The model zoo¶
Twelve architectures share one interface: Model(in_features, out_features, ...) maps coordinates (..., in_features) in \([-1, 1]\) to (..., out_features). They are re-implementations under one training loop, at settings taken from the authors' code, and the benchmarks compare them at equal parameter counts.
-
A fixed basis per axis, a low-rank combiner across axes, and a head.
-
MLPs whose prior sits in the first layer or in the activation: RFF, PE-MLP, SIREN, Gauss, WIRE, FINER, MFN.
-
Learnable feature grids and factorizations: Instant-NGP, TensoRF, GA-Planes, F-INR.
FUTON¶
| Model | Class | Idea | Reference |
|---|---|---|---|
| FUTON | FUTON |
A fixed basis per axis, a low-rank CP combiner across axes, and a linear or MLP head | Ashtari et al., arXiv 2026 |
The benchmarks run it with its default basis, the cosine, as FUTON in the results and the configs; the basis ablation compares the sinc and Lanczos bases and three more with it. The FUTON page explains every part.
Coordinate networks¶
MLPs whose prior sits in the first layer or in the activation.
| Model | Class | Idea | Reference |
|---|---|---|---|
| RFF | RFF |
Random Fourier features \([\sin(2\pi \boldsymbol{B}\boldsymbol{x}), \cos(2\pi \boldsymbol{B}\boldsymbol{x})]\), \(\boldsymbol{B} \sim \mathcal{N}(0, \sigma^2)\), then a ReLU MLP | Tancik et al., NeurIPS 2020 |
| PE-MLP | PEMLP |
NeRF's positional encoding \([\boldsymbol{x}, \sin(2^k \boldsymbol{x}), \cos(2^k \boldsymbol{x})]\), then a ReLU MLP | Mildenhall et al., ECCV 2020 |
| SIREN | SIREN |
Sine layers \(\sin(\omega (\boldsymbol{W}\boldsymbol{x} + \boldsymbol{b}))\) with SIREN initialization | Sitzmann et al., NeurIPS 2020 |
| Gauss | Gauss |
Gaussian layers \(\exp(-(s (\boldsymbol{W}\boldsymbol{x} + \boldsymbol{b}))^2)\) | Ramasinghe and Lucey, ECCV 2022 |
| WIRE | WIRE, RealWIRE |
Complex, or real, Gabor wavelet layers \(\cos(\omega z_1) \exp(-(s z_2)^2)\) | Saragadam et al., CVPR 2023 |
| FINER | FINER |
Variable-periodic sines \(\sin(\omega (\lvert z \rvert + 1) z)\) | Liu et al., CVPR 2024 |
| MFN | MFN |
Multiplicative Gabor filter network: each stage multiplies a Gabor filter of the coordinates into the hidden state | Fathony et al., ICLR 2021 |
- FreSh (Kania et al., ICLR 2025) is not an architecture but a choice of SIREN's first-layer frequency:
inro.fresh_selectpicks it per signal, before training, as the candidate whose untrained output has the spectrum nearest the signal's, andinro.SIREN(first_omega=...)takes it, its later layers keepingomega. MLPunderlies most of them: a plain ReLU network by default, or a network of custom layers (layer_class=inro.SineLayer,inro.GaussLayer,inro.FinerLayer,inro.RealGaborLayer) with their own activations.WIREis the complex-valued network of the paper. The experiments useRealWIRE, the real Gabor variant from the same paper and code, which trains about 1.5× faster at equal parameter count and has real weights.
Grid and tensor models¶
Learnable feature grids, read by interpolation and decoded by a ReLU MLP, and a factorization into one network per axis.
| Model | Class | Idea | Reference |
|---|---|---|---|
| Instant-NGP | InstantNGP, HashEncoding |
Multiresolution hash encoding (16 levels, 2 features per level by default) and a bias-free ReLU MLP, mirroring tiny-cuda-nn | Müller et al., SIGGRAPH 2022 |
| TensoRF | TensoRF |
Factorized grids: CP (products of line grids) or VM (line-plane products, 3D) | Chen et al., ECCV 2022 |
| GA-Planes | GAPlanes, MultiVector, FeatureGrid |
Line, plane and volume grids combined by an expression such as "[e1 o e2, e12]" |
Sivgin et al., ICML 2025 |
| F-INR | FINR |
One SIREN per coordinate, its outputs combined per channel by a CP or tensor-train decomposition, with no decoder | Vemuri et al., WACV 2026 |
FeatureGrid is the building block: a learnable grid of any dimension sampled with align_corners=True on \([-1, 1]\). MultiVector combines several with elementwise products, sums and concatenation, and GAPlanes and TensoRF add the decoder.
Shared conventions¶
| Topic | Rule |
|---|---|
| Widths and depths | hidden_features is the hidden width and hidden_layers the number of hidden layers, so hidden_layers=h means h + 1 weight layers (MFN: h Gabor filters, h - 1 hidden linear layers and one output layer). WIRE and RealWIRE divide the nominal width by \(\sqrt{2}\) to keep the parameter count comparable to real networks |
| Output activations | Every model takes output_activation, a callable applied last. The benchmarks use torch.tanh for targets in \([-1, 1]\) and none inside a RadianceField |
| Equal capacity | inro.count_parameters(model, trainable_only=False) is how the configs were matched. The image models hold 197k to 209k parameters, the occupancy models about 132k, and the NeRF density networks about 68k (75k with the shared color network) |
| Learning rates | Each model has its own, the best of {3e-1, 1e-1, 3e-2, 1e-2, 3e-3, 1e-3, 3e-4, 1e-4} on two signals of the task, as each config's header says |