Jasonleex1995/ReasonText-SR-outputs

ReasonText SR Outputs

Super-resolution outputs on the ReasonText benchmark, for reproducing the results in Tables 1 and 4 of the paper "Reading the Unreadable: Text-Aware Image Super-Resolution Needs Reasoning" (NeurIPS 2026).

🌐 Project Page · 📄 arXiv (coming soon) · 💻 Code

Contents

table1/       # Table 1: 12 SR models, each with its own built-in captioner (baselines)
  A1_Real-ESRGAN/  A2_SwinIR/                        # A: GAN-based
  B1_DiffBIR/  B2_SUPIR/  B3_FaithDiff/  B4_DiT4SR/  # B: multi-step diffusion
  C1_ResShift/  C2_SinSR/  C3_OSEDiff/               # C: few-step diffusion
  D1_PURE/                                           # D: AR-based
  E1_TeReDiff/  E2_UniT/                             # E: text-aware SR

table4_rtc/   # Table 4: 6 text-conditioned backbones + RTC (our method)
  B2_SUPIR/  B3_FaithDiff/  B4_DiT4SR/               # B: multi-step diffusion
  C3_OSEDiff/                                        # C: few-step diffusion
  E1_TeReDiff/  E2_UniT/                             # E: text-aware SR

Every model folder contains 513 PNGs (512×512) whose filenames match the ReasonText image stems, so each output aligns 1:1 with the benchmark's HR/, LR/, and annotations. table1/* corresponds to Table 1; table4_rtc/* corresponds to Table 4 (+RTC).

Citation

If you find our work useful for your research, please consider citing our paper:

<fill in BibTeX>

License

These images were produced by running each super-resolution model on ReasonText's LR images. Each SR model is subject to its own license; please refer to the respective model's repository for terms.