Implicit neural representations (INRs) have emerged as powerful tools for encoding signals, yet dominant MLP-based designs often suffer from slow convergence, and how their accuracy depends on their size is known only empirically. We introduce FUTON (Fourier Tensor Network), which models signals as generalized Fourier series whose coefficients are parameterized by a low-rank tensor decomposition, read out by a linear or MLP head. FUTON combines complementary inductive biases: Fourier bases capture smoothness and periodicity, while the low-rank parameterization enforces low-dimensional spectral structure. Beyond universal approximation, we show that with a linear head, the best FUTON's error splits exactly into the truncation of the signal's series and the best low-rank error of its coefficients, which we bracket from the samples before training, with a FUTON that attains the upper bound. On image fitting, occupancy volumes, and novel view synthesis, against twelve INRs of matched size, FUTON achieves the best average quality, reaches the final quality of each task's strongest baseline 4-34x sooner than that baseline, infers fastest, and trains in the least memory.
Figures & tables
Figure 1: FUTON’s forward pass for one pixel of an image ( C=2 , D=3 ). (1) A fixed basis, here cosines, expands each coordinate xc into Kc functions, at spectral resolution K=(K1,K2) . (2) The combiner’s factor matrices U(c) project each expansion onto R components, multiplied across axes into g(x) : a CP decomposition of the coefficient tensor. (3) A learned head, here an MLP, maps g(x) to the signal. Arrays hold exact values for random factors, at sizes far below practice.
Figure 2: FUTON’s coefficient tensor. (a) A Kodak image and its coefficient tensor in the cosine basis at Kc=32 , visualized as RGB images with the lowest frequencies at the top left: (b) in full, 3,072 coefficients (23.79 dB), and (c) at rank R=8 , the certified FUTON of Prop. 1 , built without training, 536 parameters (19.89 dB). (d, e) The norm of their difference per frequency, and per pixel between the images they represent.
Figure 3: The error bound and the bandwidth. (a) PSNR relative to Adam’s FUTON per rank at κ=1/2 , averaged over the 24 Kodak images and the five volumes, ± one standard error: the ceiling and the certified FUTON bracket the best FUTON. (b) Mean PSNR on Kodak against the relative bandwidth κ at 197k parameters, the rank adjusted to keep the size, the default circled.
Figure 4: Training speed. Mean quality during training against median training time, with one standard error, for the featured models; radiance fields on validation views. Figures 13 and 15 plot every model’s final quality against its costs.
Quality
Training
Inference
Model
PSNR
SSIM
LPIPS
Time
Memory
Throughput
Memory
(dB) ↑
↑
↓
(s) ↓
(GB) ↓
(MPix/s) ↑
(MB) ↓
RFF [ 18 ]
31.28
0.846
0.271
29.8
2.08
30.7
1208
MFN [ 20 ]
36.72
0.938
0.129
57.3
7.67
26.8
3224
SIREN [ 6 ]
33.57
0.906
0.192
35.7
3.25
34.4
805
FreSh (SIREN) [ 21 ]
34.88
0.922
0.150
39.3
3.25
34.5
805
Table 1: Images : quality averaged over the 24 Kodak images, and median training time, inference throughput, and peak GPU memory, at about 197k parameters. Best bold , second underlined , among the models shown; Tab. 11 adds standard errors and the other three baselines.
IoU (%) ↑
Training
Inference
Model
Armadillo
Dragon
Happy
Lucy
Thai
Mean
Time
Memory
Throughput
Memory
Buddha
Statue
(s) ↓
(GB) ↓
(MVox/s) ↑
(GB) ↓
RFF
99.71
99.65
99.59
99.78
99.28
99.60
46.5
4.15
48.9
3.33
MFN
99.23
97.47
97.50
99.10
98.74
98.41
108.5
15.07
38.9
5.49
SIREN
98.78
99.18
99.41
98.53
97.94
98.77
70.0
6.48
51.0
2.26
FreSh (SIREN)
99.34
99.67
99.65
99.15
98.83
99.33
263.6
6.48
51.3
2.26
Table 2: Occupancy volumes : IoU on each shape and the mean, and median costs, marked as in Tab. 1 , at about 131k parameters. Table 12 adds standard errors and the other baselines.
