arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

FFGaussType Rapidly Turns Text and Icon Geometry Into Compact Neural Gaussian Assets | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Computer VisionEnglishPublished

FFGaussType Rapidly Turns Text and Icon Geometry Into Compact Neural Gaussian Assets

November 22, 2026arXiv: 2611.26431v1

FFGaussType is a new pipeline that converts glyphs and SVG outlines into neural Gaussian assets in a single forward pass, avoiding the thousands of steps normally needed for per‑asset fitting. The method wraps the geometry in a point encoder that maps anchors sampled on the surface to a single 64×64 feature plane. A shared decoder then predicts Gaussian parameters with surface normals for each camera. During training, the loss combines cosine alignment, structural similarity, silhouette L1, and regularizers on Gaussian extents and opacity.

When tested on a held‑out set of 2,657 assets, FFGaussType achieved a mean normal error of 1.618° and a silhouette IoU of 0.9851, outperforming other feed‑forward converters. In a forward‑conversion benchmark on an RTX A6000, the method required only 9.1 ms per asset and 97 MiB of peak memory, compared with 157.9 ms/1,675 MiB for TexGaussian, 28.7 ms/1,400 MiB for TriplaneGaussian, and 47.9 ms/2,843 MiB for LGM. One‑pass conversion also beat 3DGS at all optimization budgets in a 64‑asset subset, though it shows higher edge‑band error on sharp rim features.

Limitations include poorer fidelity on sharp rim regions and dependency on a fixed extrusion depth of ±0.125 for all glyphs. Because materials and lighting are not encoded, shading is limited to homogeneous appearance at rendering time.

These results show that FFGaussType can produce geometry‑aware Gaussian assets quickly and at low cost, enabling scalable content creation for immersive 3D scenes, while acknowledging its current constraints.

Source paper

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

Computer Visioncs.GR
View on arXiv