DyWeight: Dynamic Gradient Weighting for Few-Step Diffusion Sampling
A learnable few-step ODE solver that predicts time-varying weights to adaptively combine historical gradients, staying accurate at very large step sizes.
Ph.D. Student, Zhejiang University & Westlake University
I am a first-year Ph.D. student at Zhejiang University and Westlake University, advised by Prof. Chi Zhang in the Westlake AGI Lab. My research focuses on efficient diffusion generation — making diffusion and flow models faster to sample while preserving generation quality. Before that, I received my B.Eng. from the University of Electronic Science and Technology of China, where I was part of the Yingcai Honors College and worked with Prof. Yong Deng on uncertainty and information theory.
I grew up in Qujing, Yunnan — a laid-back city in southwest China blessed with a mild, spring-like climate all year round. If you visit, do try our local comfort food: xiaoguo erkuai (small-pot rice cake) and zheng ersi (steamed rice noodles). 🍜
Off the clock, I play a lot of basketball 🏀 and I'm an avid gamer 🎮 — CS2 (Perfect World rank A) and Honor of Kings (3rd place in UESTC's e-sports tournament).
* denotes equal contribution.
A learnable few-step ODE solver that predicts time-varying weights to adaptively combine historical gradients, staying accurate at very large step sizes.
Learns instance-aware time discretizations: a lightweight network predicts a tailored set of sampling timesteps for each input, instead of one fixed schedule.
Fixes the compute budget in advance and searches for the step-level cache policy that best preserves output quality — controllable acceleration up to 3.7× on FLUX.1-dev.
A learnable high-order ODE solver that evaluates several gradients in parallel per step and combines them with distilled weights, cutting truncation error at very few function evaluations.
A training-free framework that repurposes pretrained video diffusion models for controllable 3D and 4D generation via inference-time camera-trajectory guidance.
Improves diffusion generalization and sample quality with a weak-to-strong segmented guidance scheme that hybridizes classifier-free guidance and autoguidance.
A training-free, geometry-aware caching method that reuses stable intermediate features to accelerate 3D shape-generation diffusion (~27% faster) without corrupting geometry.
Characterizes when Deng entropy grows linearly with the scale of the frame of discernment, and shows the slope is exactly the information fractal dimension of the mass function.
Defines the information fractal dimension for the permutation mass function of a Random Permutation Set, and shows the dimension at maximum RPS entropy is 2.
A data-synthesis framework for real-world multi-turn function calling — environment–API graph interactions, advanced tool/query synthesis, and a guided iterative chain — that brings a 4B model to SOTA among similar-sized models on BFCLv3.
HardGen builds a dynamic API graph from agent failure cases to synthesize genuinely hard tool-use traces and refine verifiable chain-of-thought — letting a 4B model rival much larger systems on tool use.