Figure 5: Image fitting on kodim19, with PSNR and two regions magnified; Fig. 14 shows more.
Figure 6: Occupancy volumes on Lucy, with IoU and two regions magnified; Fig. 16 shows more.
PSNR (dB) ↑
Training
Inference
Model
Chair
Drums
Ficus
Hotdog
Lego
Materials
Mic
Ship
Mean
Time
Mem.
Speed
Mem.
(s) ↓
(GB) ↓
(FPS) ↑
(GB) ↓
RFF
33.10
22.91
27.47
32.76
28.48
25.62
33.37
20.88
28.07
179.3
2.91
38.7
2.12
MFN
32.94
23.77
27.28
32.56
28.52
25.60
33.04
22.50
28.28
275.6
9.97
35.5
3.61
SIREN
32.54
24.58
27.61
32.40
29.03
26.53
33.25
22.29
28.53
215.5
5.08
38.4
2.14
FreSh (SIREN)
32.38
24.19
27.60
32.81
29.62
25.18
32.58
22.32
28.34
217.2
5.03
39.4
2.13
Table 3: Radiance fields : PSNR on each Blender scene’s test views and the mean, and median costs, marked as in Tab. 1 , speed in frames per second, at about 74k parameters. Table 13 adds SSIM, LPIPS, standard errors, and the other baselines.
Figure 7: Novel view synthesis on Lego, with mean test-view PSNR and two regions magnified; Fig. 17 shows more.
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Tensor diagram notation. Graphical representation of tensors with different orders: (a) scalar (0th-order), (b) vector (1st-order), (c) matrix (2nd-order), and (d) third-order tensor. In tensor diagrams, tensors are represented as nodes (shapes) with edges (lines) denoting indices. Dangling edges represent open (uncontracted) indices. (e) Complex tensor network from ( 17 ) with multiple contracted indices. Edges connecting different nodes represent shared indices that are summed over. (f) Trace of matrix product tr(ABCDEF) represented as a ring diagram, where the cyclic property becomes visually apparent, i.e . , any rotation of the ring yields the same value.
Figure 9: Tensor diagrams of basic operations. (a) matrix trace; (b) Hadamard product, used in ( 9 ); (c) mode- 1 product; (d) tensor product, used in ( 5 ) and ( 7 ); (e) matrix-vector multiplication, used in ( 10 ) and ( 8 ); and (f) generalized dot product, used in ( 6 ).
Figure 10: Contraction order, and FUTON’s tensor network (linear head). (a, b) Sequential contraction of a tensor network whose modes are all of size I . In (a), the matrix A and the tensor C are first contracted (mode- m product), then the resulting tensor is contracted with the matrix B , requiring O(I4) operations in total. In (b), the matrices A and B are first contracted (multiplied), then the resulting matrix is contracted with the tensor C , requiring O(I3) operations in total. (c) Tensor diagram of FUTON’s forward pass, where nodes represent tensors, connected edges represent contracted indices (summed over), and dangling edges correspond to open indices. See Sec. B.1 for tensor diagram notation. (d–f) Optimal contraction order achieving linear complexity O(R∑cKc+CR+DR) . Orange boxes indicate tensors being contracted at each step: (d) project each basis vector φ(xc) onto its factor matrix U(c) via matrix-vector multiplication ( 10 ), incurring O(R∑cKc) operations; (e) combine the projected features hc(xc) via Hadamard product ( 9 ), incurring O(CR) operations; and (f) map the resulting vector g(x) to the output via matrix-vector multiplication with V ( 8 ), incurring O(DR) operations. This contrasts with the naive approach of directly computing Φ(x)∙W , which requires O(D∏cKc) operations and scales exponentially with the input dimension C .
Figure 11: Tensor diagram of CP decomposition.
Axis-aligned
Oblique, 30∘
Diagonal, 45∘
Width δ
10−3
10−4
10−5
10−3
10−4
10−5
10−3
10−4
10−5
0.2
1
1
1
4
4
5
4
4
6
0.1
1
1
1
5
6
8
6
8
10
0.05
1
1
1
7
10
12
9
13
16
0.02
1
1
1
13
19
25
18
26
35
0.01
1
1
1
21
33
45
30
47
63
Appendix
Table 4: The edge test. Smallest rank R at which an edge’s cosine coefficients ( Kc=256 ) have relative squared error at most 10−3 , 10−4 , and 10−5 , per width δ and direction, exactly from the singular values ( 4096×4096 midpoint quadrature).
