Arbitrary-scale light field spatial super-resolution reconstructs high-resolution views at any requested scale from low-resolution multi-view observations. Gaussian primitives enable continuous reconstruction, but generating separate primitive sets for each view redundantly encodes shared scene structure across views and causes the primitive count grow linearly with the number of views. To avoid such redundancy, we propose MacroGS, a feed-forward decoder that jointly predicts a single shared Gaussian set per macro pixel rather than generating separate sets for each view. Disparity-dependent replacement and angle-conditioned appearance allow each set to be reused across all output views, keeping the predicted primitive count independent of the number of views at each requested scale. The decoder iteratively predicts per-primitive disparities, retrieves the corresponding multi-view features, and updates the primitive states. Gaussian splatting and residual refinement then reconstruct the output light field. Despite being learned without geometric supervision, the predicted disparities strongly correlate with underlying scene geometry. Experiments across five benchmarks demonstrate that MacroGS integrates effectively with diverse light field encoders, improving reconstruction quality at fixed scales while achieving leading performance across most evaluated arbitrary-scale settings. The decoder also achieves speedups of approximately 2-10× over state-of-the-art implicit light field decoders and 10-18× over the view-wise Gaussian baseline.
Figures & tables
Figure 1: Gaussian representations of a 2×2 light field at one requested scale; colours mark the K primitives. Only the primitives generated at the red pixel are drawn; every LR pixel generates its own set. (a) Applying an image Gaussian decoder [ 5 , 42 , 24 ] separately to each SAI gives UVKhw primitives ( K per pixel in each view) with separately predicted geometric parameters. (b) MacroGS jointly predicts K primitives from the angular observations of each macro pixel and shares them across all SAIs, giving Khw primitives. Each primitive’s disparity controls its displacement across views.
Figure 2: The framework of MacroGS for arbitrary-scale LFSSR. From the encoder features F , MacroGS infers one set G of K Gaussian primitives per macro pixel, shared by all views, through (a) initialization from all views and (b) R rounds of iterative cross-view reading. (c) View-conditioned rendering splats the set into every view at scale s , and a view-wise convolutional branch R adds a residual correction.
Method
EPFL [ 43 ]
HCInew [ 20 ]
HCIold [ 52 ]
INRIA [ 30 ]
STFgantry [ 46 ]
Average
×2
×4
×2
×4
×2
×4
×2
×4
×2
×4
×2
×4
LF-InterNet [ 47 ]
34.11
28.81
37.17
30.96
44.57
37.15
35.83
30.78
38.44
30.37
38.02
31.61
LF-InterNet + MacroGS
34.47 0.36
29.36 0.55
37.68 0.51
31.62 0.66
44.85 0.28
37.75 0.60
36.15 0.32
31.27 0.49
39.64 1.20
31.56 1.19
38.56 0.54
32.31 0.70
LF-DFnet [ 48 ]
34.40
28.85
37.60
31.29
44.36
37.37
36.34
30.85
39.94
31.26
38.53
31.92
LF-DFnet + MacroGS
34.71 0.31
29.53 0.68
38.03 0.43
31.74 0.45
44.90 0.54
37.96 0.59
36.65 0.31
31.68 0.83
40.61 0.67
31.92 0.66
38.98 0.45
32.57 0.65
LF-IINet [ 38 ]
34.74
29.05
37.77
31.33
44.85
37.62
36.56
31.04
39.89
31.26
38.76
32.06
Table 1: Fixed-scale comparison in PSNR (dB). Each encoder is evaluated with its original upsampler (baseline) and with MacroGS; red subscripts indicate PSNR increases. Baselines are trained for each scale, while MacroGS is fine-tuned from its arbitrary-scale model for each target scale following SEIIF [ 9 ] .
Method
Params (M)
FLOPs (G) / Time (s) / PSNR (in-scale)
FLOPs (G) / Time (s) / PSNR (out-of-scale)
×2
×3
×4
×6
×8
DistgSSR [ 50 ] (Baseline)
[0.02, 0.07]
0.4 /
0.04 /
38.94
1.0 /
0.04 /
34.70
1.7 /
0.04 /
32.12
–
–
Implicit neural fields
DistgSSR + LIIF [ 6 ]
0.09
35.6 /
0.07 /
38.90
79.9 /
0.11 /
34.81
142.0 /
0.17 /
32.42
319.5 /
0.38 /
29.39
568.0 /
0.68 /
27.47
DistgSSR + LTE [ 32 ]
0.13
22.2 /
0.07 /
38.94
47.5 /
0.11 /
34.88
82.9 /
0.16 /
32.46
184.1 /
0.32 /
29.43
325.8 /
0.57 /
27.53
DistgSSR + DIIF [ 19 ]
0.12
20.4 /
0.06 /
38.92
33.1 /
0.06 /
34.84
47.5 /
0.07 /
32.44
81.4 /
0.11 /
29.41
122.1 /
0.14 /
27.54
Table 2: Arbitrary-scale comparison: decoder FLOPs (G) / decoder time (s) / PSNR (dB) averaged over the five benchmarks; parameters count the decoder. All times are for a 5×5×32×32 input on a Tesla V100. Arbitrary-scale methods use one model per encoder for all factors; baselines use scale-specific models. Best and second-best PSNR in each block are in bold and underlined. OOM: out of memory [ 9 ] .
Figure 3: Error maps of GSASR, SEIIF and MacroGS from ×2 to ×8 , using (a) DistgSSR and (b) EPIT encoders.
Figure 4: Arbitrary-scale results of GSASR, SEIIF and MacroGS at ×1.9 , ×3.1 , ×4.3 and ×16 , using (a) DistgSSR and (b) EPIT encoders.
Figure 7
Scale
Disparity
EPFL
HCInew
HCIold
INRIA
STFgantry
Avg.
×2
Zero
35.16
38.28
45.37
36.83
41.14
39.36
Macro-pixel
35.19
38.39
45.46
36.92
41.52
39.50
Primitive
35.22
38.46
45.48
36.98
41.82
39.60
×4
Zero
30.06
31.87
38.10
32.10
32.64
32.95
Macro-pixel
30.12
31.96
38.16
32.11
32.81
33.03
Primitive
30.09
31.97
38.16
32.13
32.85
33.04
Table 4: PSNR (dB) of disparity designs with RASLF at ×2 , ×4 and ×8 . All variants retain primitive sets shared across views. Macro-pixel and Primitive assign one disparity per macro pixel and per primitive, respectively.
Variant
Params (M)
×2
×4
×8
Full MacroGS (res, K=9 , R=3 )
0.37
39.60 /
16.2
33.04 /
39.4
27.85 /
132.2
w/o res
0.29
39.22 /
8.5
32.97 /
8.7
27.88 /
9.6
K=4
0.37
39.42 /
11.6
32.92 /
34.7
27.78 /
127.1
K=16
0.37
39.62 /
22.6
33.05 /
46.0
27.86 /
139.5
R=1
0.21
39.47 /
10.9
32.95 /
34.2
27.67 /
127.0
R=5
0.53
39.64 /
21.4
33.06 /
44.6
27.88 /
137.5
Table 5: Ablation with RASLF. Entries are average reconstruction PSNR (dB) / decoder FLOPs (G); parameters also count the decoder. Each variant changes the indicated component of the full model.