Junwen Chen

International Conference on Pattern Recognition (ICPR)

Authors: Junwen Chen, Peilin Xiong, Keiji Yanai

• We propose HOI-R1, the first reinforcement learning–based MLLM framework for Human-Object Interaction Detection (HOID) without relying on external detection modules.
• HOI-R1 introduces a text-based reasoning process and HOID-specific reward functions to enable pure language-driven interaction detection.
• Experiments on HICO-DET demonstrate that HOI-R1 achieves twice the baseline accuracy and exhibits strong generalization capability.

IEEE Access, doi: 10.1109/ACCESS.2026.3659132

Authors: Junwen Chen and Keiji Yanai

• Hybrid-SOV aligns HOID with the latest object-detection paradigms by incorporating an efficient hybrid encoder and a query-selection mechanism that directly constructs HOI queries from visual features, eliminating predefined embeddings and enabling more interpretable decoding.
• When coupled with DINO-v3 foundation features, Hybrid-SOV achieves state-of-the-art accuracy with superior inference efficiency.

Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV 2025)

Authors: Keming Wu, Junwen Chen, Zhanhao Liang, Yinuo Wang, Ji Li, Chao Zhang, Bin Wang, Yuhui Yuan

• We propose a hybrid layout framework for diffusion transformers that reduces dependence on large-scale semantic layout annotations while preserving spatial accuracy in text-to-image generation.
• Our method divides layout control into two stages—anonymous layout generation and semantic refinement—followed by a quality-tuning phase to enhance visual aesthetics with minimal extra computation.
• Experiments show that our approach surpasses SiamLayout in layout adherence and visual quality, achieving over 10× higher data efficiency using limited semantic layout data.

International Conference on Document Analysis and Recognition (ICDAR 2025)

Authors: Honghui Yuan, Junwen Chen, and Keiji Yanai

• We focus on generating ancient handwritten fonts, particularly Kuzushiji, to support the preservation of Japanese historical documents.
• We propose a few-shot, training-free method using vector images to convert modern fonts into Kuzushiji style.
• Our experiments show that the proposed method effectively produces high-quality Kuzushiji fonts, outperforming previous approaches.

The 36th British Machine Vision Conference (BMVC 2025)

Authors: Peilin Xiong, Junwen Chen, and Keiji Yanai

• We address localized subject-driven image editing by proposing PosBridge, a training-free and scalable framework for inserting custom objects into target scenes.
• Our method leverages positional embedding transplant and a Corner Centered Layout strategy to guide diffusion models in maintaining structural and appearance consistency.
• Experimental results show that PosBridge outperforms existing methods in structure preservation, visual fidelity, and computational efficiency.

The Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025)

Authors: Yuxuan Luo, Yuhui Yuan, Junwen Chen, Haonan Cai, Ziyi Yue, Yuwei Yang, Fatima Zohra Daha, Ji Li, Zhouhui Lian

• We introduce knowledge image generation as a new task and present MMMG, a large benchmark of 4,456 expert-validated knowledge image–prompt pairs across multiple disciplines and levels.
• MMMG enables systematic evaluation using a unified Knowledge Graph (KG) representation and a new MMMG-Score that measures factual fidelity and visual clarity.
• Experiments on 16 leading text-to-image models reveal major reasoning limitations, while our proposed FLUX-Reason baseline demonstrates promising performance for future research.

arXiv preprint arXiv:2505.22523

Authors: Junwen Chen, Heyang Jiang, Yanbin Wang, Keming Wu, Ji Li, Chao Zhang, Keiji Yanai, Dong Chen, Yuhui Yuan

• We introduce PrismLayers and PrismLayersPro, the first large-scale, high-fidelity datasets of multi-layer transparent images with accurate alpha mattes for text-to-image generation.
• We propose a training-free synthesis pipeline and a strong multi-layer generation model, ART+, built upon LayerFLUX and MultiLayerFLUX for high-quality layered image composition.
• Experiments and user studies show that ART+ surpasses the original ART and matches the visual quality of FLUX.1-[dev], establishing a foundation for future research in editable multi-layer image generation.

2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2025)

Authors: Junwen Chen, Yingcheng Wang, Keiji Yanai

• We propose SOV-STG-VLA, a transformer-based framework for Human-Object Interaction Detection (HOID) that disentangles object detection and verb recognition.
• Our method introduces Subject-Object-Verb (SOV) decoding, Specific Target Guided (STG) denoising, and a Vision-Language Advisor (VLA) to enhance representation learning and training efficiency.
• Experiments show that SOV-STG-VLA achieves state-of-the-art performance with six times faster convergence than recent transformer-based methods.

