FreeMesh: Boosting Mesh Generation with Coordinates Merging

TL;DR

This work addresses the lack of a training-free way to evaluate mesh tokenizers for auto-regressive mesh generation. It introduces Per-Token-Mesh-Entropy ($PTME$) and a plug-in coordinate merging approach, Rearrange & Merge Coordinates ($RMC$), to compress coordinate sequences and improve geometry preservation. Across multiple tokenizers (MeshXL, MeshAnything V2, EdgeRunner), PTME correlates with generation quality, and $RMC$—especially with EdgeRunner—achieves state-of-the-art compression (e.g., $CR$ gains) and better topological fidelity. The proposed framework enables efficient tokenizer design, expands usable training data under fixed context windows, and provides a principled path toward higher-fidelity native mesh generation.

Abstract

The next-coordinate prediction paradigm has emerged as the de facto standard in current auto-regressive mesh generation methods. Despite their effectiveness, there is no efficient measurement for the various tokenizers that serialize meshes into sequences. In this paper, we introduce a new metric Per-Token-Mesh-Entropy (PTME) to evaluate the existing mesh tokenizers theoretically without any training. Building upon PTME, we propose a plug-and-play tokenization technique called coordinate merging. It further improves the compression ratios of existing tokenizers by rearranging and merging the most frequent patterns of coordinates. Through experiments on various tokenization methods like MeshXL, MeshAnything V2, and Edgerunner, we further validate the performance of our method. We hope that the proposed PTME and coordinate merging can enhance the existing mesh tokenizers and guide the further development of native mesh generation.

Figures (7)

• Figure 1: Per-Token-Mesh-Entropy (PTME) Analysis. (a) Visualization demonstrates that our Rearrange & Merge Coordinates (RMC) method significantly enhances geometric detail preservation and better topology. (b) Comparative analysis between baseline Merge Coordinates (MC) and the proposed RMC approach. MC fails to reduce PTME, while our RMC framework effectively minimizes token entropy.
• Figure 2: Comparison of token length distribution between coordinate merging techniques. while the baseline Merge Coordinates (MC) method typically requires 2 coordinates per token representation, the Rearrange & Merging Coordinates (RMC) approach achieves more efficient compression, with most coordinates being represented by a single token.
• Figure 3: Coordinate Merging Pipeline. Given a mesh, we first select a mesh tokenizer to convert the 3D structure into a 1D coordinate sequence. This sequence then undergoes rule-based rearrangement followed by token merging using the Byte Pair Encoding (BPE) algorithm. This approach can significantly reduce the length of the sequence, enabling the poly generation model to generate meshes with more faces.
• Figure 4: Comparison on point-cloud conditional generation. The figure above shows the results of generating meshes conditioned on point clouds sampled from meshes with different face numbers. Using the RMC can significantly improve the quality of the topology and the stability of generation, especially on higher face numbers.
• Figure 5: Compression ratio comparison of tokenizers with coordinate merging techniques. We systematically evaluate baseline Merge Coordinates (MC) and Rearrange & Merge Coordinates (RMC) across varying vocabulary sizes. Both methods exhibit decreasing compression ratios with expanding vocabulary, while RMC demonstrates a steeper reduction gradient than MC.