Lafite：面向三维原生纹理生成的生成式潜变量场
Lafite: A Generative Latent Field for 3D Native Texturing
把表面外观编码成稀疏结构化潜变量颜色场、与网格拓扑和 UV 参数化解耦，这一表示可直接复用于三维资产纹理生成、编辑与材质合成；其条件整流流采样范式也可迁移到其他三维场生成任务。
原文依据 learning to generate textures as a 3D generative sparse latent color field
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论直接在三维表面上生成高保真、无缝的纹理——我们将这一过程称为三维原生纹理生成——是一个基础性的开放难题，有望克服传统基于 UV 与多视角投影方法的局限。尽管前景可观，现有原生方法受制于缺乏强大且通用的潜变量表示，严重限制了生成结果的保真度。在本工作中，我们确认这一表示缺口是阻碍进展的核心瓶颈。我们提出 Lafite，一个通过学习将纹理生成为三维生成式稀疏潜变量颜色场来化解该挑战的框架。Lafite 的核心是利用变分自编码器将复杂的表面外观编码到稀疏、结构化的潜空间中，再解码为连续颜色场。这一新颖表示通过有效解耦纹理与网格拓扑及 UV 参数化，实现了前所未有的保真度，即相较最优方法在重建上取得超过 10 dB 的峰值信噪比提升。基于这一更优表示，条件整流流模型能够合成具备最优质量与连贯性的纹理。大量实验表明，Lafite 不仅为原生纹理生成树立了新标准，还支持编辑与材质合成等灵活的下游应用，为下一代三维内容创作流程铺平道路。
英文原文摘要 1452 字符
Generating high-fidelity, seamless textures directly on 3D surfaces, a process we term 3D-native texturing, is a fundamental open challenge, promising to overcome the limitations of traditional UV-based and multi-view projection methods. While promising, existing native approaches are bottlenecked by the lack of a powerful and versatile latent representation, severely limiting the fidelity of generated results. In this work, we identify this representation gap as the central roadblock to progress. We introduce Lafite, a framework that resolves this challenge by learning to generate textures as a 3D generative sparse latent color field. At its core, Lafite leverages a variational autoencoder (VAE) to encode complex surface appearance into a sparse, structured latent space and then decoded into a continuous color field. This novel representation achieves unprecedented fidelity, i.e., yielding a >10 dB PSNR improvement in reconstruction over state-of-the-art methods, by effectively disentangling texture from mesh topology and UV parameterizations. Building on this superior representation, a conditional rectified flow model synthesizes textures with state-of-the-art quality and coherence. Extensive experiments show that Lafite not only sets a new standard for native texturing but also enables flexible downstream applications like editing and material synthesis, paving the way for the next generation of 3D content creation workflows.