Topology Sculptor, Shape Refiner (TSSR) Overview

Updated 5 July 2026

TSSR is a topology-aware decomposition approach that separates structural connectivity repair from geometric fidelity refinement in 3D mesh generation.
It leverages discrete diffusion models with mask-noise and uniform-noise denoising to concurrently preserve local topology and recover global shape.
The framework extends across related methods in topology optimization, implicit editing, and shape transforms, demonstrating improved accuracy and efficiency on multiple datasets.

Searching arXiv for recent and directly relevant papers on “Topology Sculptor, Shape Refiner (TSSR)” and adjacent methods. Topology Sculptor, Shape Refiner (TSSR) denotes a topology-aware decomposition in which one stage establishes, repairs, or constrains structural connectivity and a subsequent stage improves geometric fidelity. In the most explicit arXiv usage, TSSR is a discrete diffusion framework for direct generation of artist-style 3D meshes, where mask-noise denoising functions as a topology sculpting stage and uniform-noise denoising functions as a shape refinement stage (Song et al., 24 Oct 2025). The same two-part logic also appears in earlier and adjacent work: a density-based topology optimization post-processor first extracts a smooth geometry from a raw density field and then re-optimizes it by shape optimization (Swierstra et al., 2020). This suggests that TSSR can be read both as a specific generative model and as a broader methodological pattern in topology-aware geometry processing.

1. Scope and meanings of the term

The phrase has one direct and one broader usage. Directly, it names a discrete diffusion model for high-fidelity direct 3D mesh generation (Song et al., 24 Oct 2025). More broadly, it aligns with methods that separate topology-sensitive operations from subsequent geometric or representational refinement, including density-to-CAD post-processing, topology-aware implicit editing, and topological shape transforms (Swierstra et al., 2020).

Paper	Representation	Topology/shape split
(Song et al., 24 Oct 2025)	Discrete mesh-token diffusion	Mask-noise topology sculpting; uniform-noise shape refining
(Swierstra et al., 2020)	RBF level-set + FCM	Geometry extraction; shape optimization
(Zhang et al., 2024)	SDF as cubic trivariate tensor-product B-spline	Saddle-point topology editing; local geometry editing
(Yang et al., 21 Feb 2026)	ECT pushforward measure / SampEuler	Topology-aware shape summarization; vectorized interpretation

A common misconception is that TSSR refers only to cosmetic smoothing. That is not how the relevant papers frame the problem. In the topology optimization setting, the stated objective is not merely to “beautify” a jagged density field, but to obtain unambiguous, accurate and optimized geometries from arbitrary 3D topology optimization results (Swierstra et al., 2020). In the diffusion setting, the split is likewise functional rather than stylistic: local topology and global shape are assigned to different denoising regimes (Song et al., 24 Oct 2025).

2. Discrete diffusion TSSR for direct 3D mesh generation

In its primary arXiv sense, TSSR addresses high-fidelity direct 3D mesh generation in the artist-style regime, where meshes must be both topologically valid and geometrically faithful at very high resolution (Song et al., 24 Oct 2025). The method is motivated by the limitations of autoregressive mesh generators: sequential token prediction is slow for long sequences, difficult at high spatial resolution, and constrained by sequential dependence. TSSR instead uses discrete diffusion models (DDMs), which generate tokens in parallel and allow the model to attend to all mesh tokens concurrently.

The mesh representation is token-based. A triangular mesh $M$ is represented as $N$ faces, each with 3 vertices, and each vertex has $(x,y,z)$ coordinates quantized into $V$ bins. With the Meshtron tokenizer, the mesh is flattened into a 1D token sequence of length $9N$, and the codebook size is $B=V+k$ , with $V=1024$ and $k$ accounting for special tokens such as mask, padding, BOS, and EOS (Song et al., 24 Oct 2025). This discrete formulation is the basis for the diffusion process.

TSSR’s central claim is that different corruption processes induce complementary priors. Mask corruption behaves like infilling, so the model learns to fill missing pieces, preserve local connectivity, and repair holes; this is the role of the “Topology Sculptor.” Uniform corruption replaces tokens with random valid tokens and produces a “hedgehog”-like, globally broken shape; recovering from this forces the model to learn overall shape, which is the role of the “Shape Refiner” (Song et al., 24 Oct 2025). The architecture therefore assigns $\theta_{\text{mask}}$ to topology sculpting and $\theta_{\text{uniform}}$ to shape refining.

Inference is hybrid and iterative. The process starts from a pure mask sequence $N$ 0. A mask-corrupted state $N$ 1 is denoised first, yielding a cleaner estimate. That estimate is then treated as the input to the uniform-noise branch, producing a refined sample. A classifier $N$ 2 identifies low-confidence tokens, and tokens below a threshold $N$ 3 are replaced again by $N$ 4; this loop repeats for $N$ 5 inference steps (Song et al., 24 Oct 2025). The intended effect is to retain the parallelism of diffusion while reintroducing topology-sensitive structure and global shape recovery through specialized denoisers.

