Grounded Multimodal Named Entity Recognition (GMNER) extracts entity spans, types, and visual regions from text-image pairs. Existing approaches face high annotation costs and limited generalization. We propose UniGen, a zero-shot GMNER framework leveraging large language models (LLMs). UniGen decomposes GMNER into sequential NER, Entity Verification & Grounding (EVG), and Visual Grounding (VG) stages: NER extracts entities without visual interference, EVG resolves type ambiguities using visual context, and VG predicts bounding boxes. This complementary flow mitigates cross-modal hallucination. We employ a multimodal chain-of-thought (CoT) prompting strategy with adaptive example selection, where all demonstrations are LLM-generated. Results show UniGen approaches supervised baselines without training, with significant improvements over existing prompting methods.
Hairstyle transfer aims to realistically integrate a hairstyle from a reference image onto a source image while preserving facial identity and visual coherence. Current methods struggle with complex hairstyles and subtle details due to reliance on generic encoders, leading to visual artifacts and unrealistic blending. To overcome this, we propose HairCraft, an instance-specific adaptation method explicitly designed for precise hairstyle transfer. HairCraft personalizes a generative diffusion model to each hairstyle instance using only a single reference image, allowing detailed and faithful reproduction of complex hair features. We introduce a novel soft inpainting mechanism, carefully refining hair boundaries during denoising to ensure seamless and natural integration. HairCraft also supports optional user interaction through natural-language prompts and spatial masks, enabling flexible control over hairstyle attributes and placement. Quantitative evaluations and perceptual user studies show that HairCraft outperforms existing state-of-the-art methods in realism, detail preservation, and user control.
High-throughput Whole Exome Sequencing (WES) generates high-dimensional features for genomic variants to classify with machine learning classifiers. Although traditional feature selection techniques perform better, selecting Top-k significant features remains challenging. Alternatively, the SHAP-based global interpretable plot provides importance scores for features trained on a TreeSHAP explainer. However, determining compact significant features from SHAP is also limited, and thus, this paper proposes an ensemble-XAI technique based on average importance thresholding, consensus, and majority voting that integrates the performance of traditional and SHAP-based feature selection techniques. Experimentation has been carried out on high-dimensional WES data from Asian cohorts, and the results show that the proposed method has comparable performance on various metrics with other methods on a minimal feature set and has been validated statistically.
Depth from Focus (DFF) techniques have recently shown strong potential in depth estimation by analyzing focal stacks—images captured at varying focal distances. While Convolutional Neural Networks (CNNs) have been applied to DFF, they often struggle to capture the complex defocus relationships across focal planes. To address this, we propose a Transformer-based architecture tailored for DFF, capable of modeling these intricate focus-defocus cues more effectively. We introduce the Up-Guided Channel Attention (UGCA) module to enhance feature fusion and information flow, leading to more accurate depth predictions. Additionally, a newly designed resampling module preserves fine details and reduces artifacts. Experiments across multiple datasets demonstrate that our method achieves state-of-the-art performance in both quantitative accuracy and visual quality.
