Codesota · Benchmark · IAMHome/Leaderboards/Vision & Documents/Handwriting Recognition/IAM
Unknown

IAM.

13,353 handwritten text lines from 657 writers. Standard handwriting benchmark.

Paper ↗Leaderboard ↓Lineage
§ 01 · SOTA history

Year over year.

§ 02 · Leaderboard

Results by metric.

Found a wrong score or missing run?
Use row edits to send a sourced correction into moderation.
Add / edit result ↗Report issue ↗

wer

Wer is the reported evaluation metric for IAM. Codesota tracks published model scores on this metric so readers can compare state-of-the-art results across sources and model families.

Lower is better

Trust tiers for werverifiedpapervendorcommunityunverified
RankModelTrustScoreYearLinksFix
01GPT-4o mini
GPT-4o-mini WER on IAM line-level. March 2025. From: Benchmarking Large Language Models for Handwritten Text Recognition (arxiv 2503.15195).
verified3.342026Source ↗Looks wrong?
02HTR-JAND
HTR-JAND (+LBC). Dec 2024. Table VI IAM WER. Includes Lexicon-Based Correction post-processing.
verified3.782024Source ↗Looks wrong?
03DRetHTR-base
DRetHTR-base. Feb 2026. Table 9: IAM Aachen split (IAM-A) WER.
verified6.552026Source ↗Looks wrong?
04MetaWriter
MetaWriter. CVPR 2025. IAM line-level WER.
verified10.322025Source ↗Looks wrong?
05HTR-ConvText
HTR-ConvText. Dec 2024. Table 2 IAM test set.
verified12.92024Source ↗Looks wrong?
06HTR-VT(line-level)
From paper: HTR-VT: Handwritten Text Recognition with Vision Transformer
verified14.92024Paper ↗Code ↗Looks wrong?
07HTR-VT
HTR-VT. Li et al. 2024. Table 4 IAM test set.
verified14.92024Source ↗Looks wrong?
08Leaky LP Cell
From paper: No Padding Please: Efficient Neural Handwriting Recognition
verified15.92019Paper ↗Code ↗Looks wrong?
09VAN
Vertical Attention Network (VAN). WER from comparison tables in HTR-VT (2409.08573) and HTR-ConvText (2512.05021). IAM line-level.
verified16.32022Source ↗Looks wrong?
10Decouple Attention Network
From paper: Decoupled Attention Network for Text Recognition
verified19.62019Paper ↗Code ↗Looks wrong?
11Start, Follow, Read
From paper: Start, Follow, Read: End-to-End Full-Page Handwriting Recognition
verified23.22018Paper ↗Code ↗Looks wrong?

cer

Cer is the reported evaluation metric for IAM. Codesota tracks published model scores on this metric so readers can compare state-of-the-art results across sources and model families.

