Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Transferable Features for Implicit Neural Representations
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 STRAINER, which trains a shared SIREN encoder across multiple signals with per-signal decoders, then uses the transferred encoder as initialization for fitting new in-domain or out-of-domain signals and inverse problems.
Problem
Implicit neural representations are typically trained on a single signal and their learned features do not transfer, so each new signal must be fit from scratch or from costly meta-learned initializations.
Approach
Share the initial encoder layers of several INRs (SIRENs) across multiple training signals while keeping separate decoder layers per signal; at test time, initialize a new INR with the shared encoder and a random decoder, then optimize as usual.
Key results
- STRAINER-10 reaches 57.80 dB PSNR on CelebA-HQ in-domain fitting vs 44.91 dB for SIREN, about +10 dB gain (Table 1).
- On out-of-domain Kodak images at width 512, STRAINER-10 achieves 44.38 dB PSNR on Parrot vs 40.18 dB for SIREN.
- For denoising, STRAINER-10 uses 76 iterations vs 203 for SIREN at comparable PSNR (26.41 vs 26.75); for fast super-resolution it reaches 31.56 dB in 1102 iterations (about 3x faster) vs SIREN 32.10 dB in 3329 iterations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train the shared encoder on 10 images instead of one (ablation) | STRAINER · CelebA-HQ and AFHQ in-domain image fitting | PSNR: 50.34 ± 2.81 -> 57.80 ± 3.46; 51.27 ± 2.94 -> 58.06 ± 3.75 | Helped |
| Train the shared encoder on 10 images instead of one (ablation) | STRAINER · OASIS-MRI in-domain image fitting | PSNR: 57.76 ± 2.19 -> 62.80 ± 3.17 | Helped |
| Transfer the encoder trained on CelebA-HQ | STRAINER-10 · AFHQ out-of-domain image fitting | PSNR: 52.40 ± 4.21 -> 57.46 ± 3.39 | Helped |
| Transfer the encoder trained on CelebA-HQ | STRAINER-10 · OASIS-MRI out-of-domain image fitting | PSNR: 65.06 ± 1.04 -> 72.21 ± 8.73 | Helped |
| Use the CelebA-HQ green channel for training and transfer (ablation) | STRAINER-10 · OASIS-MRI out-of-domain image fitting | PSNR: 72.21 ± 8.73 -> 74.61 ± 9.96 | Helped |
| Train the encoder on Stanford Cars rather than CelebA-HQ (ablation) | STRAINER-10 · OASIS-MRI out-of-domain image fitting | PSNR: 59.50 -> 59.66 | Helped |
| Train the encoder on Flowers or Stanford Cars rather than CelebA-HQ (ablation) | STRAINER-10 · AFHQ out-of-domain image fitting | PSNR: 57.46 -> 56.98; 57.46 -> 56.88 | Hurt |
| Share five encoder layers instead of three (ablation) | STRAINER-10 for super-resolution · Super-resolution (Fast) | PSNR: 32.10 -> 31.56 | Hurt |
| Use a three-layer shared encoder instead of a five-layer shared encoder (ablation) | STRAINER-10 for super-resolution · Super-resolution (HQ) | PSNR: 32.10 -> 32.43 | Helped |
| Increase the number of shared encoder layers (ablation) | STRAINER · Image fitting | PSNR | Helped |
| Use transferred encoder initialization | STRAINER-10 · CelebA-HQ denoising | PSNR: 26.75 ± 1.67 -> 26.41 ± 1.39 | Hurt |
| Use transferred encoder initialization | STRAINER-10 (Fast) · Super-resolution (Fast) | PSNR: 32.10 -> 31.56 | Hurt |
Surprising findings
- Encoder trained on CelebA-HQ transfers out-of-domain and outperforms the in-domain STRAINER-10 on OASIS-MRI and AFHQ.
- Fine-tuning SIREN on a single other image is worse than STRAINER with a single training image, indicating one-image encoders are insufficient.
Limitations
- Occasional PSNR drops during test-time fitting due to tuned encoder with random decoder.
- Unclear what features transfer and how they compare to transfer learning in CNNs and Transformers.
- Evaluation is mostly on small sets of images (10 training images per dataset) and a single shared architecture (SIREN with 6 layers).
Experimental setup
Datasets: CelebA-HQ AFHQ OASIS-MRI Stanford Cars Flowers Kodak DIV2K ShapeNet Thai Statue
Benchmarks: CelebA-HQ AFHQ OASIS-MRI Kodak DIV2K
Metrics: PSNR SSIM LPIPS IOU
Baselines: SIREN SIREN fine-tuned Meta-learned 5K TransINR IPC WIRE FINER DINER
Builds on: SIREN MAML Reptile SplineCAM
Closest papers by meaning
- Implicit Representations via Operator Learning (ICML 2024 · similarity 0.89)
- Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals (NeurIPS 2025 · similarity 0.89)
- VI^3NR: Variance Informed Initialization for Implicit Neural Representations (CVPR 2025 · similarity 0.88)
- SINR: Sparsity Driven Compressed Implicit Neural Representations (CVPR 2025 · similarity 0.88)
- MIRE: Matched Implicit Neural Representations (CVPR 2025 · similarity 0.88)
- Towards Lossless Implicit Neural Representation via Bit Plane Decomposition (CVPR 2025 · similarity 0.87)
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.