GLYPH-SR: Glyph-Aware Scene Text SR

Updated 5 July 2026

GLYPH-SR is a scene-text super-resolution framework that explicitly restores glyph structures using vision-language guidance and OCR cues for improved readability.
It employs a dual-branch ControlNet with a ping-pong scheduler to alternate between text-focused and image-focused conditioning for balanced recovery.
The method yields significant OCR F1 score improvements over traditional approaches while noting limitations in multilingual and extreme imaging scenarios.

Searching arXiv for papers on GLYPH-SR, TIGER, implicit glyph representation, and Glyph symbolic regression. GLYPH-SR most commonly denotes a recent line of scene-text image super-resolution research in which glyph structure is treated as an explicit restoration target rather than as generic texture. In its most specific 2025 usage, GLYPH-SR is a vision-language-guided latent diffusion framework that seeks to achieve both high-quality image super-resolution and high-fidelity text recovery by combining OCR-derived text cues, scene-level conditioning, a dual-branch ControlNet, and a ping-pong guidance scheduler (Sung et al., 30 Oct 2025). Closely related work frames the same problem as glyph-structure-guided super-resolution, most notably the two-stage TIGER pipeline, which restores glyphs first and enhances the full image later (Luo et al., 24 Oct 2025). The surrounding literature also includes glyph-centered implicit shape modeling for resolution-independent rendering (Liu et al., 2021) and an older, unrelated usage of “Glyph” for a symbolic regression toolkit (Quade et al., 2018).

1. Terminology and scope

The available literature suggests that “GLYPH-SR” is not a uniform label with a single historical meaning. In the scene-text restoration literature, it refers either to a specific framework—“GLYPH-SR: Can We Achieve Both High-Quality Image Super-Resolution and High-Fidelity Text Recovery via VLM-guided Latent Diffusion Model?”—or more broadly to the idea that super-resolution should explicitly reconstruct glyph structure before or during full-image enhancement (Sung et al., 30 Oct 2025). A closely aligned embodiment of that broader idea is TIGER, introduced as “Restore Text First, Enhance Image Later: Two-Stage Scene Text Image Super-Resolution with Glyph Structure Guidance,” which explicitly decouples glyph restoration from image enhancement through a two-stage “text-first, image-later” pipeline (Luo et al., 24 Oct 2025).

A second, related but distinct strand is “Learning Implicit Glyph Shape Representation,” which does not introduce or use the acronym GLYPH-SR. Its contribution is an implicit, resolution-independent representation of glyph outlines; practically, it functions like glyph super-resolution because one can render arbitrarily high-resolution glyph images from a compact, continuous shape model (Liu et al., 2021). This is glyph-centered shape modeling rather than scene-text image super-resolution.

A further terminological caveat is necessary because the data also uses “GLYPH-SR” as a clarifying label for the Glyph symbolic regression toolkit. In that context, the toolkit is referred to simply as “Glyph,” and GLYPH-SR can be understood as the same project: a Python package for genetic programming based symbolic regression with a ZeroMQ-enabled remote evaluation model for experiments and simulations (Quade et al., 2018). That usage is unrelated to image restoration.

2. Problem formulation in scene-text super-resolution

The central problem addressed by GLYPH-SR is that conventional super-resolution systems often optimize for distortion or perceptual quality while remaining insensitive to character-level integrity. In natural images, scene-text on signs, storefronts, labels, and license plates often occupies a tiny fraction of pixels but carries the most actionable information. Conventional SR systems trained and evaluated for distortion or perceptual quality largely ignore character-level integrity because these metrics are global and area-dominated (Sung et al., 30 Oct 2025).

This leads to two failure modes. The first is hallucination: characters look sharp but read incorrectly. The second is conservative restoration: blurry text is preserved to avoid artifacts, with minimal SR gains (Sung et al., 30 Oct 2025). TIGER describes the same issue as a fundamental trade-off between image quality and textual readability. Generative SR methods can preserve natural-image fidelity while corrupting text, and even small structural errors such as missing or merged strokes can change semantics entirely (Luo et al., 24 Oct 2025).

GLYPH-SR therefore treats scene-text SR as a bi-objective problem. Its explicit aim is to maximize readability, measured through OCR fidelity, and visual realism, measured through no-reference image-quality assessment metrics (Sung et al., 30 Oct 2025). TIGER operationalizes the same objective by restoring precise text structures first and then using them to guide subsequent full-image super-resolution, so that restored glyphs constrain the image synthesis process rather than being repaired only after the fact (Luo et al., 24 Oct 2025).

A common misconception is that text in scene images can be treated as another class of high-frequency texture. The recent literature rejects that assumption. In both GLYPH-SR and TIGER, text is instead modeled as semantically constrained structure whose stroke topology, spacing, baseline, and localization must be preserved to support downstream OCR and human reading (Sung et al., 30 Oct 2025).

3. Architecture of the 2025 GLYPH-SR framework

GLYPH-SR augments a latent diffusion SR backbone with structured, OCR-informed text guidance and a time-varying scheduler that alternates between text-centric and image-centric conditioning (Sung et al., 30 Oct 2025). The input is a low-resolution image $I_{LR}$ . An LR-robust conditioner extracts multi-scale features used by a pretrained latent diffusion model in VAE latent space. Two vision-language conditions are then formed: a scene-level caption $S_{IMG}$ describing global attributes, and OCR-derived position-text pairs $\{(S_{text}^k, S_{pos}^k)\}$ for detected text instances. Each pair is verbalized into a structured phrase and encoded by a frozen text encoder.

The control mechanism is a dual-branch Text-SR Fusion ControlNet, or TS-ControlNet. Its SR-ControlNet branch is frozen and produces per-block residuals that preserve global structure and texture from $I_{LR}$ and $S_{IMG}$ . Its Text-ControlNet branch is trainable and produces glyph-centric residuals conditioned on OCR-derived text-position cues. Residuals are injected into the latent diffusion U-Net at each block. The mixed residual at block $l$ and step $t$ is

$\tilde r_l(t)= s_{ctrl} \left[(1-\lambda_t) r_l^{TXT} + \lambda_t r_l^{SR}\right].$

Scene and text embeddings are fused in the same time-dependent manner:

$e^t = (1-\lambda_t)e_{txt} + \lambda_t e_{img},$

with classifier-free guidance applied as

$\hat{\epsilon}_t = (1+\omega) D_\theta(z_t,t,e^t) - \omega D_\theta(z_t,t,\varnothing).$

The defining control policy is the ping-pong scheduler. For diffusion steps descending from $S_{IMG}$ 0 to $S_{IMG}$ 1, the scheduler alternates between text-focused and scene-focused guidance according to

$S_{IMG}$ 2

A binary “square-wave” schedule worked best, with default $S_{IMG}$ 3, so the guidance toggles every step (Sung et al., 30 Oct 2025). Text-focused steps inject precise glyph cues; interleaved image-focused steps stabilize global appearance. Sampling is performed by an EDM sampler in latent space, and the final latent is decoded by the VAE decoder to produce $S_{IMG}$ 4.

The training objective is likewise specialized. HR targets are encoded to $S_{IMG}$ 5, forward noising follows the standard DDPM form, and the denoising objective for the mixed condition $S_{IMG}$ 6 is

$S_{IMG}$ 7

Only the text branch in TS-ControlNet is updated; the diffusion backbone, VAE, text encoders, and SR branch are frozen (Sung et al., 30 Oct 2025). The paper does not introduce an explicit OCR recognition loss. OCR enters as a conditioning signal rather than as a direct supervision term.

4. Training protocol, benchmarks, and reported results

The GLYPH-SR training strategy is deliberately surgical. The LDM backbone, described as Juggernaut-XL style SDXL latent diffusion, the VAE encoder/decoder, the LR-robust SR conditioning branch, and the text encoders are frozen; only the Text-ControlNet branch is trainable (Sung et al., 30 Oct 2025). The rationale is to preserve global SR fidelity while selectively improving text legibility.

Training uses a factorized synthetic corpus with four subsets that independently vary text fidelity and global image quality: Positive-Text / High-Quality, Negative-Text / High-Quality, Positive-Text / Low-Quality, and Negative-Text / Low-Quality. Each sample is encoded as a composite condition that combines a latent image code, scene caption, OCR-derived text-position conditions, and a guide prompt token indicating the subset. Text-position pairs for conditioning are always extracted from the positive-text, high-quality set so that the model does not learn to copy erroneous text (Sung et al., 30 Oct 2025).