Lower is better

Trust tiers for cerverifiedpapervendorcommunityunverified
RankModelTrustScoreYearLinksFix
01HTR-JAND
HTR-JAND with Lexicon-Based Correction (LBC) post-processing. Dec 2024. Joint Attention Network + Knowledge Distillation + curriculum learning. Table VI IAM comparison. Note: without LBC the model reaches ~2.34% CER (Table V ablation). IAM split not explicitly stated.
verified1.232024Source ↗Looks wrong?
02GPT-4o mini
GPT-4o-mini evaluated zero-shot on IAM line-level handwriting. March 2025. Outperforms Transkribus supermodel. From: Benchmarking Large Language Models for Handwritten Text Recognition (arxiv 2503.15195).
verified1.712026Source ↗Looks wrong?
03DRetHTR-base
DRetHTR-base: Decoder-only Retentive Network for HTR. Feb 2026. Table 9/11: IAM Aachen split (IAM-A), line-level. 1.6-1.9x faster inference and 38-42% less memory than Transformer baseline at same accuracy.
verified2.262026Source ↗Looks wrong?
04DTrOCR 105M
From paper: DTrOCR: Decoder-only Transformer for Optical Character Recognition
verified2.382023Paper ↗Code ↗Looks wrong?
05 Self-Attention + CTC + language model
From paper: Rethinking Text Line Recognition Models
verified2.752021Paper ↗Looks wrong?
06TrOCR-large
TrOCR-large (BEiT-large + RoBERTa-large). Microsoft. Table 4 in Li et al. 2023. IAM line-level test split. SOTA at publication.
verified2.892023Source ↗Looks wrong?
07TrOCR-large 558M
From paper: TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models
verified2.892021Paper ↗Code ↗Looks wrong?
08Transformer + CNN
From paper: Rethinking Text Line Recognition Models
verified2.962021Paper ↗Looks wrong?
09MetaWriter
MetaWriter: Personalized HTR via meta-learned prompt tuning. CVPR 2025. Table in paper: IAM line-level standard partition. Writer-adaptive; updates <1% of parameters at test time.
verified3.362025Source ↗Looks wrong?
10TrOCR-base
TrOCR-base (BEiT-base + RoBERTa-base). Microsoft. Table 4 in Li et al. 2023. IAM line-level test split.
verified3.422023Source ↗Looks wrong?
11TrOCR-base 334M
From paper: TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models
verified3.422021Paper ↗Code ↗Looks wrong?
12HTR-ConvText
HTR-ConvText: CNN+Transformer hybrid (ConvText block), 65.9M params, no pre-training. DAIR-Group, Dec 2024. Table 2: IAM line-level test set (6482/976/2915 split). Best among no-pretraining methods at publication.
verified4.002024Source ↗Looks wrong?
13TrOCR-small 62M
From paper: TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models
verified4.222021Paper ↗Code ↗Looks wrong?
14Transformer w/ CNN (+synth)
From paper: Pay Attention to What You Read: Non-recurrent Handwritten Text-Line Recognition
verified4.672020Paper ↗Looks wrong?
15HTR-VT
HTR-VT (Vision Transformer for HTR, no pre-training or synthetic data). Li et al. 2024. Table 4 IAM test set.
verified4.702024Source ↗Looks wrong?
16HTR-VT(line-level)
From paper: HTR-VT: Handwritten Text Recognition with Vision Transformer
verified4.702024Paper ↗Code ↗Looks wrong?
17VAN
Vertical Attention Network (VAN). Coquenet et al. IEEE TPAMI 2022. IAM line-level CER from comparison tables in HTR-VT (2409.08573) and HTR-ConvText (2512.05021).
verified5.002022Source ↗Looks wrong?
18Easter2.0
From paper: Easter2.0: Improving convolutional models for handwritten text recognition
verified6.212022Paper ↗Code ↗Looks wrong?
19FPHR+Aug Paragraph Level (~145 dpi)
From paper: Full Page Handwriting Recognition via Image to Sequence Extraction
verified6.302021Paper ↗Code ↗Looks wrong?
20Start, Follow, Read
From paper: Start, Follow, Read: End-to-End Full-Page Handwriting Recognition
verified6.402018Paper ↗Code ↗Looks wrong?
21Decouple Attention Network
From paper: Decoupled Attention Network for Text Recognition
verified6.402019Paper ↗Code ↗Looks wrong?
22FPHR+Aug Line Level (~145 dpi)
From paper: Full Page Handwriting Recognition via Image to Sequence Extraction
verified6.502021Paper ↗Code ↗Looks wrong?
23Leaky LP Cell
From paper: No Padding Please: Efficient Neural Handwriting Recognition
verified6.602019Paper ↗Code ↗Looks wrong?
24FPHR Paragraph Level (~145 dpi)
From paper: Full Page Handwriting Recognition via Image to Sequence Extraction
verified6.702021Paper ↗Code ↗Looks wrong?
25Transformer w/ CNN
From paper: Pay Attention to What You Read: Non-recurrent Handwritten Text-Line Recognition
verified7.622020Paper ↗Looks wrong?
Lineage

IAM in context.

See full ocr benchmarks lineage →
None — this is where the lineage begins.
This benchmark (1)
active2002-09
IAM
Successors (2)
saturated2015-08
ICDAR 2015
From clean handwritten text to incidental scene text — same 'read the pixels' task, fundamentally different visual domain. Spawned the decade of detection-then-recognition pipelines.
saturated2019-05
FUNSD
From transcription to structure: FUNSD reframed OCR as 'find the question, link to its answer' rather than 'recognise every character'. The shift that produced LayoutLM and the entire form-understanding line.
§ 04 · Submit a result

Add to the leaderboard.

← Back to Handwriting Recognition