I am currently a Ph.D. candidate in Machine Learning at The University of Tokyo, Matsuo Lab. My work focuses on visual generative models, including image and video generation and editing, image-to-image translation, vision stylization, and efficient deep learning systems.
I expect to graduate in Fall 2026 and am actively seeking summer internship or full-time opportunities in visual generative modeling. Please feel free to contact me at secret_wang@outlook.com.
我目前在东京大学松尾研究室攻读人工智能博士。我的研究方向与经验主要关于视觉生成模型，包括图像与视频生成和编辑、图像到图像转换、视觉风格化以及移动端深度学习系统。
我预计将于 2026 年秋季毕业，目前正在积极寻找视觉生成模型方向的暑期实习或全职机会。欢迎通过 secret_wang@outlook.com 与我联系。
Abstract: This work studies distribution shift in reference-based sketch colorization, where real references can be spatially or semantically misaligned with sketches. It separates reference representations into modular stages for embedding guidance, background detail transfer, and global style transfer, improving visual quality while reducing spatial artifacts.
Abstract: The paper targets spatial entanglement in high-resolution sketch colorization. It introduces a dual-branch feature alignment framework with Gram regularization, plus anime-specific attribute control and texture-transfer modules, to improve resolution, controllability, and reference consistency.
Abstract: This work learns a portrait style from a single artistic reference. It combines geometric alignment, content and style LoRA optimization, orthogonal adaptation, and ControlNet guidance to produce high-quality stylized portraits with a smaller computation and parameter budget than common inversion or diffusion baselines.
Abstract: The method builds geometric correlations between portrait photos and style samples using facial landmarks and differentiable thin-plate-spline alignment. Its lightweight GAN framework improves training data efficiency and computational cost, enabling real-time portrait stylization on mobile devices.
Abstract: Inspired by professional animation pipelines, this diffusion framework uses sketches for spatial guidance and RGB images for color reference. It separates foreground and background reference signals with spatial masks and split cross-attention LoRA modules to reduce artifacts from mismatched references.
Abstract: This paper analyzes why Adam converges quickly but often generalizes worse than SGD. It separates the roles of adaptive learning rate and momentum in saddle-point escape and flat-minima selection, then proposes Adaptive Inertia, a parameter-wise adaptive momentum framework with strong empirical generalization.
Abstract: This framework converts digital illustrations into manga by mimicking studio workflows: line drawings, regular screentones, and irregular screen textures. The generated layers can be composed into manga images and further edited, supported by a large artist-annotated dataset.
Abstract: The paper introduces a controllable image cartoonization framework by decomposing cartoons into surface, structure, and texture representations. Separate learning objectives for each representation make the model adjustable across cartoon styles and artist requirements.