Spatial-Preserving Encoder
A four-block ResNet-style encoder with GroupNorm and SiLU keeps a 4×4 feature grid before projecting to a 384-dimensional CIFAR latent or 1024-dimensional CelebA latent.
Hypernetworks for Implicit Neural Representations
Built a hypernetwork that converts an image into the weights of a coordinate-based neural representation in one forward pass, then evaluated how SIREN, Fourier-feature, and ReLU INRs behave across reconstruction, adaptation, classification, and super-resolution.
26.06 dB on CIFAR-10
44.21 dB on CelebA
Inner-loop refinement separates the hypernetwork's one-pass amortization quality from the representational ceiling of each coordinate network.
Research Question
An implicit neural representation models an image as a continuous function from coordinates to RGB values. The representation can be queried at arbitrary resolution, but fitting a separate coordinate network for every image normally requires thousands of optimization steps.
This project tests whether a shared hypernetwork can amortize that process by predicting all INR parameters directly from an input image, while preserving the high-frequency detail and continuous-query behavior that make INRs useful.
Model Architecture
Per-Layer Hypernetwork Heads
Independent two-layer generator heads map the image latent into the weight and bias deltas required by every layer of the target coordinate network.
Base + Delta Weights
The model learns a shared base INR and predicts image-conditioned corrections instead of generating raw weights from scratch.
Wℓ = Wbaseℓ + αℓ ΔWℓ(z)
Coordinate-Based Decoder
SIREN, Fourier-feature, and ReLU MLP backbones receive normalized 2D coordinates and emit RGB values, enabling continuous evaluation beyond the training grid.
26.06
CIFAR-10 reconstruction from a single hypernetwork forward pass.
44.21
CelebA SIREN after 50 test-time optimization steps.
62.0%
Classification accuracy from generated INR weights, versus 40.1% from raw pixels.
4.2
Single-image SIREN inference latency after the hypernetwork is trained.
Experimental Results
A SIREN hypernetwork reaches 26.06 dB PSNR, 0.9035 SSIM, and 0.0180 LPIPS across the CIFAR-10 test set.
Generated INR weights are the strongest linear feature, showing that the predicted parameters encode class-relevant structure.
SIREN converges smoothly, Fourier shows modestly higher variance, and ReLU plateaus earliest under spectral bias.
The direct model is limited by amortization, not only by INR capacity
Direct CelebA performance clusters tightly between 22.21 and 23.00 dB because all three architectures receive the same compressed 1024-dimensional image latent. Once the generated weights are refined per image, their representational differences become visible.
44.21 dB
41.06 dB
33.29 dB
Residual SIREN focuses neural capacity on high-frequency correction
A bicubic upsample supplies the low-frequency face structure, while a bounded SIREN residual learns details around eyes, hair, and the mouth. The residual design prevents the network from wasting capacity rebuilding the smooth base image.
I_pred(x, y) = Bicubic(I_LR)(x, y) + s · tanh(f_H(I)(x, y))
33.00 dB
10.06×
The residual INR improves bicubic from 29.94 to 33.00 dB and reduces LPIPS from 0.0956 to 0.0095.
The same coordinate network can be queried at 256×256 even though training supervision ended at 128×128.
Arbitrary-resolution querying works without retraining
A hypernetwork trained with 64×64 inputs and 128×128 targets is evaluated directly on a denser 256×256 coordinate grid. The residual INR preserves its advantage over bicubic, demonstrating that resolution is a query-time choice rather than a fixed output shape.
31.14 dB
0.32 dB
What I Learned
Insight 01
Amortization and representation are separate bottlenecks
A strong coordinate network cannot recover details that the encoder discards. Inner-loop refinement was the clearest way to distinguish compression loss from the true capacity of SIREN, Fourier, and ReLU.
Insight 02
Generated weights are useful representations
The generated INR parameters outperformed raw pixels under a linear classifier, showing that weight space can expose semantic directions that are not obvious in the original image vector.
Insight 03
Inductive bias controls high-frequency behavior
SIREN's learned sinusoidal basis produced the highest refined quality and the most graceful resolution generalization, while ReLU's spectral bias created an earlier performance ceiling.