3. Architecture and topological regularization

The method is organized around three components: decoupled training and hybrid inference, an improved Hourglass backbone, and a connection loss (Song et al., 24 Oct 2025). The design objective is to make non-causal diffusion workable for long mesh sequences without sacrificing topological validity.

The improved Hourglass architecture addresses the information bottleneck induced by sequence compression. The paper adds bidirectional cross-attention, described as multi-level conditioning from the compressed level to the uncompressed level, so that information can flow more effectively across representation scales. This is paired with multi-level rotational positional embeddings (RoPE), using a structured positional tuple

$N$ 6

rather than a flat 1D position index (Song et al., 24 Oct 2025). The motivation is mesh-specific: coordinate-level structure supports precise value prediction, vertex-level grouping supports shared geometry, and face-level structure supports local topology.

The training objective is likewise decomposed. TSSR uses a mask denoising loss, a uniform denoising loss, and a classifier loss, combined as

$N$ 7

and then extended to

$N$ 8

A unified model handles both denoising modes through a task flag $N$ 9, with $(x,y,z)$ 0 for mask denoising and $(x,y,z)$ 1 for uniform denoising, while a lightweight classifier head serves as $(x,y,z)$ 2 (Song et al., 24 Oct 2025).

The connection loss is the explicit topology prior. It is based on the mesh-specific constraint that tokens belonging to the same shared vertex should agree. Shared-vertex groups are constructed, logits for repeated occurrences of the same vertex are collected by coordinate dimension, and a group-wise consistency penalty is averaged into a batch-level loss (Song et al., 24 Oct 2025). The stated purpose is to suppress disconnected vertices, non-manifold-like inconsistencies, fragile triangles, and related topology errors. A recurrent misconception in diffusion-based geometry generation is that parallel token prediction is necessarily incompatible with topology; TSSR’s answer is not to restore causality, but to impose topological agreement at the level of shared-vertex token groups.

4. Reported results, efficiency, and limitations

The reported training corpus contains 400K meshes from Objaverse-XL, ShapeNet, and other licensed datasets, and the model size is 2B (Song et al., 24 Oct 2025). Training is described as 20 days on 32 NVIDIA A100 40GB GPUs for the first stage, followed by 7 days on 16 NVIDIA H20 96GB GPUs when scaling to face counts in the range $(x,y,z)$ 3. The optimizer is AdamW, the learning-rate schedule is linear with warmup, the peak learning rate is $(x,y,z)$ 4, and inference uses $(x,y,z)$ 5 steps.

Evaluation follows prior mesh generation work and uses Hausdorff Distance (HD), $(x,y,z)$ 6, $(x,y,z)$ 7, Normal Consistency (NC), and F-score (F1). Baselines include MeshAnythingV2, BPT, TreeMeshGPT, and DeepMesh (Song et al., 24 Oct 2025).

Dataset	Distance metrics	Surface metrics
ShapeNet	HD 0.050; $(x,y,z)$ 8 0.013; $(x,y,z)$ 9 0.035	NC 0.954; F1 0.976
Thingi10K	HD 0.106; $V$ 0 0.032; $V$ 1 1.183	NC 0.969; F1 0.775
Objaverse	HD 0.101; $V$ 2 0.014; $V$ 3 0.863	NC 0.959; F1 0.744

The paper reports the best overall results on all three datasets and emphasizes particularly strong F1 and normal consistency (Song et al., 24 Oct 2025). For 2K-face meshes on a single H20 GPU, the reported generation times are 73 s for DeepMesh, 79 s for BPT, and 51 s for TSSR, which the paper attributes to parallel token generation.

The ablations support the two main architectural changes. On ShapeNet, adding cross-attention yields HD 0.1338 and $V$ 4 0.025, while adding both cross-attention and RoPE yields HD 0.050 and $V$ 5 0.013 (Song et al., 24 Oct 2025). The connection-loss ablation is more ambiguous: the rendered table is internally inconsistent with the textual discussion, but the accompanying qualitative claim is explicit that the connection loss improves topology quality and suppresses fragile or invalid triangles. The stated limitation is training inconvenience: DDM training requires padding sequences to accommodate large batch sizes, which increases computational cost; infilling is suggested as future work.

5. Earlier topology-optimization workflow associated with TSSR

A closely related use of the topology-sculptor/shape-refiner pattern appears in the fully automated post-processing of density-based topology optimization results (Swierstra et al., 2020). The problem is that SIMP-type topology optimization produces jagged boundaries and intermediate densities, which are awkward for CAD and inaccurate for analysis. The proposed remedy is a two-stage pipeline: geometry extraction followed by shape optimization.

