Yuta Oshima

Iā€™m a second-year Ph.D. student at the University of Tokyo, advised by Professor Yutaka Matsuo. I am also a research intern at Sakana AI and AIRoA.

My Ph.D. research aims to unlock the potential of image and video generation AI and expand the domains where it can be used. Just as LLMs are now indispensable across countless domains, image and video generation should follow the same path. To this end, I am currently working on š—±š—¶š—³š—³š˜‚š˜€š—¶š—¼š—» š—ŗš—¼š—±š—²š—¹š˜€ for video generation and world models, and š—ŗš˜‚š—¹š˜š—¶š—ŗš—¼š—±š—®š—¹ š—”š—œ š—®š—“š—²š—»š˜š˜€ for image generation.

selected publications

  1. Preprint
    autoref.jpg
    AutoRef: Harness Optimization for Agentic Multi-Reference Image Generation
    Yuta Oshima, Ku Onoda, Yusuke Iwasawa, Masahiro Suzuki, Yutaka Matsuo, and Hiroki Furuta
    2026
    arXiv Code
  2. CVPR 2026 Main
    multibanana_teaser.png
    MultiBanana: A Challenging Benchmark for Multi-Reference Text-to-Image Generation
    Yuta Oshima, Daiki Miyake, Kohsei Matsutani, Yusuke Iwasawa, Masahiro Suzuki, Yutaka Matsuo, and Hiroki Furuta
    In the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026
    arXiv Code Hugging Face
  3. TMLR 2026
    worldpack.png
    WorldPack: Dynamic Frame Compression for Long-context Video World Modeling
    Yuta Oshima, Yusuke Iwasawa, Masahiro Suzuki, Yutaka Matsuo, and Hiroki Furuta
    In Transactions on Machine Learning Research, 2026
    arXiv
  4. NeurIPS 2025
    dlbs_teaser.gif
    Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search
    Yuta Oshima, Masahiro Suzuki, Yutaka Matsuo, and Hiroki Furuta
    In Neural Information Processing Systems (NeurIPS), 2025
    arXiv Code Website