TL;DR
SeamGPT is a novel auto-regressive transformer model that predicts surface cutting seams for meshes, improving the coherence and quality of UV unwrapping and mesh segmentation tasks.
Contribution
It formulates surface cutting as a next token prediction problem using a GPT-style model, enabling more semantically coherent mesh decompositions.
Findings
Achieves state-of-the-art results on UV unwrapping benchmarks.
Improves mesh segmentation with cleaner boundary predictions.
Handles both manifold and non-manifold meshes effectively.
Abstract
Surface cutting is a fundamental task in computer graphics, with applications in UV parameterization, texture mapping, and mesh decomposition. However, existing methods often produce technically valid but overly fragmented atlases that lack semantic coherence. We introduce SeamGPT, an auto-regressive model that generates cutting seams by mimicking professional workflows. Our key technical innovation lies in formulating surface cutting as a next token prediction task: sample point clouds on mesh vertices and edges, encode them as shape conditions, and employ a GPT-style transformer to sequentially predict seam segments with quantized 3D coordinates. Our approach achieves exceptional performance on UV unwrapping benchmarks containing both manifold and non-manifold meshes, including artist-created, and 3D-scanned models. In addition, it enhances existing 3D segmentation tools by providing…
Peer Reviews
Decision·Submitted to ICLR 2026
The idea of framing surface cutting as an auto-regressive sequence prediction task is novel and well-motivated. The integration with PartField yields particularly clean and semantically coherent part boundaries, leading to visually impressive segmentation outcomes. The approach also demonstrates solid generalization across datasets and diverse mesh types.
The method section would benefit from additional toy visualizations to clarify the intuition behind the sequence representation and the quantization/tokenization strategy. Some architectural details (such as hierarchy levels and quantization schemes) could be illustrated more intuitively to help readers grasp the overall process.
1. The paper proposes a new paradigm for surface cutting, which formulates surface cutting as a next token prediction task. The idea is novel, and the results are pretty good. 2. Surface cutting is a very important task in 3D understanding. It essentially finds the best way to geometrically segment a 3D surface into parts (with different criteria). With the part information, it potentially boosts a variety of downstream tasks, such as semantic segmentation (as demonstrated), texture editing, ren
1. As the method is trained purely supervised by ground truth cuttings, the quality of the ground truth cuttings matters a lot, and the model might be sensitive to the poor samples. As the authors mentioned, a rigorous filtering process was applied to clean the data. Thus, scaling up the dataset may be laborious. 2. Some details about the paper are not clearly described, which I will mention in the question section.
- Overall, it's a good idea and a low hanging fruit to approach seam cutting through a GPT-like architecture. - I believe that the results are accurate and good quality can be achieved this way. - The paper clearly describes the approach and except some doubts on implementation (see below) the work seems to be reproducible.
- Abstract claims exceptional performance. This is not validated by the experiments. Please tone down. - Missing baselines (See Q1). - UV texturing can introduce bad seams. There are some new works that discuss and alleviate this problem. See: Foti, S., Zafeiriou, S., & Birdal, T. Uv-free texture generation with denoising and geodesic heat diffusion. NeurIPS 2024. More on this in Q3 below. - Ordering (for example yzx) is rotation dependent and there seems to be no treatment of this. See Q4. -
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
