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.
Figures & tables
Figure 2: Geometry-to-Gaussian conversion. Top: the encoder Eθ maps surface anchors from the extruded outline to a single 642 feature plane through a point transformer, normalized bilinear splatting and two residual convolutions, and the decoder Dϕ combines bilinearly queried displacement and attribute features with the anchor coordinate ai to predict K neural Gaussians per anchor, which pass an opacity filter and are rasterized for normal and silhouette supervision of both networks. Bottom left: opposite caps project to the same plane coordinate and feature, and the raw depth in ai separates them. Bottom center: the shared five-head decoder predicts offsets, anchor scales, normals, opacity, and covariance, where only opacity and covariance depend on the camera direction and normals are independent of covariance. Bottom right: rendered normals feed deferred PBR shading under user-assigned materials and lighting.
Method
Input
Normal ( ∘ ) ↓
Edge-band ( ∘ ) ↓
Δ normal
IoU ↑
Full test split: 2,657 assets
TexGaussian [ 41 ]
Mesh → points/normals
1.957
13.86
+0.339
0.9595
TriplaneGaussian [ 49 ]
Anchors + 1 view
5.926
18.84
+4.308
0.9810
LGM [ 35 ]
4 views
12.873
32.51
+11.255
0.9576
FFGaussType
Anchors
1.618
8.98
–
0.9851
64-asset fitting subset
Table 1: Conversion fidelity on held-out views. Asset-averaged normal error over foreground pixels and their edge band, Δ normal relative to FFGaussType on the same population, and silhouette IoU. Retrained feed-forward methods receive the listed inputs, while fitting methods optimize against 32 available views per asset for the stated step budgets. Red and orange cells mark the best and second-best values per column within each block. The Δ column is relative and left unmarked.
Method
Forward (ms) ↓
Assets/s ↑
Peak GPU (MiB) ↓
TexGaussian
157.9
6.3
1,675
TriplaneGaussian
28.7
34.9
1,400
LGM
47.9
20.9
2,843
FFGaussType
9.1
109.9
97
Table 2: Forward conversion cost. Time, throughput and peak memory per asset from prepared inputs to Gaussian parameters, batch one, on a single RTX A6000. Red and orange cells mark the best and second-best values.
Figure 3: Qualitative comparison on five test assets. Rows follow increasing edge fraction (left labels), using the fitting subset’s selection rule and one held-out view per asset. The first column shows reference normals, and subsequent columns pair method normals with angular error clipped at 45∘ . RGB encodes world-space normal components (x,y,z) , while the color bar applies only to the error maps.
Figure 4: Seen and unseen font families. Reference and predicted world-space normals for three training (seen) and three test (unseen) font families at held-out views.
Figure 5: Plane prediction and decoder sharing. Held-out normal error on 32 training assets over three seeds, against training cycles and wall-clock time, with per-seed errors over the final cycles. Statistics in (b) give the mean ± s.d. over three seeds at the last checkpoint. A predicts planes with a shared decoder, while B and C fit planes with shared and separate decoders, respectively.
Variant
Latent (kB)
Normal ( ∘ )
IoU
Per-anchor features
–
1.581 ± 0.029
0.983 ± 0.0016
Tri-plane, fewer channels
393.2
1.683 ± 0.045
0.982 ± 0.0018
Tri-plane, lower resolution
394.3
1.876 ± 0.162
0.978 ± 0.0031
Plane, no depth
393.2
1.757 ± 0.070
0.975 ± 0.0024
Plane + γ(z)
393.2
1.675 ± 0.132
0.984 ± 0.0010
Plane + raw z (ours)
393.2
1.668 ± 0.057
0.984 ± 0.0020
Table 3: Latent layout and depth feature. Validation normal error and silhouette IoU, reported as means ± standard deviations over 3 training seeds. Latent size is the raw 16-bit plane payload.
