Hong-Phuc Lai

I'm a Research Engineer at VinMotion, working on Vision-Language-Action and World Action Models. Previously, I was a Predoctoral Research Resident at VinAI Research – Qualcomm AI Research in Hanoi, Vietnam, supervised by Principal Scientist Dr. Anh Tran, where I focused on making large generative models efficient and practical, with interests spanning Diffusion Models, Flow Matching, and Reinforcement Learning. I received my B.Sc. in Automation Control Engineering from Hanoi University of Science and Technology (HUST).

Email  /  CV  /  LinkedIn  /  GitHub  /  Blog

Research

I currently work on Vision-Language-Action and World Action Models. Previously, I focused on making generative models efficient, robust, and practical under real-world constraints, at the intersection of Diffusion Models, Flow Matching, and Reinforcement Learning — with a recurring focus on identifying the true bottleneck in a generative pipeline and designing lightweight algorithms to address it. Some papers are highlighted.

Addresses the VAE decoding bottleneck in high-resolution image generation, enabling further speed improvements beyond PixelRush. It speeds up megapixel synthesis beyond the denoiser, delivering ~2× faster 4K generation with better image quality than a state-of-the-art pipeline.

An ultra-fast, training-free method for high-resolution image generation via one-step diffusion, achieving state-of-the-art quality at up to 30× faster speed than prior work.

A hybrid transformer combining pyramid spatial pooling with spatial attention for accurate medical image segmentation.

Preprints

Next3DGen is a unified post-training framework for text- and image-conditioned 3D generation, spanning single-stage and multi-stage models. It improves quality, condition alignment, and 3D consistency over pretrained models and substantially outperforms existing post-training methods while remaining training-efficient.

Black-box reward post-training for already-distilled few-step flow models. SwiftAlign adapts a deployed sampler to new rewards without reward gradients or a new distillation run, while keeping the deterministic ODE path used at inference.

Writing

Short notes on research. Full list on the blog.

Post-Training After the Model Is Already Fast
September 28, 2026
Why alignment still matters once a flow model has already been distilled to a few steps.

The Bottleneck Moves Past the Denoiser
September 14, 2026
A note on where high-resolution generation actually spends its time, from PixelRush to UltraFlash.

Template from Jon Barron's website.