ICDAR 2025 Competition on End-to-End Document Image Machine Translation Towards Complex Layouts

Yaping Zhang; Yupu Liang; Zhiyang Zhang; Zhiyuan Chen; Lu Xiang; Yang Zhao; Yu Zhou; Chengqing Zong

cs.CVMar 10, 2026

ICDAR 2025 Competition on End-to-End Document Image Machine Translation Towards Complex Layouts

Authors: Yaping Zhang, Yupu Liang, Zhiyang Zhang, Zhiyuan Chen, Lu Xiang, Yang Zhao, Yu Zhou, Chengqing Zong

Organizations: Institute of Automation, Chinese Academy of Sciences · University of Chinese Academy of Sciences, Beijing, China · Fanyu AI Laboratory, Zhongke Fanyu Technology Company Ltd., Beijing, China.

Abstract

Document Image Machine Translation (DIMT) seeks to translate text embedded in document images from one language to another by jointly modeling both textual content and page layout, bridging optical character recognition (OCR) and natural language processing (NLP). The DIMT 2025 Challenge advances research on end-to-end document image translation, a rapidly evolving area within multimodal document understanding. The competition features two tracks, OCR-free and OCR-based, each with two subtasks for small (less than 1B parameters) and large (greater than 1B parameters) models. Participants submit a single unified DIMT system, with the option to incorporate provided OCR transcripts. Running from December 10, 2024 to April 20, 2025, the competition attracted 69 teams and 27 valid submissions in total. Track 1 had 34 teams and 13 valid submissions, while Track 2 had 35 teams and 14 valid submissions. In this report, we present the challenge motivation, dataset construction, task definitions, evaluation protocol, and a summary of results. Our analysis shows that large-model approaches establish a promising new paradigm for translating complex-layout document images and highlight substantial opportunities for future research.

Figures & tables

Figure 1: Document image examples of the DIMT 2025 challenge. (a)(b): Web document examples in Track 1; (c)(d): arXiv document examples in Track 2.

Dataset	Images	Language	Granularity	Document Type	Year
DocVQA [ 21 ]	12,767	English	Word	Industrial Reports	2021
VisualMRC [ 24 ]	10,197	English	Word	Website	2021
RDVQA [ 22 ]	8,514	English	Word	Data Analysis Report	2022
CS-DVQA [ 23 ]	600	English	Word	Industry Documents	2022
DUDE [ 25 ]	28,709	English	Word	Cross-Domain	2023
DITrans [ 8 ]	1,675	English	Word	Report, News, Ad.	2023

Table 1: Existing Visually-rich Document Image Datasets.

Figure 2: An example of input and output for Track 1 is shown in (a). The purple rectangular box highlights the text areas that need to be reordered and translated, with the words and their corresponding bounding boxes (indicated by the red boxes) serving as the model’s input. For clarity, only the area within the rectangular box is illustrated, but in practice, the entire page is considered. (b) The results of reordering and translating the text in the rectangular box area.

Figure 3: An input and output example of Track 2. (a) is the original document image. (b) is the translated text in markdown format after rendering. The blocks with the same color represent the corresponding original text and translated text.

Track	Dataset	# of Examples
Track	Dataset	Train	Valid	Test
Track 1	DIMT-WebDoc-300K	300K	1K	1K
Track 2	DIMT-arXiv-124K	124K	1K	1K

Table 2: Statistical data information of the DIMT 2025 challenge.

Track	Subtrack	Participate	Submission
Track 1 OCR-based DIMT	LLM	18	6
	Small	16	7
	Total	34	13
Track 2 OCR-free DIMT	LLM	20	9
	Small	15	5
	Total	35	14

Table 3: The statistics of the competition submission

Rank	Team Name	Team Members	Institute	BLEU
Baseline	-	-	-	26.34
1	Hw-tsc	Zhanglin Wu, Tengfei Song, Ning Xie, Weidong Zhang, Pengfei Li, Shuang Wu, Chong Li, Junhao Zhu, and Hao Yang	Huawei Translation Service Center	70.48
2	Lucky Star	Zhentao Guo, Zining Wang, Tongkun Guan, Hao Sun, Chen Duan, and Kai Zhou	Beijing Institute of Technology, Meituan,Shanghai Jiao Tong University, Chinese Academy of Sciences	34.35
3	Abantika Bose	Abantika Bose	Friedrich-Alexander-Universität Erlangen-Nürnberg	10.64

Table 4: The results of Track 1.1 OCR-based DIMT-LLM

Rank	Team Name	Team Members	Institute	BLEU
Baseline	-	-	-	38.81
1	Hw-tsc	Zhanglin Wu, Tengfei Song, Ning Xie, Weidong Zhang, Pengfei Li, Shuang Wu, Chong Li, Junhao Zhu, and Hao Yang	Huawei Translation Service Center	66.16
2	CMC ATI	Doan Kien Thai, An Nguyen Thai, and Vu Son Lam Nguyen	CMC ATI	39.73
3	sunmk	Mai Sun, Jiazi Wang, and Long Yuan	Beijing Jiaotong University	37.06
4	Lucky Star	Zhentao Guo, Zining Wang, Tongkun Guan, Hao Sun, Chen Duan, and Kai Zhou	Beijing Institute of Technology, Meituan, Shanghai Jiao Tong University, Chinese Academy of Sciences	33.61
5	Abantika Bose	Abantika Bose	Friedrich-Alexander-Universität Erlangen-Nürnberg	9.40

