I like taking apart big, intimidating models until I understand exactly why they work — then putting the smallest runnable version back together so other people can too.

These days that means embodied AI: vision-language-action policies, diffusion / flow matching, and 3D vision. I'm a direct-PhD student at Peking University, but most of what I actually do lives in the repos below.

A recurring theme in my work: complex papers shouldn't stay locked inside their original codebases.

• 🤖 pi-zero-minimal — Physical Intelligence's π0, stripped down to a minimal runnable VLA. No engineering scaffolding, just the core idea you can read in one sitting.
• 🧩 CV_Milestones — clean-room re-implementations of landmark papers (DiT, 3DGS, …) — the version I wish existed when I was first reading them.

And when re-implementing isn't enough, I write things down:

• 🌊 Flow-Diffusion — diffusion, score matching, and SDEs, all derived from one flow-matching lens. The unified picture I wanted but couldn't find.
• 📐 3D_Vision — camera models → epipolar geometry → SfM, built from the ground up.

On the research side, I've spent time on:

• a diffusion foundation model unifying five image-restoration tasks — and rebuilding its sampler with flow matching for a ~30× speedup (co-first author, Nature Communications)
• using a VLM as a reward signal to make few-step text-to-image models actually follow instructions, trained end-to-end and differentiable
• feed-forward 3D reconstruction with Transformers, self-supervised where ground truth doesn't exist

I'm always up for a good conversation about generative models, embodied agents, or why your sampler is slow.

📫 jzx417889065@stu.pku.edu.cn