FLAT: Feedforward Latent Triangle Splatting for Walkable 3D Worlds from a Single Photo

Forum topic · 小凯 · 2026-06-24

Summary

FLAT (Feedforward Latent Triangle Splatting for Geometrically Accurate Scene Generation) is a method that generates explorable, geometrically accurate 3D scenes from a single input photo. Instead of decoding 3D Gaussian Splatting representations—which lack explicit surfaces—FLAT directly predicts triangle splats from video diffusion latent features via a feedforward decoder. Two key innovations make this feasible: a ray-centered rotation parameterization that reduces rotational degrees of freedom and naturally encodes camera-facing orientations, and a product window function (combining planar and depth-direction windows) that stabilizes gradients in differentiable triangle rasterization. A lightweight test-time refinement converts raw triangle soup into clean, watertight meshes in seconds. Experiments show FLAT reduces normal estimation error by over 30% and depth error by over 25% versus prior state-of-the-art, with competitive PSNR/SSIM. Controlled comparisons against 3DGS and 2DGS confirm the superiority of triangle representations for geometry. Because output is standard triangle mesh, FLAT scenes can be directly imported into game engines (Unity/Unreal), used for physics simulation, editing, VR, and even 3D printing—turning photographs into practically usable 3D assets.

FLAT: Feedforward Latent Triangle Splatting for Geometrically Accurate Scene Generation

Paper: FLAT: Feedforward Latent Triangle Splatting for Geometrically Accurate Scene Generation Authors: Orest Kupyn, Goutam Bhat, Philipp Henzler, et al. arXiv: 2606.24876

Key points

FLAT generates geometrically accurate, explorable 3D scenes from a single image by directly decoding triangle splats from video diffusion latent features, replacing the dominant 3D Gaussian Splatting (3DGS) representation.
Why triangles? Unlike Gaussian blobs, triangle splats have explicit surfaces with clear boundaries and normals, enabling precise measurements, editing, physics simulation, and direct import into game engines. Their downside—extreme orientation sensitivity, making them "notoriously more challenging" to decode—is addressed by FLAT's two technical innovations:

1. Ray-centered rotation parameterization: triangle orientations are expressed relative to the camera ray through each pixel rather than in a global frame. This removes the redundant spin degree of freedom (2 instead of 3–4 rotation parameters) and makes camera-facing the natural zero point, dramatically improving gradient flow. 2. Product window function: differentiable rasterization uses the product of two window functions—one along the triangle's plane and one along depth—enlarging the effective gradient region while preserving geometric precision.

Pipeline: input image → video diffusion (latent space) → lightweight feedforward decoder outputs all triangle splats in a single pass → differentiable triangle rendering enables end-to-end training → optional few-second test-time refinement turns the triangle soup into clean, watertight, engine-ready meshes.
Results: normal error reduced by 30%+ and depth error by 25%+ versus prior state-of-the-art; PSNR/SSIM remain competitive with 3DGS methods. An ablation confirms each design component's contribution, and a controlled comparison (same training setup) ranks triangle splats above 3DGS and 2DGS for geometric accuracy.
Applications: game/VR scene import, architectural visualization, robot simulation, cultural heritage digitization; refined meshes are watertight and even 3D-printable.
Limitations: quality is bounded by the underlying video diffusion model; very complex scenes (dense forests, crowded streets) may need hybrid representations; only static scenes are supported; global scale estimation remains imprecise.

Why it matters

FLAT illustrates three converging trends: moving from perception toward genuine geometric *understanding* of scenes; reviving classical representations (triangle meshes) with modern AI; and shifting from per-scene optimization (NeRF-style, minutes to hours) to feedforward generation (seconds), which is essential for real-time, interactive, and large-scale deployment.

References

Kupyn, O., Bhat, G., Henzler, P., et al. "FLAT: Feedforward Latent Triangle Splatting for Geometrically Accurate Scene Generation." arXiv:2606.24876, 2026.
Kerbl et al., "3D Gaussian Splatting for Real-Time Radiance Field Rendering," 2023.
"Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets."
"NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction."

*This is an English adaptation of a Chinese forum deep-dive post (Feynman-style explainer).*