Proceedings of the 2nd International Workshop on Multimedia Content Generation and Evaluation: New Methods and Practice (McGE '24)

Authors: Jing Yang, Junwen Chen, Keiji Yanai

• We propose RecipeSD, a Stable Diffusion–based method that incorporates recipe text information to enhance food image synthesis.
• Using a pretrained recipe encoder and the Image-like Recipe Transformation (IRT) with our CookNet model, we effectively inject detailed recipe semantics into the diffusion process.
• Experiments show that RecipeSD generates high-quality, recipe-aligned food images, outperforming existing cross-modal synthesis methods.

Proceedings of the 5th ACM Workshop on Intelligent Cross-Data Analysis and Retrieval

Authors: Jing Yang, Junwen Chen, Keiji Yanai

• We propose a Cross-Modal Embedding Fusing Decoder (Cross Decoder) to enhance cross-modal recipe retrieval.
• By integrating the Cross Decoder into a GAN–transformer framework and using dynamic margin loss, we improve embedding reliability and retrieval performance.
• Experiments on the Recipe1M dataset show that our method outperforms state-of-the-art approaches in both retrieval accuracy and image generation quality.

Proceedings of the 5th ACM International Conference on Multimedia in Asia (MMAsia '23)

Authors: Jingbin Xu, Junwen Chen, Keiji Yanai

• We address the limitations of existing Panoptic Scene Graph (PSG) methods by proposing a new one-stage framework called Contextual Associated Triplet Queries (CATQ).
• CATQ decodes subject, object, and relation features through separate branches, guided by instance information and enhanced with a Triplet Context Fusion Block.
• Experiments demonstrate that CATQ significantly outperforms state-of-the-art methods, achieving 34.8 Recall@20 and 20.9 mRecall@20 with only half the training time.

Proceedings of the 8th International Workshop on Multimedia Assisted Dietary Management (MADiMa '23)

Authors: Yingcheng Wang, Junwen Chen, Keiji Yanai

• We focus on eating and diet multimedia analysis, emphasizing the need to detect eating activities in videos and images.
• To address the lack of eating-specific data, we introduce HowToEat, a large-scale dataset with 66 days of videos and 95k annotated images across 12 eating scenarios.
• Based on this dataset, we develop an eating analysis system that simultaneously detects hand-object interactions and eating actions using a single model.

2023 18th International Conference on Machine Vision and Applications (MVA)

Authors: Junwen Chen and Keiji Yanai

• We address the Human-Object Interaction (HOI) detection task by proposing QAHOI, a transformer-based method using query-based anchors for end-to-end HOI prediction.
• QAHOI employs a multi-scale architecture to capture spatial variations in object size and location, improving interaction recognition accuracy.
• Experiments on the HICO-DET benchmark show that QAHOI with a transformer backbone significantly outperforms state-of-the-art methods.

International Conference on Multimedia Modeling

Authors: Jing Yang, Junwen Chen, Keiji Yanai

• We propose TNLBT (Transformer-based Network for Large Batch Training), a simple yet effective framework for cross-modal recipe retrieval and image generation.
• TNLBT employs Transformer-based encoders for both image and text embedding, combines self-supervised and contrastive losses, and leverages large-batch training to enhance cross-modal learning.
• Experiments on Recipe1M demonstrate that TNLBT significantly outperforms state-of-the-art methods, and we confirm that large-batch training improves recipe embedding learning.

Proceedings of the 4th ACM International Conference on Multimedia in Asia (MMAsia '22)

Authors: Junwen Chen, Keiji Yanai

• We propose Parallel Query Network (PQNet) for Human-Object Interaction (HOI) detection, which separates human and object localization into parallel decoding branches.
• PQNet employs two transformer decoders for subject and object embeddings and introduces a verb decoder that fuses their representations through attention mechanisms.
• Experiments show that PQNet outperforms previous methods while requiring only half the training epochs.

International Symposium on Neural Networks

Authors: Junwen Chen, Yi Lu, Yaran Chen, Dongbin Zhao, Zhonghua Pang

• We propose ContourRend, a segmentation method designed to refine object contours and achieve clearer, more complete segmentation results.
• ContourRend integrates a contour renderer with a GCN-based segmentation model, focusing on high-resolution prediction around contour pixels.
• Experiments on the Cityscapes dataset show that ContourRend achieves 72.41% mIoU, outperforming the baseline Polygon-GCN by 1.22%.