cstr/text-super-resolution-gguf

Text Super-Resolution & Restoration GGUF Models

Lightweight super-resolution and image restoration models converted to GGUF for CrispEmbed OCR preprocessing.

Models

File	Architecture	Params	Scale	Size	License	Paper
tbsrn-telescope-f16.gguf	TBSRN (text-line SR)	1.13M	2x	2.2 MB	Apache-2.0	CVPR 2021
pan-x4-f16.gguf	PAN (pixel attention)	272K	4x	0.5 MB	Apache-2.0	ECCV 2020W
hat-sr-x4-f16.gguf	HAT (hybrid attention transformer)	21M	4x	40 MB	MIT	CVPR 2023
dat-light-x2-f16.gguf	DAT-light (dual aggregation transformer)	830K	2x	38 MB	Apache-2.0	ICCV 2023
restormer-denoise-f16.gguf	Restormer (denoising)	26M	1x	50 MB	Apache-2.0	CVPR 2022

TBSRN Telescope (text-line SR)

• Task: Enhance individual detected text lines before recognition
• Input: Text-line crop resized to 16x64 -> Output: 32x128 (2x)
• Source: PaddleOCR sr_telescope (Apache-2.0)

PAN (whole-image 4x SR)

• Task: Upscale full document pages (rescues 75dpi text)
• Input: Any RGB image (tiled) -> Output: 4x upscale
• Source: PaddleGAN pan_x4 (Apache-2.0)

HAT (hybrid attention transformer, 4x SR)

• Task: High-quality 4x upscaling (CVPR 2023 SOTA on multiple SR benchmarks)
• Input: Any RGB image (tiled) -> Output: 4x upscale
• Architecture: Swin Transformer + overlapping cross-attention + channel attention
• Source: XPixelGroup/HAT (MIT)

DAT-light (dual aggregation transformer, 2x SR)

• Task: High-quality 2x upscaling with dual spatial+channel attention
• Input: Any RGB image (tiled) -> Output: 2x upscale
• Architecture: Split-channel windowed spatial attention + L2-normalized transposed channel attention + AIM + SGFN
• Source: zhengchen1999/DAT (Apache-2.0)

Restormer (image denoising/restoration)

• Task: Remove noise from document scans
• Input: Any RGB image -> Output: Denoised (same size)
• Architecture: Multi-Dconv head transposed attention, U-Net encoder-decoder
• Source: swz30/Restormer (Apache-2.0)

Parity Verification

All models pass the CrispEmbed diff harness (Python reference vs C++ engine):

Model	cos_sim	Status
TBSRN	0.999985	PASS
PAN	0.999654	PASS
HAT	0.999990	PASS
DAT-light	0.999956	PASS
Restormer	1.000000	PASS

Usage with CrispEmbed

from crispembed import CrispPanSr, CrispDatSr

# PAN: 4x upscale
sr = CrispPanSr("pan-x4-f16.gguf")
out, ow, oh = sr.process(pixels, width, height)

# DAT: 2x upscale (higher quality)
sr = CrispDatSr("dat-light-x2-f16.gguf")
out, ow, oh = sr.process(pixels, width, height)

# CLI
crispembed --pan-model pan-x4-f16.gguf --pan-sr input.png > output.ppm
crispembed --dat-model dat-light-x2-f16.gguf --dat-sr input.png > output.ppm

License

Apache-2.0 for all models except HAT (MIT). Both licenses are permissive.

Provenance and EU AI Act Art. 53 note

• Upstream model: multiple upstreams (TBSRN/Telescope, PAN, …) listed per file in the card.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.