Generative AI / Visual ComputingMPhil · ShanghaiTech
I'm an MPhil student at ShanghaiTech University working on generative models for graphic design—jointly modeling imagery, layout, typography, and semantics. I also explore 2D/3D and vector graphics, world models, and diffusion- and flow-based generation.
Selected research / 研究
Three projects tracing a path from joint image editing and layout generation, through explicit interaction between image and layout priors, to full structured text-to-design generation with multimodal pretrained priors.
01Under review · IEEE TMM · 2026
CELG jointly edits a background image and generates its foreground layout, allowing visual content and spatial composition to co-adapt to a design intention.
Its Search-Compare-Edit data pipeline and coupled flow-matching models turn content-aware layout generation into an iterative editing problem.
Manuscript under review; no public preprint or code release yet.
02Under review · IEEE TVCG · 2026
InterIL makes the interaction between pretrained image and layout priors explicit through a learnable bidirectional communication module.
The joint diffusion process generates both modalities together while preserving the strengths of each frozen backbone.
Public arXiv preprint; manuscript under review at IEEE TVCG.
03Under review · ACM TOG · 2026
PriDe extends structured generation from image-layout interaction to full text-to-design synthesis, jointly modeling imagery, layout, typography, and editable structure.
It reuses multimodal priors from frozen text-to-image and language models to learn from only 10K layered designs.
Manuscript under review; no public preprint or code release yet.
Research outlook
The point where a model stops being a black box and becomes a material you can work with.
Broader directions I am actively exploring beyond the publication record above.
Diffusion, flow matching, and learned generative processes for controllable visual synthesis.
Models that move between images, assets, scenes, and spatial representations.
Editable layers, layout, typography, imagery, and semantics generated as one coherent artifact.
Scalable geometry and native graphic representations that remain directly editable.
Generative representations that capture the structure and dynamics of interactive environments.
Research at the boundary of computer vision, graphics, design systems, and interaction.
A working philosophy / 工作哲学
Understand the problem space. Understand your toolkit. Make something, learn from contact with reality, and let the direction evolve.
理解问题，也理解自己的工具箱。保持认真，但不要僵化。
Field notes / writing
Small observations from work in progress. Useful enough to keep; unfinished enough to revisit.
About this place
I'm an MPhil student at ShanghaiTech University working on generative models for graphic design and structured visual generation. My current projects study how imagery, spatial layout, typography, and semantic intent can be generated and edited together.
My broader interests include 2D/3D generation, vector and SVG graphics, world models, and diffusion- and flow-based generative modeling. Outside the lab, I like drawing and building interactive systems with Unity, Unreal, and Bevy.
Have a thought to share?
Ideas, collaborations, strange problems, or a paper I should read.
yangshr2024@shanghaitech.edu.cn my inbox is open :)
