Abstract
Scene text image super-resolution must preserve character identity and glyph geometry, including approximate symmetry where present and asymmetric details that distinguish characters. Erroneous optical character recognition (OCR) priors can distort characters during generative reconstruction. We combine constrained large language model (LLM) correction and Mamba-based visual sequence modeling in a conditional residual diffusion method. The LLM proposes replacements within OCR candidate and edit constraints; deterministic validation rejects invalid suggestions and retains the original transcription. Corrected text, position, and low-resolution structural priors are fused by sample-level reliability gating and injected through cross-attention. A bottleneck Mamba block models horizontal visual dependencies. Under separately trained TextZoom and Real-CE protocols, the method achieves peak signal-to-noise ratios of 25.06 and 24.38 dB and exact-match accuracies of 83.4% and 50.2%, respectively. Five-seed paired ablations show that constrained OCR correction improves full-model reconstruction by 0.07 dB in PSNR and 0.6 percentage points in exact-match accuracy on both TextZoom and Real-CE, with the corresponding 95% confidence intervals excluding zero. Compared with the tested lightweight Transformer, Mamba reduces reconstruction latency by 7.82 ms with comparable quality. Independent recognition evaluation supports improved readability. The results support constrained correction and complementary visual conditioning.