cstr / text-super-resolution-gguf

Text Super-Resolution GGUF Models

Lightweight super-resolution models converted to GGUF for CrispEmbed OCR preprocessing. Two models targeting different use cases:

Models

TBSRN Telescope (text-line SR)

• File : tbsrn-telescope-f16.gguf (2.2 MB)
• Architecture : TBSRN — Recurrent residual blocks + 4-head multi-head attention (FeatureEnhancer) + PixelShuffle 2x
• Params : 1.13M (inference-only, frozen transformer/GRU/STN excluded)
• Input : Text-line crop resized to 16x64
• Output : 32x128 (2x upscale)
• Source : PaddleOCR sr_telescope (Apache-2.0)
• Paper : Scene Text Telescope: Text-Focused Scene Image Super-Resolution (Chen et al., CVPR 2021)
• Use case : Enhance individual detected text lines before recognition

PAN (whole-image SR)

• File : pan-x4-f16.gguf (0.5 MB)
• Architecture : PAN — 16x SCPA blocks (self-calibrated convolution + pixel attention) + nearest-neighbor upsample + bilinear residual
• Params : 272K
• Input : Any size RGB image (tiled processing for large images)
• Output : 4x upscale
• Source : PaddleGAN pan_x4 (Apache-2.0)
• Paper : Efficient Image Super-Resolution Using Pixel Attention (Zhao et al., ECCV 2020 Workshop)
• Use case : Upscale full document pages before OCR

Usage with CrispEmbed

// TBSRN: text-line super-resolution
tbsrn_sr_context * ctx = tbsrn_sr_init("tbsrn-telescope-f16.gguf", 4);
uint8_t * output; int ow, oh;
tbsrn_sr_process(ctx, line_crop, w, h, &output, &ow, &oh);
// output is 32x128 RGB
tbsrn_sr_free_image(output);
tbsrn_sr_free(ctx);

// PAN: whole-image 4x upscale
pan_sr_context * ctx = pan_sr_init("pan-x4-f16.gguf", 4);
uint8_t * output; int ow, oh;
pan_sr_process(ctx, image, w, h, 128, 16, &output, &ow, &oh);
// output is 4*w x 4*h RGB
pan_sr_free_image(output);
pan_sr_free(ctx);

Parity Verification

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

PAN's higher max_abs is due to uint8 quantization in the output path (values outside [0,1] are clamped).

Conversion

Convert from PaddlePaddle .pdparams :

# TBSRN
python models/convert-tbsrn-to-gguf.py \
    --model sr_telescope_train/best_accuracy.pdparams \
    --output tbsrn-telescope-f16.gguf --fp16

# PAN
python models/convert-pan-to-gguf.py \
    --model pan_x4.pdparams \
    --output pan-x4-f16.gguf --fp16

License

Apache-2.0 (both source models and converted GGUF files).