Evaluation is reported on SVT, SCUT-CTW1500, and CUTE80 at $S_{IMG}$ 8 and $S_{IMG}$ 9, using OpenOCR, GOT-OCR, and LLaVA-NeXT to compute OCR F1. No-reference IQA is measured with MANIQA, CLIP-IQA, and MUSIQ (Sung et al., 30 Oct 2025). The headline claim is an improvement of up to +15.18 percentage points OCR F1 over a diffusion/GAN baseline on SVT $\{(S_{text}^k, S_{pos}^k)\}$ 0 with OpenOCR, where GLYPH-SR achieves 48.79 versus 33.61 for SUPIR (Sung et al., 30 Oct 2025).

Setting	OCR F1 (OpenOCR / GOT-OCR / LLaVA-NeXT)	IQA (MANIQA / CLIP-IQA / MUSIQ)
SVT $\{(S_{text}^k, S_{pos}^k)\}$ 1	67.54 / 71.72 / 73.22	47.75 / 59.40 / 70.99
SCUT-CTW1500 $\{(S_{text}^k, S_{pos}^k)\}$ 2	38.26 / 36.96 / 42.90	70.33 / 57.88 / 70.31
CUTE80 $\{(S_{text}^k, S_{pos}^k)\}$ 3	73.09 / 55.62 / 85.01	49.77 / 65.93 / 69.96
SVT $\{(S_{text}^k, S_{pos}^k)\}$ 4	48.79 / 56.16 / 58.54	47.40 / 56.78 / 69.93
SCUT-CTW1500 $\{(S_{text}^k, S_{pos}^k)\}$ 5	11.09 / 14.71 / 14.67	61.94 / 48.21 / 63.43
CUTE80 $\{(S_{text}^k, S_{pos}^k)\}$ 6	63.66 / 45.65 / 73.71	47.75 / 65.85 / 68.85

Against baselines including BSRGAN, DiffBIR, DiffTSR, InvSR, PiSA-SR, Real-ESRGAN, StableSR, and SUPIR, GLYPH-SR attains top or second-best OCR F1 in 5 of 6 settings and top or second-best among MANIQA, CLIP-IQA, and MUSIQ in 26 of 30 cases. On SVT $\{(S_{text}^k, S_{pos}^k)\}$ 7 it leads all six metrics. On CUTE80 $\{(S_{text}^k, S_{pos}^k)\}$ 8 it leads all SR metrics and the OpenOCR F1 (Sung et al., 30 Oct 2025).

The ablations identify the role of OCR content and localization. Using both $\{(S_{text}^k, S_{pos}^k)\}$ 9 and $I_{LR}$ 0 gives the best legibility and geometry. Text-only conditioning preserves semantic correctness but yields warped baselines or kerning; position-only conditioning yields correct layout but misspelled words; removing both causes hallucinations and distortions resembling generic SR priors (Sung et al., 30 Oct 2025). Binary ping-pong scheduling outperforms continuous ramps tied to noise level at similar perceptual quality.

TIGER is the closest neighboring framework in both motivation and design. It introduces a two-stage “text-first, image-later” paradigm: Stage 1 performs glyph restoration, taking a low-resolution image and text region detections with OCR text strings and outputting a high-fidelity text mask $I_{LR}$ 1 positioned in original image coordinates; Stage 2 performs full-image super-resolution using the low-resolution image and the restored glyph mask as inputs (Luo et al., 24 Oct 2025). In Stage 1, each text crop is reconstructed by a latent diffusion model into a clean structural mask, and the restored regions are assembled into the full-image mask. In Stage 2, the glyph mask is encoded and injected into a ControlNet-like SR network that guides denoising of the low-resolution latent toward the high-resolution latent.

