Confidence-Aware Ensemble and Long-Word Refinement for Artistic Text Recognition

WordArt (artistic text) features highly customized fonts, textures, and layouts, making WordArt-oriented scene TExt Recognition (WATER) substantially more challenging than general Scene Text Recognition (STR). Existing STR datasets and methods, typically built around regular scene text and fixed-template inputs, struggle to scale to WATER. Thus, we aim to advance this task from both data and model perspectives. On the data side, we construct a 2M synthetic dataset, WATER-S, with the scale improved by hundreds of times compared to existing artistic text data. WATER-S consists of two complementary subsets. One rendered by an upgraded rendering pipeline (SynthWordArt), which provides highly accurate and controllable synthetic WordArt data. The other is generated by combining Qwen3-VL for prompt mining and Z-Image for image synthesis, which improves the coverage of realistic and diverse data. On the model side, we propose WATERec. It adopts an visual encoder supporting arbitrary-shaped inputs and an autoregressive decoder to model complex layouts, structurally breaking the bottleneck of fixed-template STR on WordArt. Experiments show that this architecture outperforms prior STR methods, achieving state-of-the-art performance on irregular texts such as WordArt. Together with WATER-R, carefully reorganized from existing real STR data, our strong baseline with the new synthetic data and model design reaches 90.40% accuracy on WordArt-Bench, surpassing both general-purpose and OCR-specialized vision-language models by a large margin. Code and data are available at https://github.com/YesianRohn/WATER.

Scene text recognition is reported as 89--97% accurate on the six standard benchmarks, and the problem is widely treated as saturated. We present an alternative reading. When the same test images are stratified jointly by ground-truth word rarity and character n-gram novelty against a reference corpus, accuracy at the rare-word x rare-trigram corner of the resulting 5x5 grid drops 10--18 pt below the q3/q3 centre across nine English specialised recognisers, and the same direction (corner below centre) holds on all 13 of 13 (language, model) pairs we test across four writing systems (Latin, Han, Han+kana, Arabic). The drop is not a capacity bottleneck. A 6x vision-backbone scale-up (CLIP4STR-Base 158M -> CLIP4STR-Huge 1.0B, OpenCLIP ViT-H/14 LAION-2B) leads every benchmark in aggregate accuracy yet leaves the stress corner unchanged (86.9 -> 86.5, within paired-bootstrap noise). Four converging probes--layer-wise probing, confidence-when-wrong, attention re-balancing, and a cross-script commit-vs-abstain error split--localise the failure to the autoregressive decoder's lexical prior. We then ask how much of the gap existing techniques recover. Of 16 non-architectural mitigations, the largest mean q5/q5 gain is +1.3 pt and none clears the paired-bootstrap noise floor; the only intervention that does is the architectural shift from autoregressive to CTC decoding (SVTRv2, +2.5 pt, p=0.02, n=474). A confidence-routed AR-CTC ensemble adds a directionally consistent +0.6 pt that stays within noise, and its dominant learned coefficient is each model's own minimum-softmax confidence--independently echoing the mechanism above. No configuration we test improves both the compositional corner and aggregate accuracy. The rare-input long tail thus points to architectural change rather than added capacity.

Scene Text Recognition (STR) models are trained almost exclusively on word crops of at most 25 characters, yet real deployments (signage, product labels, dense captions) require reading much longer text. This paper diagnoses that failure and then closes it. The diagnosis separates out-of-length failure into two simultaneously extrapolating axes (the encoder's width axis and the decoder's time axis) and shows that encoder width, not decoder length, is the dominant failure mode. Representation-side fixes bring only partial relief: training-free rotary rescalings recover at most 2-4 points of character error rate (CER), and a weighted fine-tuning recipe recovers 6-8 points while improving standard-benchmark accuracy, yet word accuracy on the Long Text Benchmark (LTB) stays near zero, because the residual gap lies in the decoding mechanism rather than the representation. We then close that gap at inference time, on an unmodified word-level checkpoint: the long image is sliced into overlapping crops at the model's training width, each decoded independently and in-distribution, and the reads stitched by geometry-anchored edit-distance alignment. This procedure reaches 42.79-43.05% bucket-average word accuracy on LTB across two base checkpoints, matching the published state of the art (41.57%) and beating it by 11-12 points on the hardest bucket, at wall-clock parity with plain decoding; applied unchanged to the public PARSeq checkpoint it reaches 47.11%. Once chunking is applied fine-tuning no longer helps: the decoding-side fix alone matches purpose-built architectures. We release the diagnosis harness and implementation.