Research map / Architectures and efficiency / RNNs and normalization
ASMR: Activation-Sharing Multi-Resolution Coordinate Networks for Efficient Inference
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 ASMR, a SIREN-based coordinate network that shares activations across multi-resolution coordinate grids with hierarchical modulations, cutting inference MACs to near O(1) in depth while matching or beating SIREN reconstruction quality.
Problem
Implicit neural representations (coordinate networks) have inference cost (MACs) that grows with depth and width, which is costly under hardware-constrained inference; existing low-MAC methods (hybrid grids, KiloNeRF, LoE) trade off parameters or quality.
Approach
Decompose coordinates into multi-resolution grid indices, add position-dependent bias modulations per level, and perform activation sharing so each layer evaluates only on the C_i coarse grid points and upsamples, making per-sample MAC roughly independent of depth.
Key results
- On Kodak, ASMR reaches 33.087 dB PSNR with 1.85K MACs versus SIREN's 32.398 dB at 132.4K MACs, and is over 3x faster in latency on server, desktop and Raspberry Pi.
- On the 8192x8192 Pluto image, ASMR applied to SIREN cuts MACs by about 175x (to 6K) and improves PSNR from 31.66 dB to 33.10 dB; the ASMR-BIG variant gives over 500x MAC reduction.
- On 3D ShapeNet occupancy fitting, ASMR matches SIREN IoU (89.57 vs 89.27) with 0.039M MACs versus 2.100M.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ASMR hierarchical modulation and activation sharing instead of vanilla SIREN | SIREN · Kodak | PSNR (dB): 32.398$\pm$3.331 -> 33.087$\pm$2.704 | Helped |
| Use ASMR hierarchical modulation and activation sharing instead of vanilla SIREN | SIREN · Kodak | MACs (K): 132.4 -> 1.85 | Helped |
| Use ASMR instead of SIREN | SIREN · Pluto 8192×8192 image | PSNR (dB): 31.66dB to 33.10dB | Helped |
| Use ASMR instead of SIREN | SIREN · Pluto 8192×8192 image | MACs: more than 1M MACs -> 6K MACs | Helped |
| Use ASMR instead of SIREN | SIREN · ShapeNet occupancy grids | IoU: 89.27$\pm$6.43 -> 89.57$\pm$6.72 | No clear effect |
| Use ASMR instead of SIREN | SIREN · Cat video | PSNR (dB): 34.641 $\pm$ 1.159 -> 31.506 $\pm$ 1.401 | Hurt |
| Use ASMR instead of SIREN | SIREN · UVG | PSNR(dB): 33.6 $\pm$ 3.6 -> 32.3 $\pm$ 4.2 | Hurt |
| Permute the partition bases from [4,4,4,8] to [4,4,8,4] (ablation) | ASMR · Cameraman | PSNR (dB): 37.755 -> 37.442 | No clear effect |
| Permute the partition bases from [4,4,4,8] to [4,8,4,4] (ablation) | ASMR · Cameraman | PSNR (dB): 37.755 -> 37.776 | No clear effect |
| Permute the partition bases from [4,4,4,8] to [8,4,4,4] (ablation) | ASMR · Cameraman | PSNR (dB): 37.755 -> 37.574 | No clear effect |
| Move the larger partition base to the final level ([4,4,4,8] instead of [4,4,8,4]) (ablation) | ASMR · Cameraman | MACs (K): 4.42 -> 1.34 | Helped |
| Change the partition bases from [4,4,4,8] to [4,4,8,4] (ablation) | ASMR · Cameraman | SSIM: 0.930 -> 0.925 | No clear effect |
Surprising findings
- ASMR performs significantly worse on continuous signals such as signed distance fields, indicating a strong bias toward rasterized data.
- Instant-NGP fails to learn shared structure in the CIFAR-10 meta-learning setting, while ASMR does.
Limitations
- Performs poorly on smooth continuous signals such as SDFs, producing noisy artifacts and losing analytical gradient cleanliness.
- Expressivity still depends on network width, unlike hash-grid methods that scale capacity without widening the MLP.
- Ultra-low MAC gains depend on the chosen partition bases; LoE and KiloNeRF could not be fully reproduced without official code.
Experimental setup
Datasets: Cameraman Kodak LibriSpeech Pluto Cat video UVG ShapeNet CIFAR-10
Benchmarks: Cameraman Kodak LibriSpeech Pluto Cat video UVG ShapeNet CIFAR-10
Metrics: PSNR SSIM MACs parameter count latency IoU
Baselines: SIREN KiloNeRF Level-of-Experts FFN MFN WIRE Instant-NGP
Builds on: SIREN COIN++ KiloNeRF Level-of-Experts Instant-NGP
Closest papers by meaning
- Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals (NeurIPS 2025 · similarity 0.86)
- Representing Sounds as Neural Amplitude Fields: A Benchmark of Coordinate-MLPs and a Fourier Kolmogorov-Arnold Framework (AAAI 2025 · similarity 0.86)
- End-to-End Implicit Neural Representations for Classification (CVPR 2025 · similarity 0.85)
- VI^3NR: Variance Informed Initialization for Implicit Neural Representations (CVPR 2025 · similarity 0.85)
- A sampling theory perspective on activations for implicit neural representations (ICML 2024 · similarity 0.85)
- MIRE: Matched Implicit Neural Representations (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.