Table 5: The results of Track 1.2 OCR-based DIMT-Small

Rank	Team Name	Team Members	Institute	BLEU
Baseline	-	-	-	48.63
1	Hw-tsc	Zhanglin Wu, Tengfei Song, Ning Xie, Weidong Zhang, Pengfei Li, Shuang Wu, Chong Li, Junhao Zhu, and Hao Yang	Huawei Translation Service Center	60.78
2	360ailab	Junhui Yu	360 AI Research Institute	57.78
3	Lucky Star	Zhentao Guo, Zining Wang, Tongkun Guan, Hao Sun, Chen Duan, and Kai Zhou	Beijing Institute of Technology, Meituan, Shanghai Jiao Tong University, Chinese Academy of Sciences	44.41

Table 6: The results of Track 2.1 OCR-free DIMT-LLM

Rank	Team Name	Team Members	Institute	BLEU
Baseline	-	-	-	27.03
1	Intime & HY	Gengluo Li, Huawen Shen, Chengquan Zhang, Xingyu Wan, Pengyuan Lyu, Liang Wu, Zhuohao Chen, Han Hu, and Yu Zhou	Chinese Academy of Sciences, Tencent, Nankai University	59.96
2	Hw-tsc	Zhanglin Wu, Tengfei Song, Ning Xie, Weidong Zhang, Pengfei Li, Shuang Wu, Chong Li, Junhao Zhu, and Hao Yang	Huawei Translation Service Center	59.56
3	Blue Jays	Cameron Carpenter, David Etter, Matt Lee, Paul McNamee, Kevin Duh, and Kenton Murray	Johns Hopkins University	58.41
4	Lucky Star	Zhentao Guo, Zining Wang, Tongkun Guan, Hao Sun, Chen Duan, and Kai Zhou	Beijing Institute of Technology, Meituan, Shanghai Jiao Tong University, Chinese Academy of Sciences	44.41

Table 7: The results of Track 2.2 OCR-free DIMT-Small

OCR-based
Model Type	Team Name	Method		SFT	DPO	Parameters	Extra Data
	hw-tsc	InternVL2.5-8B-MPO	70.48	$\checkmark$	$\checkmark$	8B
	Lucky Star	LayoutLM+qwen2.5-7B	34.35	$\checkmark$	×	7.3B
LLMs	Abantika Bose	LLaMA-402B	10.64	×	×	402B
	hw-tsc	InternVL2.5-1B-MPO	66.16	$\checkmark$	$\checkmark$	1B
	tadkt	LayoutLMv3+Qwen2.5-0.5B-Instruct	39.73	$\checkmark$	×	633M

Table 8: The method reports submitted by the participating teams

Image Text Translation

See all 8 papers

IMTBench: A Multi-Scenario Cross-Modal Collaborative Evaluation Benchmark for In-Image Machine Translation
Mar 11, 2026Jiahao Lyu, Pei Fu, Zhenhang Li +7Image Text Translation Multimodal MT
Institute of Information Engineering, Chinese Academy of Science, China · MiLM Plus, Xiaomi Inc, China · School of Cyber Security, University of Chinese Academy of Science, China +3
ForMaT: Dataset for Visually-Grounded Multilingual PDF Translation
May 15, 2026Michał Ciesiółka, Dawid Wiśniewski, Adrian Charkiewicz +1Multimodal MT Machine Translation
Laniqo, Pozna´n, Poland · Faculty of Mathematics and Computer Science, Adam Mickiewicz University, Pozna´n, Poland · Pozna´n University of Technology
LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR
Jan 20, 2026Said Taghadouini, Adrien Cavaillès, Baptiste AubertinVision-Language Models Multilingual OCR
LightOn

Mar 11, 2026J/K move · Enter open · S save

IMTBench: A Multi-Scenario Cross-Modal Collaborative Evaluation Benchmark for In-Image Machine Translation

Jiahao Lyu, Pei Fu, Zhenhang Li, Weichao Zeng+6

Institute of Information Engineering, Chinese Academy of Science, China · MiLM Plus, Xiaomi Inc, China · School of Cyber Security, University of Chinese Academy of Science, China+3

Image Text Translation Multimodal MT Machine Translation MT Evaluation

Open paper AskPDF arXiv

End-to-end In-Image Machine Translation (IIMT) aims to convert text embedded within an image into a target language while preserving the original visual context, layout, and rendering style. However, existing IIMT benchmarks are largely synthetic and thus fail to reflect real-world complexity, while current evaluation protocols focus on single-modality metrics and overlook cross-modal faithfulness between rendered text and model outputs. To address these shortcomings, we present In-image Machine Translation Benchmark (IMTBench), a new benchmark of 2,500 image translation samples covering four practical scenarios and nine languages. IMTBench supports multi-aspect evaluation, including translation quality, background preservation, overall image quality, and a cross-modal alignment score that measures consistency between the translated text produced by the model and the text rendered in the translated image. We benchmark strong commercial cascade systems, and both closed- and open-source unified multi-modal models, and observe large performance gaps across scenarios and languages, especially on natural scenes and resource-limited languages, highlighting substantial headroom for end-to-end image text translation. We hope IMTBench establishes a standardized benchmark to accelerate progress in this emerging task.

Abstract

Image Text Translation

Explore similar work

IMTBench: A Multi-Scenario Cross-Modal Collaborative Evaluation Benchmark for In-Image Machine Translation

ForMaT: Dataset for Visually-Grounded Multilingual PDF Translation

LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR