cstr/parseq-GGUF

PARSeq — Scene Text Recognition (GGUF)

GGUF conversions of PARSeq (ECCV 2022) for use with CrispEmbed.

PARSeq is a scene text recognition model that reads text from natural images (signs, labels, documents). It recognizes 94 printable ASCII characters (digits, letters, punctuation).

Architecture

• Encoder: 12-layer pre-LN ViT (patch 4×8, input 32×128 RGB, 128 tokens, GELU FFN)
• Decoder: 1-layer two-stream Transformer (XLNet-style position queries + context self-attention, then cross-attention to encoder memory)
• Head: Linear → 95 classes (94 printable ASCII chars + EOS)
• Inference: Autoregressive greedy decode (max 25 characters)

Variants

File	Variant	Params	Size	Notes
parseq-f32.gguf	Base	24M	91 MB	Full precision
parseq-q8_0.gguf	Base	24M	24 MB	Best quantized
parseq-q4_k.gguf	Base	24M	13 MB	Smallest base
parseq-tiny-f16.gguf	Tiny	6M	12 MB	Half precision
parseq-tiny-q8_0.gguf	Tiny	6M	6 MB	Smallest overall

All quantization levels produce identical output on test images.

Usage

# CLI
crispembed -m parseq-q8_0.gguf --ocr image.png

# Auto-download
crispembed -m parseq --auto-download --ocr image.png

from crispembed import CrispMathOcr
ocr = CrispMathOcr("parseq-q8_0.gguf")
text = ocr.recognize("sign.png")

Benchmark (94-char, PARSeq-base)

Dataset	Accuracy
IIIT5k	99.1%
SVT	97.9%
IC13-1015	98.1%
IC15-2077	89.2%
SVTP	96.9%
CUTE80	98.6%

Source

• Paper: Scene Text Recognition with Permuted Autoregressive Sequence Models (ECCV 2022)
• Code: baudm/parseq (Apache-2.0)
• Converted with models/convert-parseq-to-gguf.py from CrispEmbed

Provenance and EU AI Act Art. 53 note

• Upstream model: baudm/parseq.
• 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.