Research map / Architectures and efficiency / RNNs and normalization

Towards Croppable Implicit Neural Representations

NeurIPS 2024 · Architectures and efficiency › RNNs and normalization · approach: Sparse dictionary representations · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Local-Global SIRENs (and a Local-Global INCODE variant), an INR architecture that splits the signal into partitions with per-partition local sub-networks plus a shared global sub-network, enabling cropping by deleting local weights without retraining.

Problem

Implicit neural representations are black boxes, so encoded signals are hard to edit post-training; in particular, cropping an encoded signal with a proportional reduction in weights is not supported without retraining or distillation.

Approach

Combine per-partition local SIREN sub-networks (removable for cropping) with a global sub-network whose features are merged into each local sub-network via a per-coordinate concat-plus-linear merge operator, recovering global context lost in pure INR-per-partition encoding without a distillation step.

Key results

• On 25 DIV2K images with 16x16 partitions, Local-Global SIREN reaches 34.13 dB PSNR (SSIM 0.971) vs 33.57 dB for SIREN and 31.73 dB for SIREN-per-Partition.
• On the cat video, Local-Global SIREN with (5,4,4) partitions reaches 30.95 dB and trains in 386 s vs 2354 s for SIREN, with a 3.5% global weight share.
• Local-Global INCODE improves over INCODE on DIV2K, 39.03 dB vs 38.79 dB PSNR (SSIM 0.974 vs 0.963).

Measured effects reported in the paper

Surprising findings

• Local-Global SIREN with partition factors (1,1) outperforms the plain SIREN on the Cameraman image (32.52 dB vs 31.17 dB), which the authors attribute to the two-network merged architecture rather than local learning.
• Distilling a full SIREN into a SIREN-per-Partition via upsampled novel pixels reduced accuracy (31.64 dB vs 31.90 dB without distillation).

Limitations

• Cropping is limited to partition granularity, so it is less flexible than arbitrary weight removal.
• Torch bmm implementation latency is unstable across GPUs.
• Extension requires fine-tuning after adding local sub-networks.

Experimental setup

Datasets: DIV2K Cameraman Bach Cello Suite audio clip Counting audio clip cat video Lucy (Stanford 3D Scanning Repository) Kaggle Lung Nodule Analysis CT image

Benchmarks: DIV2K subset image encoding Cameraman image encoding audio encoding (Bach, Counting) video encoding (cat video) 3D shape encoding (Lucy) image denoising image super-resolution CT reconstruction

Metrics: SSIM PSNR MSE IoU training time

Baselines: SIREN SIREN-per-Partition INCODE INCODE-per-Partition MINER

Builds on: SIREN INCODE KiloNeRF

Closest papers by meaning

• Implicit Representations via Operator Learning (ICML 2024 · similarity 0.88)
• Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals (NeurIPS 2025 · similarity 0.87)
• VI^3NR: Variance Informed Initialization for Implicit Neural Representations (CVPR 2025 · similarity 0.86)
• Implicit Neural Representations and the Algebra of Complex Wavelets (ICLR 2024 · similarity 0.86)
• Learning Transferable Features for Implicit Neural Representations (NeurIPS 2024 · similarity 0.86)
• End-to-End Implicit Neural Representations for Classification (CVPR 2025 · similarity 0.86)