Ji Ha Jang
Hi! 👋 I'm Ji Ha Jang. I'm currently pursuing an integrated PhD course in Electrical and Computer Engineering (ECE) at Seoul National University (SNU), advised by Prof. Se Young Chun. I earned my B.S. degree in ECE at Seoul National University.
Research Keywords Multimodal AI Generative AI Commonsense AI
News
Research
I'm interested in multimodal, generative, commonsense AI, and low-level computer vision. My work is driven by a deep curiosity about how AI can better understand and interact with the complexities of the world, combining various modalities. Highlighted papers are representative works.
We propose HyFL-CLIP (Hyperbolic fine-tuning of CLIP) for enhancing long-context image-text alignment in CLIP. HyFL-CLIP models hierarchical part-to-whole semantics in hyperbolic space by linking long descriptions, short textual components, and images. It achieves robust performance on long-context retrieval, perturbation-robust retrieval, intra-modality retrieval, and short-text retrieval benchmarks.
We propose UNCHA for enhancing hyperbolic VLMs. UNCHA models part-to-whole semantic representativeness with hyperbolic uncertainty, assigning lower uncertainty to more representative parts and higher uncertainty to less representative ones. UNCHA achieves state-of-the-art performance on zero-shot classification, retrieval, and multi-label classification benchmarks.
We propose RoMaP, a novel framework for local 3D Gaussian editing that enables precise and flexible part-level modifications. RoMaP introduces a geometry-aware 3D mask prediction module and a regularized SDS loss to constrain edits to target regions while preserving context.
We present INTRA, a novel framework for affordance grounding which enables training without egocentric images, grounds different parts for different interactions on the same object, and enables free-form text input.
We propose PODIA-3D, a novel pipeline that uses pose-preserved text-to-image diffusion-based domain adaptation for 3D generative models.