Its losses are correspondingly split. Stage 1 combines text-control diffusion loss and segmentation-oriented supervision:

$I_{LR}$ 2

with

$I_{LR}$ 3

Stage 2 uses an image reconstruction loss and an edge loss based on Sobel filtering:

$I_{LR}$ 4

$I_{LR}$ 5

TIGER is trained on synthetic data built upon LSDIR and on real data including Real-CE and UltraZoom-ST. UltraZoom-ST is presented as the first real-world scene text dataset with extreme zoom $I_{LR}$ 6, containing 5,036 LR–HR pairs and 49,675 text lines, with focal lengths 14 mm, 35 mm, 85 mm, and 200 mm, and evaluation splits for $I_{LR}$ 7, $I_{LR}$ 8, and $I_{LR}$ 9 (Luo et al., 24 Oct 2025). On UltraZoom-ST overall, TIGER reports PSNR 25.48, SSIM 0.830, LPIPS 0.196, DISTS 0.156, FID 20.01, PSNR_cr 22.22, SSIM_cr 0.814, LPIPS_cr 0.228, DISTS_cr 0.212, and OCR-A 0.430. On Real-CE $S_{IMG}$ 0 it reports PSNR 24.12, SSIM 0.839, LPIPS 0.164, DISTS 0.125, FID 38.72, PSNR_cr 23.43, SSIM_cr 0.864, LPIPS_cr 0.173, DISTS_cr 0.167, and OCR-A 0.673 (Luo et al., 24 Oct 2025). The reported ablations also show that TIGER glyph restoration outperforms standard font guidance, SAM-TS extraction, and LDM guidance conditioned on LR when Stage 2 is fixed.

A different relation to glyph super-resolution appears in “Learning Implicit Glyph Shape Representation.” There, a glyph is modeled as the union of $S_{IMG}$ 1 primitives, each enclosed by $S_{IMG}$ 2 quadratic curves, with $S_{IMG}$ 3 and $S_{IMG}$ 4 in the reported experiments (Liu et al., 2021). The learned implicit quadratic for curve $S_{IMG}$ 5 is

$S_{IMG}$ 6

and the training approximation forms

$S_{IMG}$ 7

$S_{IMG}$ 8

Because $S_{IMG}$ 9 is defined for any continuous coordinate, glyphs can be rendered at arbitrary resolution. The paper reports SSIM 0.9039 for reconstruction versus 0.7752 for a VAE baseline, and in one-shot font style transfer reports SSIM 0.6518, L1 21.62, and LPIPS 0.094 with 1.95M parameters (Liu et al., 2021). This work is not scene-text SR, but it demonstrates that explicit glyph geometry can support sharp, resolution-independent reconstruction.

6. Limitations, deployment considerations, and caveats

The principal limitations of GLYPH-SR are tied to OCR dependence and distributional scope. The method is validated on Latin-centric benchmarks, so multilingual scripts such as CJK, Arabic, and Devanagari are not explicitly addressed. Under severe blur, motion, occlusion, or very low light, even precise text guidance can fail if OCR cannot provide reliable strings or boxes. Because the system conditions on OCR output, hallucination can still occur when OCR or text prompts are wrong; in that case the model may faithfully render incorrect guidance (Sung et al., 30 Oct 2025).

TIGER exhibits closely related constraints. It relies on OCR to detect text boxes and extract transcripts for Stage 1, so failures in OCR propagate into glyph restoration. Extremely small text at $l$ 0 remains difficult, and the paper identifies inference speed as a practical concern because diffusion in Stage 1 requires multiple steps; real-time on mobile remains challenging (Luo et al., 24 Oct 2025). TIGER therefore recommends fast OCR detectors such as PP-OCRv3 or PP-OCRv5, batching Stage 1 across detected regions, and tile-based inference for Stage 2.

Both lines of work identify immediate application domains in surveillance, autonomy, document-in-the-wild analysis, and retail analytics, where downstream perception depends on readable text rather than on image realism alone (Sung et al., 30 Oct 2025). Both also note ethical or privacy concerns: enhanced readability may reveal sensitive information, and generated text should not be over-trusted in safety-critical settings without guidance logging, confidence estimation, or related audit mechanisms (Luo et al., 24 Oct 2025).

A final caveat is terminological. “GLYPH-SR” in current vision literature refers to glyph-aware scene-text super-resolution or to the specific 2025 VLM-guided latent diffusion model. It should not be confused with the older Glyph symbolic regression toolkit, whose focus is GP-based symbolic regression, multi-objective optimization, symbolic constant optimization, ZeroMQ-based remote experimentation, and applications such as control of the chaotic Lorenz system (Quade et al., 2018). This suggests that, in research usage, the term is best interpreted contextually: in computer vision it denotes glyph-structured scene-text super-resolution, whereas in symbolic regression it denotes an unrelated experimental software framework.