Latent
Budget
Normal ( ∘ ) ↓
Flat ( ∘ ) ↓
Edge-band ( ∘ ) ↓
IoU ↑
Plane (ours)
none
1.70
1.00
8.90
0.984
1 s
1.36
0.65
8.69
0.984
10 s
1.31
0.57
8.98
0.982
60 s
1.35
0.56
9.37
0.981
Per-anchor features
none
1.61
0.98
8.12
0.984
1 s
1.21
0.55
8.00
0.983
Table 4: Refinement on the 64-asset subset. Held-out normal error over all foreground pixels, flat regions and the edge band, and silhouette IoU, before and after 1, 10 and 60 s. Plane and per-anchor features use separately selected step sizes. Red and orange cells mark the best and second-best distinct values.
Figure 6: Coded delivery on 64 test assets. Mean held-out normal error and silhouette IoU versus combined anchor-and-plane size over eight HEVC QPs, for default predictions, 10 s refined planes and the rate-aware variant.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A1 : Camera coverage and paired supervision. (a) The 40-camera rig, with polar angle encoded by radius and azimuth by angle. Red marks the eight held-out (scoring) views, and gray marks the 32 available (pool) views. (b) Projected surface anchors and reference world-space normal maps with silhouette outlines for one glyph at four held-out views.
Method
Input
Adaptation
TexGaussian [ 41 ]
Surface points and normals from a regenerated extrusion mesh, organized in the released octree
Normal head replaces color; text, CLIP and material heads removed. Leaf-center placement and identity rotation; our objective and rasterizer; exponential-moving-average weights (rate 0.999) used for scoring.
TriplaneGaussian [ 49 ]
Our anchors and one rendered normal view
Generative point stage bypassed. We implement a training loop for the second-stage regressor, applying its loss family to normals. Geometry and normals are rotated back to world coordinates for scoring.
LGM [ 35 ]
Four rendered normal views at ring azimuths
Diffusion front end omitted; four-view regressor retrained for normal output with the released optimizer settings and learning rate 4×10−4 .
Appendix
Table B1 : Feed-forward baseline adaptations. Inputs retain each converter’s geometry or image conditioning, and rendered input views come from the 32 available views. All three converters are retrained with normal targets.
Method
Budget
Normal ( ∘ ) ↓
Flat ( ∘ ) ↓
Edge-band ( ∘ ) ↓
IoU ↑
FFGaussType (ours)
one pass
1.701
0.997
8.90
0.9840
3DGS fitting
500
5.185
3.530
21.95
0.8357
1k
4.400
2.841
20.18
0.8700
2k
2.123
0.887
14.74
0.9405
4k
1.931
0.731
14.18
0.9577
8k
1.849
0.701
13.61
0.9681
Appendix
Table B2 : Per-asset fitting across seven budgets. Normal, flat-region and edge-band errors and silhouette IoU on eight held-out views of the same 64 test assets. The one-pass reference and fitting results each use one seed. Budgets are optimization steps. Red and orange cells mark the best and second-best values across budgets within each fitting method.
Method
Setting
Storage
Normal ( ∘ ) ↓
Flat ( ∘ ) ↓
Edge-band ( ∘ ) ↓
IoU ↑
FFGaussType (ours)
one pass
401 kB
1.701
0.997
8.90
0.9840
Mesh2Splat
64
339 kB
4.888
2.222
32.43
0.9372
128
1.36 MB
3.126
1.348
21.41
0.9617
256
5.43 MB
2.307
0.602
19.73
0.9590
520 (default)
22.5 MB
2.371
0.544
20.89
0.9535
GaMeS
unsubdivided, 2, 32k
15.2 kB
15.472
13.605
35.12
0.9437
Appendix
Table B3 : Gaussians attached to the extrusion mesh. Median per-asset storage and mean normal, flat-region, and edge-band errors and silhouette IoU on eight held-out views of the 64 test assets. Storage counts half-precision Gaussian parameters without color for the baselines, and our plane and anchors for FFGaussType. Mesh2Splat settings are conversion resolutions, and GaMeS settings give the mesh, Gaussians per face, and optimization steps. Red and orange cells mark the best and second-best values in each error and IoU column.
