FuguReport

FLAT: Feedforward Latent Triangle Splatting for Geometrically Accurate Scene Generation

Authors Orest Kupyn, Goutam Bhat, Philipp Henzler, Fabian Manhardt, Christian Rupprecht, Federico Tombari
Affiliations Google / University of Oxford / Technical University of Munich
Categories Method / Feedforward Networks / Latent triangle splatting method, Evaluation / Geometric Accuracy Evaluation / Benchmark comparison for geometric precision, Application / Game Engine Compatible Rendering / Triangle soup to opaque rendering conversion
License CC BY 4.0

Abstract Overview

FLAT is a feedforward method that decodes triangle splats directly from frozen camera-conditioned video diffusion latents to generate explorable 3D scenes from a single image in one forward pass. The paper focuses on explicit surface-aligned primitives rather than volumetric Gaussians, aiming for better geometric faithfulness and stronger compatibility with standard graphics pipelines. To make feedforward prediction of flat primitives stable, the method combines a ray-centered triangle parameterization with a product-based differentiable rendering window that improves gradient flow. On RealEstate10K and DL3DV, the triangle variant shows substantially better geometric accuracy than comparable Gaussian-based variants while maintaining competitive novel-view rendering quality, and a lightweight refinement stage converts the output into an opaque, game-engine-ready representation.

Novelty

The paper claims the first direct single-pass decoding of triangle splats from compressed video diffusion latents for feedforward scene generation. Its main technical distinctions are a ray-centered local triangle parameterization with constrained shape and residual orientation, plus a new product window function for differentiable triangle rendering. It also provides a controlled comparison of 3DGS, 2DGS, and triangle splats under the same latent-decoding setup.

Results

In scene-generation evaluation, the triangle representation achieves the strongest geometric quality among the compared FLAT variants, reaching cosine normal similarity 0.853 versus 0.587 for 2DGS and 0.116 for 3DGS, while keeping visual metrics in a competitive range. In mesh conversion, the triangle-based pipeline produces much stronger opaque-mesh rendering quality than extracted meshes from Gaussian representations; on RealEstate10K it reports 21.23 PSNR versus 15.89 for 2DGS TSDF and 14.18 for GS2Mesh from 3DGS. The ablations further indicate that the proposed rotation parameterization and window function are important for stable training.

Key Points

  1. FLAT replaces volumetric Gaussian outputs with directly decoded triangle splats to obtain more explicit and geometrically meaningful surface representations.
  2. The method’s stability depends on two core design choices: a ray-centered triangle parameterization that avoids degenerate or poorly oriented triangles, and a product window function that improves rendering gradients.
  3. A lightweight post-processing step converts the predicted semi-transparent triangle soup into a coherent opaque mesh that is more suitable for game-engine and real-time rendering workflows.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.