The geometry extraction stage represents the boundary implicitly using a level-set function $V$ 6 built from Gaussian radial basis functions,

$V$ 7

with $V$ 8, where $V$ 9 is the centroid of element $9N$0, $9N$1 is the basis-function weight, and $9N$2 is a global shift (Swierstra et al., 2020). The optimized density field is converted into this level-set representation by solving a linear system $9N$3, with $9N$4. For computational efficiency the Gaussian RBFs are truncated to a practical support diameter of 7 elements, which makes $9N$5 sparse.

The sign of the level-set function defines solid and void, and a smooth Heaviside projection reconstructs density from the implicit geometry. Because the unshifted field is positive everywhere if the topology-optimization densities are positive, the global shift $9N$6 is chosen so that the zero contour satisfies the target volume fraction; the paper solves this with Newton–Raphson and then fixes $9N$7 for the remainder of the process (Swierstra et al., 2020). This is described not as a visualization trick but as a true geometry extraction step: it removes gray regions, preserves the target volume, and yields a clean implicit boundary.

The second stage treats the RBF weights as design variables and re-optimizes the extracted geometry. Rather than remeshing, the paper uses p-FEM and the Finite Cell Method (FCM), with local analysis refinements near the structural boundary: one level of quadtree refinement in 2D and one level of octree refinement in 3D, extending about one element beyond the current boundary (Swierstra et al., 2020). The paper’s distinctive technical point is that bounds on basis-function weights imply bounds on the maximum possible level-set slope, which in turn guarantees sufficient sensitivity information throughout the shape optimization process. The explicit motivation is to avoid the “frozen boundary” problem caused by overly steep level-set transitions.

This workflow is not characterized as passive smoothing. The paper states that Stage 2 and Stage 3 do not merely smooth the topology-optimization result; they recover feasibility and improve performance. Reported examples include the MBB beam, cantilever beam, 3D MBB beam, and 3D cantilever beam, and the final designs generally outperform the original Stage 1 topology-optimization designs by about 10–15% (Swierstra et al., 2020). In that specific sense, the 2020 pipeline already instantiated the logic later named “Topology Sculptor” and “Shape Refiner.”

6. Topology-aware editing and topological shape transforms

A second adjacent line of work is real-time topology-aware implicit editing (Zhang et al., 2024). Here the central observation is that saddle points of the signed distance function (SDF) reveal where topology can change deliberately. The SDF is parameterized as a cubic trivariate tensor-product B-spline field $9N$8, critical points are located by subdivision-based root finding coupled with Newton’s method, and saddle points are classified through the Hessian eigenvalue signature. The topological interpretation is explicit: 1-saddle points correspond to unstable regions where one can break apart an existing connection or join two disconnected ends, while 2-saddle points correspond to tunnel-like or hole-like regions where one can fill a hole or open a hole. Users select a saddle point, and the system adds a compactly supported B-spline deformer aligned with the three Hessian eigenvectors. Combined with ray marching, the system operates at 30 FPS; the paper also reports around 30 FPS on a GTX 1060 and around 60 FPS on an RTX 3060. In a TSSR-like sense, this is topology sculpting through field manipulation, followed by localized geometric refinement.

A different but conceptually related development is SampEuler, a topological shape transform for thymus structures (Yang et al., 21 Feb 2026). SampEuler is not a mesh generator or an editor; it is a topology-aware shape descriptor derived from the Euler characteristic transform (ECT). For a geometric simplicial complex $9N$9, the ECT computes directional Euler characteristic curves, and SampEuler replaces the direction-indexed family by the empirical measure

$B=V+k$ 0

where $B=V+k$ 1 are sampled i.i.d. from the ECT pushforward measure (Yang et al., 21 Feb 2026). The paper proves convergence of SampEuler to the ECT pushforward measure in Wasserstein distance, establishes stability of the pushforward measure, and develops a vectorization and visualization framework that makes the descriptor interpretable. On the MPEG-7 10-class subset, the reported accuracies are 86.23% for ECT, 51.02% for DETECT, 91.67% for SampEuler, and 94.39% for vectorized SampEuler. On the whole MPEG-7 dataset, the reported accuracies are 71.3% for ECT, 36.4% for DETECT, 91.4% for SampEuler, and 86.8% for vectorized SampEuler. On thymus age classification, SampEuler is best or near-best among the tested methods, while the vectorized representation enables SHAP-based localization of the filtration intervals and Euler-characteristic patterns that drive classification (Yang et al., 21 Feb 2026).

Taken together, these adjacent lines of work indicate that the topology-sculptor/shape-refiner idea is not confined to one representation. In direct mesh generation it separates local topology recovery from global shape recovery; in density-based design it separates geometry extraction from shape optimization; in implicit editing it separates saddle-guided topology change from local surface deformation; and in topological data analysis it separates topology-aware transformation from interpretable vectorization. The shared theme is not a single algorithmic primitive, but a structured allocation of topology-sensitive and geometry-sensitive operations.