PolyFlow: Continuous Topology Embedding Flow Matching for Artist-style Mesh Generation

Instead of building 3D meshes one piece at a time like a slow, sequential sequence, PolyFlow builds the whole shape at once using a fast mathematical process. It turns the tricky, discrete connections between points into smooth numbers that a computer can easily refine. This allows for fast and precise 3D model creation.

The bottleneck of sequential decoding

In 3D content creation, generating polygonal meshes—the standard representation for games and film—traditionally relies on autoregressive (AR) Transformers. These models treat a mesh like a sentence. They predict one vertex or face at a time in a strict serial order. This approach can produce high-quality, "artist-like" topology. However, it suffers from a massive computational penalty. Because each token depends on the previous ones, the process is inherently slow. It often takes tens of seconds or several minutes to produce a single asset.

The fundamental problem is that mesh connectivity is discrete. An edge either exists or it does not. This binary nature makes it difficult to use continuous generative frameworks like flow matching (a method that learns to transform noise into data via velocity fields). These frameworks thrive on adding and removing smooth noise. While continuous models excel at synthesizing images or point clouds in parallel, they struggle to "denoise" a discrete connection. Consequently, the industry has faced a trade-off. One could use AR models for high-quality topology at the cost of extreme latency. Alternatively, one could use faster continuous models that produce "triangle soups"—unstructured collections of faces that lack purposeful connectivity.

Lifting topology into continuous space

The authors of PolyFlow resolve this incompatibility by "lifting" discrete connectivity into a continuous latent space. The architecture functions in two distinct stages, as outlined in .

First, the researchers develop a topology embedder. This is a lightweight neural network. It takes ground-truth vertex positions and normals as input. It maps each vertex to a $d$-dimensional continuous embedding. The critical innovation here is the use of a spacetime distance metric. Instead of using standard Euclidean distance, the embedder is trained so that the original discrete adjacency matrix can be recovered. This is done by looking at the "distances" between these embeddings. Specifically, an edge is predicted to exist between two vertices if their spacetime distance falls below a certain threshold $\tau$. This effectively turns a binary, discrete property into a continuous coordinate.

Second, the model employs a Transformer-based flow-matching framework. Rather than predicting tokens sequentially, PolyFlow treats the entire mesh as a single joint state $z = [p, n, e]$. Here, $p$ represents positions, $n$ represents normals, and $e$ represents the continuous topology embeddings. During training, the model learns to map Gaussian noise to this unified state. During inference, the user specifies a target vertex count $\hat{V}$. The model denoises all $\hat{V}$ vertices simultaneously in parallel using an Ordinary Differential Equation (ODE) solver. As seen in, the model progressively resolves the mesh. Early steps establish the global silhouette. Later steps refine local surface details and stabilize the connectivity.

Parallelism delivers significant speedups

The primary evidence for PolyFlow's utility lies in its massive reduction in inference latency. The authors report that PolyFlow achieves a speedup of "tens of times" over autoregressive baselines like BPT. At a scale of 4,000 vertices, PolyFlow completes generation in 5.88 seconds on a single NVIDIA A100. In contrast, the BPT baseline is estimated to take approximately 554.5 seconds. This represents a jump from nearly nine minutes down to just a few seconds. Even compared to FastMesh, a more efficient two-stage pipeline, PolyFlow is roughly 6.2$\times$ faster.

Beyond speed, the paper demonstrates superior geometric fidelity. On the Toys4K benchmark, the authors measure Chamfer Distance (CD) and Hausdorff Distance (HD). These metrics quantify how closely a generated surface matches the ground truth. PolyFlow reports a CD of 0.008 and an HD of 0.021. This outperforms the strongest AR baseline (BPT) by 43% in CD and 40% in HD. The authors suggest this improvement stems from the fact that parallel flow matching avoids "serial error accumulation." In AR models, a single mistake early in the sequence can cause errors in later tokens.

Furthermore, the model offers explicit control over resolution. As demonstrated in, the user can specify a vertex budget.

In the example, this ranges from 250 to 3,000 vertices. The model produces geometrically consistent outputs. It captures finer details as the budget increases. This is a significant departure from AR models. In those models, the sequence length is typically a byproduct of tokenization rather than a controllable parameter.

Constraints of the embedding dimension

While the results are impressive, there are technical nuances regarding the topology embedder. The effectiveness of the continuous representation is sensitive to the dimensionality ($d$) of the embedding space.

The authors perform an ablation study on this dimension. They show that if $d$ is too low (e.g., $d=8$), the embedding space lacks capacity. This leads to "collapsed" geometries with tangled faces . Conversely, increasing the dimension to $d=64$ provides almost no benefit over $d=32$. In fact, the authors observe that higher dimensions can increase the learning difficulty for the downstream flow Transformer. This slightly degrades the end-to-end Hausdorff Distance. Therefore, the choice of $d=32$ is a compromise between reconstruction fidelity and generative stability.

Additionally, the edge recovery process relies on a post-hoc thresholding step. While the authors note that a "normal-guided winding correction" helps ensure consistent face orientation, the process still relies on the precision of the spacetime distance calculation.

The verdict: A new standard for mesh synthesis

If you are working in a production pipeline where 3D asset generation must be integrated into interactive workflows, PolyFlow is a major step forward. It successfully breaks the tension between "high-quality topology" and "fast generation." It does this by rethinking how topology is represented mathematically. By moving from discrete token prediction to continuous state transformation, the authors have applied the efficiency of modern flow models to the domain of polygonal meshes.

I would classify this as a "yes" for practitioners looking to replace slow, autoregressive retopology tools with high-throughput alternatives. The ability to specify exact vertex counts and the dramatic reduction in latency makes it highly actionable. Whether it can scale to extremely high-resolution professional assets remains to be seen. However, the architectural foundation is robust.

Figures from the paper