Research map / Image generation / Super-resolution

PRISM: Prior Rectification and Uncertainty-Aware Structure Modeling for Diffusion-Based Text Image Super-Resolution

NeurIPS 2026 · Image generation › Super-resolution · approach: Perceptual and real-world SR · 0 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 PRISM, a single-step diffusion Text-SR framework that rectifies LQ text priors via flow matching toward a privileged LQ/HQ prior space and adds uncertainty-aware structural residual control for stroke boundaries.

Problem

Text image super-resolution under severe degradation, where text-aware conditions inferred from low-quality inputs are unreliable and global priors do not determine fine stroke boundaries, causing character-identity errors.

Approach

Builds a privileged prior from paired LQ/HQ latents at training time, learns a flow-matching transport from LQ-only embeddings to this prior space (FMPR), and adds an uncertainty-aware structural residual encoder (SURE) that gates Sobel-supervised edge cues by predicted variance, all within one diffusion pass on SD 2.1.

Key results

• On BTL-test at x4, PRISM reaches LPIPS 0.2314, FID 12.57, ACC 42.12%, NED 0.6644, best among compared methods.
• On RealCE-val at x4, PRISM improves ACC from 60.62% to 65.19% and FID from 74.52 to 47.83 over the second-best results.
• PRISM runs one diffusion step at about 0.08 s per image, versus 10.70 s for DiffTSR (200 steps) and 5.27 s for TeReDiff (50 steps).

Surprising findings

• PRISM has a lower PSNR than some baselines (e.g., TBSRN, TATT at x2 on BTL-test), which the authors attribute to the perception-distortion tradeoff.

Limitations

• Scope limited to Chinese-English text-line SR with moderate to long aspect ratios; not full-scene or dense multi-line text.
• Language coverage of BTL is mainly Chinese, English and digits.
• Lower PSNR than several baselines, reflecting perception-distortion tradeoff.

Experimental setup

Datasets: BTL-train TextZoom RealCE CTR SA-Text BSRGAN degradation Real-ESRGAN degradation

Benchmarks: BTL-test RealCE-val

Metrics: PSNR LPIPS FID ACC NED

Baselines: TSRN TBSRN TATT MARCONet DiffTSR StyleSRN TeReDiff

Builds on: Stable Diffusion 2.1 ControlNet LoRA MARCONet rendering strategy flow matching

Closest papers by meaning

• Text-Aware Real-World Image Super-Resolution via Diffusion Model with Joint Segmentation Decoders (NeurIPS 2025 · similarity 0.89)
• SRSR: Enhancing Semantic Accuracy in Real-World Image Super-Resolution with Spatially Re-Focused Text-Conditioning (NeurIPS 2025 · similarity 0.88)
• Text-guided Explorable Image Super-resolution (CVPR 2024 · similarity 0.87)
• The Power of Context: How Multimodality Improves Image Super-Resolution (CVPR 2025 · similarity 0.87)
• Noise-Started One-Step Image Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement (NeurIPS 2026 · similarity 0.86)
• RAP-SR: RestorAtion Prior Enhancement in Diffusion Models for Realistic Image Super-Resolution (AAAI 2025 · similarity 0.86)