Head
Input
Output width
Hidden activation
Output transformation
Offset
xidisp
18
Sine
Linear
Anchor scales
xidisp
6
ReLU
Linear
Opacity
[xiatt;di(v)]
6
ReLU
Sigmoid
Covariance
[xiatt;di(v)]
42
ReLU
Sigmoid (scale); unit normalization (rotation)
Normal
xiatt
18
ReLU
Unit normalization
Appendix
Table C1: Decoder heads and output transformations. All heads have one hidden layer of width 96. Transformations apply after the final linear layer. The covariance head produces three scale factors and four quaternion components per candidate Gaussian.
Figure D1 : Architecture controls on validation examples. Five assets ordered by edge fraction, shown at one held-out view using one training seed. The first two rows show projected anchors and reference normals, and each subsequent cell pairs rendered normals (left) with angular error (right), clipped at 45∘ . The tri-plane rows shown here omit depth conditioning, distinct from the depth-conditioned tri-plane controls in Tab. 3 .
Variant
Normal ( ∘ ) ↓
Flat ( ∘ ) ↓
Edge-band ( ∘ ) ↓
IoU ↑
Reference
1.675
0.930
8.92
0.9842
1282 plane
1.631
0.928
8.49
0.9841
60k steps
1.630
0.885
8.88
0.9853
Appendix
Table D1 : Resolution and training length. Means over three seeds on the validation split. The reference uses a 642 plane, γ(z) , and 30k steps. Red and orange cells mark the best and second-best distinct values per metric.
Variant
Normal ( ∘ ) ↓
Flat ( ∘ ) ↓
Edge-band ( ∘ ) ↓
IoU ↑
Reference
1.712
0.960
9.03
0.9845
Octree-CNN
2.005
1.149
10.27
0.9838
No SSIM term
1.908
1.147
9.31
0.9857
No scale regularizer
1.579
0.869
8.48
0.9864
Appendix
Table D2 : Encoder and loss controls. Single-seed results on the same validation split. The reference uses the point transformer and full objective, and each control changes the named component. Red and orange cells mark the best and second-best values per metric.
Default
Refined (10 s)
Rate-aware
QP
Total (kB)
Normal ( ∘ ) ↓
IoU ↑
Total (kB)
Normal ( ∘ ) ↓
IoU ↑
Total (kB)
Normal ( ∘ ) ↓
IoU ↑
Uncoded
–
1.701
0.9840
–
1.313
0.9825
–
1.635
0.9854
7
52.0
1.702
0.9840
54.7
1.316
0.9824
67.8
1.635
0.9854
12
42.6
1.706
0.9839
45.3
1.321
0.9824
57.8
1.636
0.9854
17
32.7
1.719
0.9838
35.3
1.342
0.9823
47.5
1.636
0.9854
22
23.9
1.755
0.9835
26.3
1.409
0.9822
38.3
1.642
0.9853
Appendix
Table F1 : HEVC operating points on 64 test assets. Mean total size of the coded plane and anchor files, held-out normal error and silhouette IoU at each QP. The uncoded row evaluates each variant’s plane before quantization. Refined planes are the exported 10 s states.
Codec
Total (kB) ↓
Normal ( ∘ ) ↓
IoU ↑
PNG
108.6
1.701
0.9840
JPEG XL
72.4
1.701
0.9840
JPEG 2000
95.0
1.701
0.9840
AVIF
92.3
1.701
0.9840
Appendix
Table F2 : Lossless image delivery on 64 test assets. Each codec exactly preserves the ten-bit plane codes, giving identical fidelity. Total storage includes the unchanged anchor file. Red and orange cells mark the smallest and second-smallest totals. PNG stores the codes in a 16-bit container, and JPEG 2000 uses its reversible mode.
Bosch Research Hildesheim, Germany · University of Bonn Bonn, Germany · Lamarr Institute for Machine Learning and Artificial Intelligence Bonn, Germany