Research map / Image generation / Super-resolution
Boosting Flow-based Generative Super-Resolution Models via Learned Prior
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.
Adds a conditional learned prior (a latent module predicting a latent code from the LR image) in front of frozen flow-based SR models (LINF, SRFlow) to fix grid artifacts, exploding inverses, and fixed-temperature limits.
Problem
Flow-based super-resolution models suffer from grid artifacts, exploding inverses in invertible networks, and suboptimal quality from a fixed sampling temperature.
Approach
A lightweight latent module (dense-block encoders plus a UNet latent generator) predicts a conditional latent code from the LR image and an initial prior (upsampled-LR latent) in a single pass; this code is fed to a frozen pretrained flow model, replacing random Gaussian sampling and removing the need to tune temperature.
Key results
- SRFlow-LP reaches 27.51 dB PSNR and 0.109 LPIPS on DIV2K 4x, versus 27.09 dB and 0.121 for SRFlow (tau=0.9).
- RRDB-LINF-LP reaches 28.00 dB PSNR and 0.105 LPIPS on DIV2K 4x, versus 27.33 dB and 0.112 for RRDB-LINF (tau=0.8).
- SRFlow-LP eliminates exploding inverses (0% vs 6.7% for SRFlow at tau=1.0), and LINF-LP improves LPIPS at out-of-distribution scales such as 8x on Urban100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add a conditional learned prior at inference | SRFlow · DIV2K validation 4× | PSNR: 27.09 -> 27.51 | Helped |
| add a conditional learned prior at inference | SRFlow · DIV2K validation 4× | LPIPS: 0.121 -> 0.109 | Helped |
| add a conditional learned prior at inference | EDSR-baseline-LINF · DIV2K validation 4× | PSNR: 27.02 -> 27.64 | Helped |
| add a conditional learned prior at inference | RRDB-LINF · DIV2K validation 4× | LPIPS: 0.112 -> 0.105 | Helped |
| add a conditional learned prior at inference | RRDB-LINF · Urban100 8× | LPIPS: 0.354 -> 0.331 | Helped |
| add a conditional learned prior | SRFlow · DIV2K validation 4× | %Inf: 0.8 -> 0 | Helped |
| use EDSR-baseline rather than UNet as the latent generator (ablation) | EDSR-baseline-LINF-LP · DIV2K validation 4× | LPIPS: 0.119 -> 0.121 | Hurt |
| use Swin-T rather than UNet as the latent generator (ablation) | EDSR-baseline-LINF-LP · DIV2K validation 4× | LPIPS: 0.119 -> 0.120 | Hurt |
| use EDSR-baseline rather than UNet as the latent generator (ablation) | RRDB-LINF-LP · DIV2K validation 4× | LPIPS: 0.105 -> 0.110 | Hurt |
| use Swin-T rather than UNet as the latent generator (ablation) | RRDB-LINF-LP · DIV2K validation 4× | LPIPS: 0.105 -> 0.105 | No clear effect |
| omit the initial prior and use only LR image features (ablation) | LINF-LP · DIV2K 4× SR | LPIPS: 0.1297 -> 0.201 | Hurt |
| omit LR image features and use only initial-prior features (ablation) | LINF-LP · DIV2K 4× SR | LPIPS: 0.1297 -> 0.125 | Helped |
Surprising findings
- Using only the L_percep loss without the latent L1 loss gives the best LPIPS for LINF-LP, while latent-only L1 loss yields averaged, lower-quality predictions.
- Using only LR image features as input to the latent module degrades LPIPS versus the baseline (0.201 vs 0.130 on DIV2K 4x).
Limitations
- Gains are proportional to the capacity of the underlying flow model, with weaker backbones (EDSR-baseline) showing smaller improvements.
- Latent L1 loss must be used as regularization for SRFlow, and the loss weighting differs between the two integrations (lambda 0 vs 0.1).
Experimental setup
Datasets: DIV2K Flickr2K Set5 Set14 B100 Urban100
Benchmarks: DIV2K validation Set5 Set14 B100 Urban100
Metrics: PSNR SSIM LPIPS LR-PSNR exploding inverse rate (%Inf)
Baselines: LINF SRFlow HCFlow++ ESRGAN RankSRGAN SRDiff SROOE MetaSR LIIF LTE
Builds on: LINF SRFlow UNet VGG perceptual loss
Closest papers by meaning
- Efficient Conditional Diffusion Model with Probability Flow Sampling for Image Super-resolution (AAAI 2024 · similarity 0.89)
- HiFlow: Training-free High-Resolution Image Generation with Flow-Aligned Guidance (NeurIPS 2025 · similarity 0.88)
- STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis (NeurIPS 2025 · similarity 0.88)
- Solving Inverse Problems with FLAIR (NeurIPS 2025 · similarity 0.87)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.87)
- One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation (ICML 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.