Li-Xuan Alex Peng

Li-Xuan Alex Peng

MS student · Vision Science Lab, NTHU · Embodied AI Lab, NTU

I am a master's student at National Tsing Hua University, advised by Prof. Min Sun, and I work with Prof. Tsung-Wei Ke at National Taiwan University. Before that I was a research assistant at CITI, Academia Sinica, with Dr. Jun-Cheng Chen, and received a B.S. in Electrical Engineering and Computer Science from National Central University.

I am seeking Visiting Student / PhD opportunities for 2027!

Research

I study how machines infer physical dynamics from what they see. Instead of training end-to-end predictors, I decompose visual representations along the causal structure of the physical process, so that geometry, material and dynamics can each be checked and adapted. My recent work, DELPHI and ADYTON establish a way to learn 4D deformable dynamics from RGB-D videos. We aim to build a general 4D Physics World Model for robotic decision-making.

DELPHI · single-step full-geometry prediction, direct-hit collisionpaused

News

Publications

4D Physics Dynamics From Vision

Beyond the Visible: Learning Dense 4D Motion in Contact-Rich Deformable Worlds

Li-Xuan Peng, Po Yi Wu*, Yu-Tsen Tsai*, Tsung-Wei Ke, Min Sun

Preprint 2026 · first author · corresponding author

A vision-conditioned Eulerian displacement field that predicts dense 4D motion everywhere: visible surface, occluded interior, and ambient off-object space.

project

@inproceedings{peng2026delphi,
  title     = {Beyond the Visible: Learning Dense 4D Motion in
               Contact-Rich Deformable Worlds},
  author    = {Peng, Li-Xuan and Wu, Po Yi and Tsai, Yu-Tsen and Ke, Tsung-Wei and Sun, Min},
  booktitle = {arXiv},
  year      = {2026}
}

Generative Diffusion Models

VLM as an Editing Pilot: VLM-Steered Velocity Control for Flow-Based Image Editing

Kuan-Shen Chiu*, Jia-Wei Liao*, Yu-Chieh Chao, Li-Xuan Peng, Wei-Yun Ma, Cheng-Fu Chou

WACV 2027

A training-free framework in which a VLM picks among coupling-controlled velocities at each step of flow-based editing, improving instruction alignment over FlowEdit on MagicBrush and PIE-Bench with FLUX and SD3.

paperarXivcode

@misc{chiu2026v2cedit,
  title         = {VLM as an Editing Pilot: VLM-Steered Velocity Control
                   for Flow-Based Image Editing},
  author        = {Chiu, Kuan-Shen and Liao, Jia-Wei and Chao, Yu-Chieh
                   and Peng, Li-Xuan and Ma, Wei-Yun and Chou, Cheng-Fu},
  year          = {2026},
  eprint        = {2611.10066},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

DiffRGD: An Inference-Time Diffusion Guidance Through Riemannian Gradient Descent

Jia-Wei Liao, Li-Xuan Peng, Mei-Heng Yueh, Min Sun, Cheng-Fu Chou, Jun-Cheng Chen

ECCV 2026

Inference-time guidance that keeps diffusion latents on the sphere induced by their Gaussian distribution, solved by Riemannian gradient descent; plugs into any pre-trained diffusion model.

paperarXivprojectcode

@inproceedings{liao2026diffrgd,
  title     = {DiffRGD: An Inference-Time Diffusion Guidance
               Through Riemannian Gradient Descent},
  author    = {Liao, Jia-Wei and Peng, Li-Xuan and Yueh, Mei-Heng
               and Sun, Min and Chou, Cheng-Fu and Chen, Jun-Cheng},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026},
}

Zero-shot Geometry-Aware Diffusion Guidance for Music Restoration

Jia-Wei Liao, Pin-Chi Pan, Li-Xuan Peng, Sheng-Ping Yang, Yen-Tung Yeh, Cheng-Fu Chou, Yi-Hsuan Yang

NeurIPS 2025 Workshop on AI for Music

Diffusion Geodesic Guidance moves noisy samples along geodesics with a closed-form spherical interpolation, improving on training-free baselines across four music restoration tasks.

paper

Pixel Is Not A Barrier: An Effective Evasion Attack for Pixel-Domain Diffusion Models

Chun-Yen Shih*, Li-Xuan Peng*, Jia-Wei Liao, Ernie Chu, Cheng-Fu Chou, Jun-Cheng Chen

AAAI 2025 · co-first author

An evasion attack that protects images from pixel-domain diffusion editing such as SDEdit, and extends to latent diffusion models.

paperarXiv

@misc{shih2024pixelbarriereffectiveevasion,
  title={Pixel Is Not A Barrier: An Effective Evasion Attack for Pixel-Domain Diffusion Models},
  author={Chun-Yen Shih and Li-Xuan Peng and Jia-Wei Liao and Ernie Chu and Cheng-Fu Chou and Jun-Cheng Chen},
  year={2024},
  eprint={2408.11810},
  archivePrefix={arXiv}
}

DiffQRCoder: Diffusion-based Aesthetic QR Code Generation with Scanning Robustness Guided Iterative Refinement

Jia-Wei Liao, Li-Xuan Peng*, Winston Wang*, Tzu-Sian Wang*, Ju-Hsuan Weng, Cheng-Fu Chou, Jun-Cheng Chen

WACV 2025

Raises QR scan success from 60% with ControlNet alone to 99%.

paperarXivproject

@misc{liao2024diffusionbased,
  title={DiffQRCoder: Diffusion-based Aesthetic QR Code Generation with Scanning Robustness Guided Iterative Refinement},
  author={Jia-Wei Liao and Li-Xuan Peng and Winston Wang and Tzu-Sian Wang and Ju-Hsuan Weng and Cheng-Fu Chou and Jun-Cheng Chen},
  year={2025},
  eprint={2409.06355},
  archivePrefix={arXiv}
}

* Equal contribution.