Sovereign 8s.is mirror of standardgalactic/SteleSR
  • Python 85.4%
  • Cuda 8.7%
  • C++ 5.9%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-07-08 11:09:34 +08:00
SteleSR Add files via upload 2026-05-08 18:56:06 +08:00
README.md Update README.md 2026-07-08 11:09:34 +08:00

SteleSR

Bridging the Semantic-Visual Gap: A Universal Super-Resolution Framework for Ancient Stele Inscriptions

Ancient stele inscriptions face irreversible information loss due to millennia of environmental degradation. This paper presents the first systematic effort to address the task of Stele Inscription Text Image Super-Resolution. General-purpose models often suffer from a "semantic-visual imbalance" in this domain, where a lack of structural priors leads to spurious topological connections that render restored text illegible. To resolve this, we propose SISTR, the first systematic benchmark and progressive paradigm that bridges the gap between synthetic noise and real-world damage via LLM-driven structural decoupling and LVM-guided style injection. Furthermore, we introduce SteleSR, a universal restoration framework that integrates Edge-aware and Text Prior losses into diverse super-resolution backbones to ensure both visual authenticity and semantic legibility. Extensive evaluations across semantic and structural metrics demonstrate that SteleSR consistently enhances various models, establishing a pioneering standard for the high-fidelity digital preservation of historical artifacts.

image

Overview

image image

Real-World SR Results

image image

Environment

  • PyTorch >= 1.7
  • BasicSR == 1.3.4.9

How To Test

The HAT model is used as an example here.

  • Refer to ./options/test/HAT for the configuration files of the models to be evaluated.
  • Prepare the testing dataset and pretrained checkpoints before inference.
  • The testing results will be saved in the ./results directory.
python basicsr/test.py -opt options/test/HAT/HAT_SRx4_ImageNet-LR.yml

How To Train

The HAT model is used as an example here.

  • Refer to ./options/train/HAT for the training configuration files.
  • Prepare the training datasets before starting training.
  • Training logs and checkpoints will be saved in the ./experiments directory.

Stage 1: Training on SISTR-Syn

python basicsr/train.py -opt options/train/HAT/train_HAT-L_SRx4_ImageNet_from_scratch_test.yml

Stage 2: Finetuning on SISTR-LVM

python basicsr/train.py -opt options/train/HAT/train_HAT-L_SRx4_finetune_from_ImageNet_pretrain_test.yml