Research map / Image generation / Image restoration
UGoDIT: Unsupervised Group Deep Image Prior Via Transferable Weights
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 UGoDIT, which trains a shared encoder with M disentangled decoders on a handful of degraded measurements, then freezes the encoder at test time and optimizes only a DIP decoder for faster, more robust unsupervised reconstruction.
Problem
Inverse imaging problems (multi-coil MRI, super resolution, non-linear deblurring) where clean fully sampled training data is scarce, and standalone DIP methods overfit noise and are slow because they optimize a full network per measurement.
Approach
Learn transferable encoder weights from only M degraded measurements using a shared encoder with M disentangled decoders and an input-adaptive autoencoding objective; at test time freeze the pretrained encoder and optimize only a randomly initialized decoder against measurement consistency.
Key results
- On 4x MRI, UGoDIT-6 reaches 36.05 dB PSNR vs 36.33 for SITCOM-MRI and 33.06 for VarNet, with run-time 48.4s vs 76.2s for SITCOM-MRI.
- On super resolution, UGoDIT-6 reaches 30.75 dB PSNR vs 29.45 for STRAINER-6 trained on clean images, while SITCOM (49000 clean training images) reaches 30.68 dB.
- Compared with aSeqDIP, UGoDIT gives at least 2 dB gain on MRI and at least 1 dB gain on non-linear deblurring; on SR it reaches 30 dB in about 125 steps versus more than 250 for aSeqDIP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use disentangled decoders rather than a shared single decoder during training (ablation) | UGoDIT · Super resolution | PSNR: approximately over a 1 dB to 1.5 dB improvement | Helped |
| Increase training images from 4 to 6 (ablation) | UGoDIT · fastMRI knee, 4x AF | PSNR: 35.45 -> 36.05 | Helped |
| Increase training images from 4 to 6 (ablation) | UGoDIT · fastMRI knee, 8x AF | PSNR: 32.04 -> 32.54 | Helped |
| Increase training images from 4 to 6 (ablation) | UGoDIT · FFHQ super resolution | PSNR: 30.55 -> 30.75 | Helped |
| Increase training images from 4 to 6 (ablation) | UGoDIT · FFHQ non-linear deblurring | PSNR: 29.89 -> 30.19 | Helped |
| Use five convolutional layers in both encoder and decoder rather than fewer layers (ablation) | UGoDIT · Super resolution | PSNR | Helped |
| Increase the number of convolutional layers beyond five (ablation) | UGoDIT · Super resolution | PSNR | No clear effect |
| Reduce the number of convolutional layers below five (ablation) | UGoDIT · Super resolution | PSNR | Hurt |
| Increase the training set beyond six images (ablation) | UGoDIT · FFHQ non-linear deblurring | PSNR | No clear effect |
| Use the selected N and K settings rather than the tested alternatives (ablation) | UGoDIT · MRI, super resolution, and non-linear deblurring | PSNR | Helped |
| Freeze the pretrained encoder at test time rather than initialize and optimize it (ablation) | UGoDIT · MRI | PSNR: around 1dB difference | Helped |
| Set lambda to 2 rather than the other tested regularization values (ablation) | UGoDIT · MRI | PSNR: approximately 1 dB | Helped |
Surprising findings
- Out-of-distribution test on brain MRI with a knee-trained encoder converges more slowly and underperforms aSeqDIP, while SR OOD on ImageNet converges faster than aSeqDIP.
Limitations
- Assumes full access to the forward model; blind setting not addressed.
- Limited to 2D image recovery; 3D extension left to future work.
- Out-of-distribution performance is weaker for MRI (brain test on knee-trained encoder), converging slower than aSeqDIP.
Experimental setup
Datasets: fastMRI knee FFHQ fastMRI brain ImageNet
Benchmarks: fastMRI knee FFHQ
Metrics: PSNR SSIM LPIPS run-time
Baselines: aSeqDIP Self-Guided DIP Deep Random Projector ES-DIP VarNet DDS SITCOM-MRI DPS SITCOM STRAINER
Builds on: Deep Image Prior aSeqDIP STRAINER MetaDIP
Closest papers by meaning
- Image Reconstruction Via Autoencoding Sequential Deep Image Prior (NeurIPS 2024 · similarity 0.89)
- Untrained Neural Nets for Snapshot Compressive Imaging: Theory and Algorithms (NeurIPS 2024 · similarity 0.88)
- A Unified Model for Compressed Sensing MRI Across Undersampling Patterns (CVPR 2025 · similarity 0.87)
- A Principled Self-Referenced Early Stopping Approach for Deep Image Prior (NeurIPS 2026 · similarity 0.87)
- DEFT: Efficient Fine-tuning of Diffusion Models by Learning the Generalised $h$-transform (NeurIPS 2024 · similarity 0.86)
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.86)
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.