Gaussians uses the same rendering supervision as LGM-style pipelines:

$\mu_m^t\in\mathbb{R}^3$9

where $f_m^t\in\mathbb{R}^C$0 are ground-truth RGB and alpha images, and $f_m^t\in\mathbb{R}^C$1 are the rendered outputs. HGG is thus optimized end-to-end under photometric supervision through the rendered avatars, rather than trained only as an auxiliary graph module.

A common misconception is that HGG is merely a graph regularizer placed on top of a Gaussian predictor. The method is more specific: it is the mechanism by which per-frame predictions are converted into an animatable, SMPL-aligned avatar. In that sense, the graph is not peripheral but constitutive of the representation.

5. Empirical performance and ablation behavior

The reported experiments indicate that HGG achieves the best results in both novel view synthesis and novel pose animation in the main quantitative comparison (Liu et al., 24 Jul 2025). In the single-view setting, it improves over ExAvatar by about 1.6 dB PSNR on novel view synthesis, reaching 23.112 PSNR versus 21.450, and it improves novel pose animation to 21.857 PSNR. In the multi-view setting, the gains are larger: the method reaches 26.536 PSNR on novel view synthesis and 24.013 PSNR on novel pose animation, outperforming GPS-Gaussian and 4DGS by a wide margin. The paper also reports an approximately 2.3 dB PSNR advantage on novel pose animation over the compared methods, emphasizing the effectiveness of HGG-derived SMPL-aligned Gaussians for animation.

The ablation study is organized around the graph’s two operators and the depth of the graph stack. Removing either the intra-node or the inter-node operation reduces performance, while stacking more HGG blocks improves results further. The intra-node operation contributes a particularly large gain, which is consistent with the claim that cross-frame aggregation is central to reconstruction quality. The inter-node operation is described as cleaning up noise and aliasing in local regions.

These results are significant because the method’s improvements are linked directly to its graph design rather than to a change in rendering objective. This suggests that the principal benefit comes from reorganizing information flow across time and across the body surface. A plausible implication is that HGG’s main contribution lies in representation learning rather than in loss engineering.

6. Position within human Gaussian modeling

HGG emerges within a rapidly expanding literature on human-centric Gaussian representations, but it occupies a distinct position. GPS-Gaussian predicts pixel-wise Gaussian parameters from source images, lifts them into 3D using depth, aggregates Gaussians from two views, and renders novel views without per-subject optimization; however, it does not explicitly model cross-frame correspondences or information sharing across timestamps (Zheng et al., 2023). HUGS represents an animatable human and scene with 3D Gaussian Splatting, initializes human Gaussians from SMPL, allows them to deviate from the body model, and jointly optimizes linear blend skinning weights, but it is person-specific and optimized per subject (Kocabas et al., 2023). HFGaussian extends generalizable Gaussian human modeling to predict 3D skeleton, 3D keypoints, and dense pose in real time, attaching feature embeddings to Gaussians and decoding them after splatting (Dey et al., 2024).

Other work has introduced graph-like or explicitly graph-structured variants of human Gaussian modeling. “Gaussian Control with Hierarchical Semantic Graphs in 3D Human Recovery” organizes Gaussians into a hierarchical semantic graph over body parts and topological relations to preserve joint boundaries and fine surface detail (Wang et al., 2024). HiFi4G uses a dual-graph mechanism consisting of a coarse deformation graph and a fine-grained Gaussian graph to obtain motion priors and temporal coherence for compact 4D human performance rendering (Jiang et al., 2023). “Generalizable Human Gaussians from Single-View Image” combines image-driven Gaussian prediction, SMPL-guided feature propagation, and diffusion-guided back-view refinement; its representation is described as a structured, prior-aware, attention-connected human Gaussian field rather than an explicit graph in the HGG sense (Chen et al., 2024).

Against this background, HGG’s distinguishing property is the joint use of all video frames through a shared SMPL-mesh graph. It is neither an unstructured Gaussian cloud nor a purely per-subject optimization scheme. Nor is it a semantic graph over body parts or a temporal regularizer over already-optimized Gaussians. Its defining operation is the reorganization of independently predicted per-frame Gaussians into vertex-centered temporal neighborhoods, followed by vertex-level attention and mesh-level message passing. That specific design is what allows the method to claim a bridge between generalizable feed-forward Gaussian prediction and pose-driven human representation learning (Liu et al., 24 Jul 2025).

7. Conceptual significance and scope

The conceptual significance of Human Gaussian Graph lies in how it reframes correspondence. Instead of seeking direct pointwise matching among Gaussians across time, it uses the SMPL mesh as a stable intermediate scaffold. Temporal evidence is first routed to canonical vertex anchors, then summarized, propagated, and written back into a canonical Gaussian set. This suggests a broader principle for human Gaussian modeling: when temporal identity of primitives is unstable, correspondence can be mediated by a shared body prior rather than by direct Gaussian-to-Gaussian tracking.

The method also clarifies what “generalizable” means in this context. Earlier feed-forward Gaussian approaches demonstrated that Gaussian attributes can be regressed without per-subject optimization, while optimization-based avatar systems showed how explicit Gaussians can support animation and efficient rendering. HGG combines these two directions by taking feed-forward per-frame predictions and converting them into a temporally fused, mesh-aligned, animatable avatar (Liu et al., 24 Jul 2025). A plausible implication is that the graph formalism provides the missing structural layer required for generalizable Gaussian prediction to support robust pose-driven animation.

The term “Human Gaussian Graph” should therefore be understood primarily as a specific dual-layer graph architecture for cross-frame human Gaussian aggregation, not as a generic synonym for any human-centered Gaussian representation. In adjacent literature, the phrase can serve as an interpretive umbrella for graph-like organizations of human Gaussians, but in the strict technical sense it refers to the HGG mechanism that couples Gaussian nodes from all frames with canonical SMPL vertex nodes to produce a temporal-invariant, SMPL-aligned Gaussian avatar (Liu et al., 24 Jul 2025).