Highlight

Research

My published research works revolve around the fields of Visual Generation (Image Inpainting, Multi-Frame Generation, Video Generation), Visual Perception (Video Object Detection), and AI for Science. Representative papers are highlighted. There are also some unpublished papers related to representation learning and memory modeling of visual features, so stay tuned. Currently, my research interests focus on generative AI. From a long-term and idealistic perspective, I'm very interested in Artificial General Intelligence.

Through retrospective analyses and iterative system upgrades, we developed GenDSA-v2, which is, to the best of our knowledge, the largest and most comprehensive low-dose imaging system to date. Leveraging a multi-center cohort of 50,000 patients from 86 hospitals worldwide, collected in collaboration with major international vendors (GE, Philips, and Siemens), we demonstrate that GenDSA-v2 achieves a threefold reduction in radiation dose without compromising diagnostic or interventional performance. Its clinical effectiveness and safety were further validated through animal studies and prospective evaluations, including cross-over observational trials and randomized controlled studies, confirming that GenDSA-v2 maintains sufficient image fidelity to visualize lesions as small as 1 mm.

Genesis is a unified framework for joint generation of multi-view driving videos and LiDAR sequences with spatio-temporal and cross-modal consistency. Genesis employs a two-stage architecture that integrates a DiT-based video diffusion model with 3D-VAE encoding, and a BEV-aware LiDAR generator with NeRF-based rendering and adaptive sampling. Both modalities are directly coupled through a shared latent space, enabling coherent evolution across visual and geometric domains. Extensive experiments on the nuScenes benchmark demonstrate that Genesis achieves SOTA performance across video and LiDAR metrics (FVD 16.95, FID 4.24, Chamfer 0.611), and benefits downstream tasks including segmentation and 3D detection, validating the semantic fidelity and practical utility of the generated data.

PixelHacker is a purely visual and general inpainting framework. Its core lies in a new paradigm named Latent Categories Guidance (LCG), which introduces category-level latent semantic distribution to guide latent space for structural and semantic consistent denoising. PixelHacker achieves 0.86B parameters—only 7% and 11% of Flux.1-Fill-dev and SD3.5 Large-Inpainting—running at 47 ms/step (512×512, single L40S GPU), over 3× faster than these baselines while maintaining SOTA performance across seven benchmarks on both natural and portrait scenes. Representatively, PixelHacker achieves FID 0.82 / LPIPS 0.088 on Places2 with 13% and 11% improvements, and FID 6.35 / LPIPS 0.229 on FFHQ with 43% and 15% improvements. This work establishes a principled and interpretable foundation for efficient, consistent, and general-purpose image inpainting.

Compared to our last job MoSt-DSA, GaraMoSt adds multi-granularity motion and structural feature modeling and modifies the overall Pipeline into a highly parallel design, which greatly improves the accuracy and reduces high-frequency and low-frequency noise under the same inference time level (for interpolating 3 frames, only increasing by 0.005s). Comprehensive beyond the SOTA natural scene, and DSA scene methods.

XS-VID dataset comprises aerial data from various periods and scenes, and extensively collects three types of objects with smaller pixel areas: extremely small (0-12^2), relatively small (12^2-20^2), and generally small (20^2-32^2). XS-VID offers unprecedented breadth and depth in covering and quantifying minuscule objects, significantly enriching the scene and object diversity in the dataset. YOLOFT enhances local feature associations and integrates temporal motion features, significantly improving the accuracy and stability of Small Video Object Detection.

GenDSA is a large-scale pretrained multi-frame generative model-based real-time and low-dose DSA imaging system, which pre-trained, fine-tuned and tested on ten million of images from 35 hospitals. Suitable for most DSA scanning protocols, GenDSA could reduce the DSA frame rate (i.e., radiation dose) to 1/3 and generates video that was virtually identical to clinically available protocols. Videos generated by GenDSA reach a comparable level to the full-sampled videos, both in terms of overall quality (4.905 vs 4.935) and lesion assessment (4.825 vs 4.860), which fully demonstrated the potential of GenDSA for clinical applications.

MoSt-DSA is the first work that uses deep learning for DSA frame interpolation, comprehensively achieving SOTA in accuracy, speed, visual effect, and memory usage. Meanwhile, MoSt-DSA is also the first method that directly achieves any number of interpolations at any time steps with just one forward pass during both training and testing. If applied clinically, MoSt-DSA can significantly reduce the DSA radiation dose received by doctors and patients when applied clinically, lowering it by 50%, 67%, and 75% when interpolating 1 to 3 frames, respectively.

Internship

Research Intern, VIVO AI Lab

2024.10~2025.08, Hangzhou, China

Research on end-side multi-modal content understanding and generation