Step
Computes
Cost
1
Qc,Σc,Ec from the thin SVD of Φc/Lc
O(LcKc2) per axis
2
Z=S×1Q1⊤⋯×CQC⊤/N
O(ND∑cKc)
3
∥s−sK∥2=∥S∥F2/N−∥Z∥F2 , exact
O(ND)
4
εR2 : the singular values of every matricization of Z , tails beyond each R
one eigendecomposition per split
5
Lem. 3 for every mode ordering, keeping the ordering with the most weight in its R heaviest terms; εR2=∥Z−Y∥F2 for their sum Y=[[U~(1),…,U~(C),V~]]
nested SVDs of Z
6
U^(c)=EcΣc−1U~(c) , V^=V~
O(R∑cKcKc′)
Appendix
Table 5: The bounds from samples. Steps 5 and 6 give the certified FUTON.
Seconds
κ
R
Truncation
Certified
Ceiling
Adam
Bound
Adam
Images
1/8
32
23.65 ± 0.06
22.85 ± 0.06
23.27 ± 0.01
23.13 ± 0.02
0.02
1.7
1/8
64
23.65 ± 0.01
23.45 ± 0.03
23.62 ± 0.01
23.57 ± 0.00
0.02
1.7
1/8
128
23.65 ± 0.00
23.64 ± 0.00
23.65 ± 0.00
23.65 ± 0.00
0.02
1.7
1/8
256
23.65 ± 0.00
23.65 ± 0.00
23.65 ± 0.00
23.65 ± 0.00
0.02
1.7
Appendix
Table 6: Means per setting over the 24 Kodak images and over the five occupancy volumes. PSNR (dB), ± one within-signal standard error: the truncation, the certified FUTON, the ceiling and Adam’s FUTON; seconds on an A100 for the bound at all six ranks of a K and for one Adam run.
Figure 12: Where Adam’s error goes, and what the bound costs. (a) Adam’s error at κ=1/2 , in percent, as Prop. 1 splits it, averaged over the 24 Kodak images and the five occupancy volumes, per rank: the truncation, which K forces; the part of the misfit up to the ceiling, which R forces with certainty; and the rest, which the bounds leave open. (b) How many times faster the bound, certified FUTON included, is than the six Adam runs it stands in for, per relative bandwidth κ , from the seconds per signal averaged over each dataset.
Adam
Dataset
R
Certified
Ceiling
1000 steps
5000 steps
Kodak
280
29.56 ± 0.03
29.84 ± 0.01
29.75 ± 0.00
29.80 ± 0.01
Occupancy
215–310
21.22 ± 0.16
26.94 ± 0.43
22.55 ± 0.10
22.94 ± 0.07
Appendix
Table 7: Training against the bound. Top: mean PSNR (dB) over each dataset, ± one within-signal standard error, at κ=1/2 and the benchmark’s rank, for the certified FUTON and the ceiling against Adam at the benchmark’s 1000 steps and at 5000. Bottom: seconds per signal for the six ranks of a K , from κ=1/8 to 2 , averaged over each dataset, for the bound, certified FUTON included, and the six Adam runs; and how far the certified FUTON stays below Adam’s, in dB, over the signals and ranks.
Reached
Pick vs. oracle
#Params (k)
Target (dB)
Adam
Certified
Possible
Same
Larger
Ceiling
Adam
Certified
Kodak images
22
24/24
24/24
24/24
88%
12%
5.2
5.2
5.2
24
23/24
23/24
23/24
70%
30%
5.2
10.3
10.3
26
22/24
22/24
22/24
77%
23%
20.7
20.7
30.9
28
18/24
18/24
18/24
67%
33%
41.2
61.7
61.8
Appendix
Table 8: The smallest setting for a target PSNR. Reached: the signals of each dataset for which some setting reaches the target by Adam, by the certified FUTON, and by the ceiling (possible). Pick vs. oracle: over the signals both reach, how often the certified pick has the parameter count of the oracle’s, which trains every setting and takes the smallest that Adam reaches, or a larger one; it is never smaller. #Params: the median parameter count of the ceiling’s pick, below which no candidate reaches the target, of the oracle’s, and of the certified pick.
