Cheng Lu
Research Scientist, MSL TBD Lab at Meta
San Francisco, California, United States
Summary
🤩
Rockstar🎓
Top SchoolCheng Lu is a research scientist at Meta's MSL TBD Lab in San Francisco with eight years of machine learning research and engineering experience and a PhD in Computer Science from Tsinghua University. Previously a Member of Technical Staff at OpenAI, he is an active open-source contributor whose work includes significant contributions to Hugging Face’s widely used diffusers library and maintaining the DPM-Solver codebase (NeurIPS 2022 oral). His technical focus is on fast, numerically stable solvers for diffusion models—implementing multistep and singlestep DPM-Solver variants, SDE adaptations, cosine noise schedules, and SDXL stability fixes within stable-diffusion pipelines. Cheng combines deep theoretical knowledge with production-minded engineering, routinely turning advanced ODE/SDE samplers into practical, high-performance components for generative models.
8 years of coding experience
2 years of employment as a software developer
Doctor of Philosophy - PhD, Computer Science, Doctor of Philosophy - PhD, Computer Science at Tsinghua University
Github Skills (18)
algorithm10
algorithms10
pytorch10
diffusion-models10
machine-learning10
stable-diffusion10
diffusers10
diffusion-probabilistic-models10
diffusion-probabilistic10
deep-learning10
datastructures9
datastructure9
data-structure9
python9
data-structures9
Programming languages (12)
Github contributions (5)
LuChengTHU/dpm-solver
Aug 2022 - Dec 2022
Official code for "DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps" (Neurips 2022 Oral)
Role in this project:
ML Engineer
ML EngineerContributions:6 reviews, 61 commits, 3 PRs in 4 months
Contributions summary:Cheng's commits primarily focus on the implementation and refinement of the DPM-Solver, a fast ODE solver for diffusion probabilistic models. Their work involves defining noise schedules, model wrappers, and the core DPM-Solver algorithm. The changes include modifications to existing code, bug fixes, and the addition of new functionalities, indicating an active role in the development and improvement of the solver. The user also demonstrates experience with model types, guidance, and sampling modes.
ode-solverneuripssolverdpmdiffusion-probabilistic
huggingface/diffusers
Nov 2022 - Dec 2022
🤗 Diffusers: State-of-the-art diffusion models for image, video, and audio generation in PyTorch and FLAX.
Role in this project:ML Engineer
ML EngineerContributions:28 reviews, 2 commits, 5 PRs in 1 month
Contributions summary:Cheng implemented and refined DPM-Solver schedulers, specifically focusing on multistep and singlestep variations for diffusion models. Their work involved integrating DPM-Solver into the stable-diffusion pipeline, addressing cosine schedules for models such as DeepFloyd-IF, and incorporating SDE (Stochastic Differential Equation) variants. This also included the addition of new features and fixing issues to optimize existing functionalities with focus on numerical stability improvements, particularly for SDXL.
pytorchartdeep-learningimage2imagestate-of-the-art
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.