FFGaussType Feed-Forward Conversion of Glyph and SVG Geometry into Compact Gaussian Assets

*Work done during an internship at Dolby Laboratories.

Overview

Overview figure. Glyph and SVG icon outlines enter FFGaussType, which converts them in 9.1 ms per asset into Gaussian-format geometry shown as normal-coloured renderings. A scene graph groups a video, the word COFFEE with asset instances F1, F2 of asset A_F and E1, E2 of asset A_E, light sources and icons. Below, the word, a star and a heart are composited into the Coffee Martini scene, with an inset showing a second view of the same setup.
Convert once, reuse anywhere. Glyph and icon outlines become Gaussian assets with surface normals in 9.1 ms each. Every asset is stored once and instanced with its own material, here inside the Coffee Martini scene[8].

Bring text and icons into volumetric videos built on 4D Gaussians. FFGaussType converts an extruded glyph or SVG outline into a Gaussian asset in a single forward pass. Its surface normals let creators restyle material and lighting freely, without converting again.

Abstract

Immersive 3D content allows text and graphical elements, such as subtitles, annotations, and icons, to be placed inside a scene rather than displayed as flat overlays. Placed this way, they respect occlusion and perspective and take on user-assigned materials and lighting while preserving their original shapes. In Gaussian scenes, however, each new glyph or icon requires its own iterative optimization, which limits scalable content creation. We present FFGaussType, which converts extruded glyph and SVG outlines into Gaussian assets in a single forward pass. A point encoder maps anchors sampled on the known surface to a single feature plane, from which a decoder shared by all assets predicts neural Gaussians carrying surface normals. This separates geometric conversion from material and lighting choices, and the feature plane supports image and video coding for compact delivery. We construct a paired dataset of 27,221 glyph and SVG icon assets with anchors and calibrated normal and silhouette renderings, holding out font families for evaluation. Conversion from prepared anchors takes 9.1 ms per asset, at a fraction of the compute and memory of prior baselines. On unseen font families and test icons, FFGaussType achieves 1.62° normal error and 0.985 silhouette IoU, the best among the compared feed-forward converters, matching the fidelity that per-asset fitting needs thousands of steps to reach. The code and dataset will be released after publication.

In a scene

Converted letters and icons, animated inside a Gaussian video scene.

Fixed camera
Moving camera

Method

Pipeline diagram. An extruded F is sampled into surface anchors; the encoder produces a single 64 by 64 feature plane with 48 channels; a shared five-head decoder queries the plane with the anchor coordinates and camera direction and outputs neural Gaussians per anchor, which are rasterized into normals and alpha for supervision. Rendered normals feed deferred PBR shading under user-assigned material, light and environment map.
Pipeline. Surface anchors are encoded into one feature plane; a shared decoder turns it into neural Gaussians with normals, which are shaded with any material and lighting.
  1. 1Encode

    A point transformer maps surface anchors to a single 64×64 feature plane. The anchors and this plane are the whole asset.

  2. 2Decode

    A small decoder shared by all assets reads the plane and predicts Gaussians carrying surface normals.

  3. 3Shade

    Creators pick materials and lights at render time and can change them whenever they like; the geometry stays as converted.

Results

FFGaussType is the most accurate of the compared feed-forward converters and the cheapest to run, at 9.1 ms and 97 MiB per asset. Ten seconds of refinement surpasses every per-asset fitting baseline in normal accuracy.

Normal error and silhouette IoU on held-out views. Best and second best per block.
Method Input Normal (°) ↓ Edge band (°) ↓ Δ normal IoU ↑
Full test split: 2,657 assets
TexGaussian[1]Mesh → points/normals1.95713.86+0.3390.9595
TriplaneGaussian[2]Anchors + 1 view5.92618.84+4.3080.9810
LGM[3]4 views12.87332.51+11.2550.9576
FFGaussTypeAnchors1.6188.98–0.9851
64-asset fitting subset
3DGS[4], 500 stepsAnchors + available views5.18521.95+3.4840.8357
3DGS, 2kAnchors + available views2.12314.74+0.4220.9405
3DGS, 32kAnchors + available views2.09913.97+0.3980.9811
Scaffold-GS[5], 2kAnchors + available views2.73316.55+1.0320.9400
Scaffold-GS, 32kAnchors + available views1.51410.95−0.1870.9863
Plane-and-decoder fitting, 32kAnchors + available views1.89713.35+0.1960.9828
Mesh2Splat[6], defaultMesh2.37120.89+0.6700.9535
GaMeS[7], 32kMesh + available views2.76415.71+1.0630.9700
FFGaussTypeAnchors1.7018.90–0.9840
FFGaussType + 10 s refinementAnchors + available views1.3138.98−0.3880.9825
Qualitative comparison grid. Five test assets, ordered by edge fraction from 0.06 to 0.12, each shown as a reference normal map followed by normal and angular-error maps for FFGaussType, TexGaussian, 3DGS fitting at 500 and 32k steps, and Scaffold-GS fitting at 32k steps. Error maps use a colour scale from 0 to 45 degrees.
Qualitative comparison. Predicted normals and angular error (0–45°) on five test assets. Rows are sorted by edge fraction, the share of pixels next to sharp creases; generally speaking, a higher edge fraction means a harder shape.
Reference and FFGaussType normal renderings for six glyphs: a 9, an R and an infinity sign from font families seen in training (ProzaLibre, Mirza, Batang), and an F, a U and an A from held-out families (SairaSemiCondensed, NotoSansChakma, NATS). Predictions match the references closely in all six.
Across font families. Reference and predicted normals for glyphs from font families seen in training and from families held out entirely.

BibTeX

@misc{feng2026ffgausstype,
  title  = {{FFGaussType}: Feed-Forward Conversion of Glyph and {SVG}
            Geometry into Compact {Gaussian} Assets},
  author = {Feng, Haoan and Wang, Tianfu and Guan, Peimeng and
            Lee, Dae Yeol and Su, Guan-ming},
  year   = {2026}
}