Model
Configuration
#Params (k)
LR
Images
RFF
3×212 ; 256 Fourier features, σ=10
199.7
3⋅10−3
PE-MLP
4×256 ; positional encoding with 10 frequencies
209.2
3⋅10−3
MFN
4×256 ; input scale 1024, α=1.25 , β=0.125
204.3
3⋅10−2
SIREN
4×256 ; ω0=30
198.9
10−3
FreSh (SIREN)
as SIREN, ω0 selected per image
198.9
10−3
Appendix
Table 9: Model configurations per task : architecture, parameters, and learning rate. “ 3×256 ” is three hidden layers of 256 units; ω0 and γ0 are the activations’ frequency and scale, Lc is the signal’s size along axis c , and M the width of FUTON’s MLP head. Radiance-field sizes include the shared color network.
Figure 13: Quality against training time (top) and inference speed (bottom) for every model; gray dots are the other baselines. On the volumes, the axes start at 98% IoU, and models below sit on the axis line; a model past an axis’s range sits at its end as a marker.
Figure 14: Image fitting on kodim01, kodim21, and kodim23, as in Fig. 5 .
Figure 15: Quality against training memory (top) and inference memory (bottom), as in Fig. 13 .
Figure 16: Occupancy volumes on Armadillo, Dragon, Happy Buddha, and Thai Statue, as in Fig. 6 .
Training
Inference
Mean over the scenes
Model
Time
Memory
Speed
Memory
PSNR
SSIM
LPIPS
(s) ↓
(GB) ↓
(FPS) ↑
(GB) ↓
(dB) ↑
↑
↓
RFF
179.3
2.91
38.7
2.12
28.07 ± 0.30
0.928 ± 0.004
0.082 ± 0.008
PE-MLP
201.0
3.00
37.0
2.10
27.94 ± 0.22
0.931 ± 0.003
0.079 ± 0.006
MFN
275.6
9.97
35.5
3.61
28.28 ± 0.16
0.932 ± 0.002
0.078 ± 0.006
SIREN
215.5
5.08
38.4
2.14
28.53 ± 0.16
0.935 ± 0.002
0.106 ± 0.007
Appendix
Table 13: Radiance fields at about 74k parameters: median costs and means over the eight Blender scenes ± one standard error (top); per scene, the PSNR of the baselines Tab. 3 omits and every model’s SSIM and LPIPS (bottom), marked among all models; 200 test views per scene.
Figure 17: Novel view synthesis on Chair, Materials, Hotdog, and Ship, as in Fig. 7 .
Figure 18: Ablations on Kodak , mean PSNR ± one standard error, as in Tab. 14 : (a) relative bandwidth and (b) rank, the defaults circled; (c–e) PSNR during training for four heads, three combiners, and six bases.
Training
Inference
κ
R
Time
Memory
Throughput
Memory
PSNR
SSIM
LPIPS
(s) ↓
(MB) ↓
(MPix/s) ↑
(MB) ↓
(dB) ↑
↑
↓
Relative bandwidth, at about 197k parameters
1/8
885
9.0
442
111.9
219
33.41 ± 0.15
0.914 ± 0.003
0.167 ± 0.004
1/4
515
6.1
392
156.9
201
36.95 ± 0.18
0.948 ± 0.003
0.118 ± 0.004
1/2
280
4.1
359
228.7
201
37.63 ± 0.51
0.955 ± 0.003
0.098 ± 0.003
Appendix
Table 14: Ablations on Kodak : means over the 24 images ± one standard error, and median costs, with the best and second best marked within each block. Top: the relative bandwidth κ at about 197k parameters, the rank adjusted to keep the size, and the rank R at κ=1/2 , from 49.5k to 789k parameters. Bottom: at equal size, MLP heads of width M=64 , 128, and 256 and a linear head, each with the largest rank that keeps the size; CP, a tensor ring (TR, 180k parameters), and Tucker, at about equal size; and six bases.