My name is David Li and I am a first-year PhD student at MBZUAI. My research interests are mainly in generative models, diffusion models, optimal transport, and related areas of machine learning. I am especially interested in making diffusion models faster through distillation approaches.
I created LiSearch for people who want to understand generative model papers through discussion. I share short paper notes, highlight key ideas, and invite researchers, students, and enthusiasts to continue the conversation in Telegram comments. Session recordings are published on YouTube, and I will be glad to see you there.
Highlights
Framework that distills diffusion language models into few-step generators while preserving entropy and generative perplexity.
ICML 2025A distillation objective for diffusion bridge models that accelerates image-to-image generation from hundreds of teacher steps to a few student steps.
ICLR 2026 (Oral)A universal inverse distillation framework that incorporates real data into matching-model distillation without GAN training.
Publications
* equal contribution
Studies zero-shot uncertainty signals for large language diffusion models by using information from iterative denoising.
Extends inverse distillation to discrete token spaces for diffusion language models, targeting 4x to 64x fewer inference steps.
Distills ResShift-style super-resolution diffusion into a single-step restoration model with competitive perceptual quality and lower memory cost.
Introduces RealUID, a universal distillation framework for matching models that uses real-data supervision without adversarial training.
Connects iterative Markovian fitting with iterative proportional fitting for solving Schrodinger bridge problems.
Workshop version of the IDLM project, presenting inverse distillation for faster diffusion language model sampling.
Workshop submission presenting the same inverse-distilled diffusion language model direction.
Distills diffusion bridge models for conditional and unconditional image-to-image translation, accelerating generation by 4x to 100x in the reported setups.
Uses anatomically guided diffusion pretraining to improve 3D medical semantic segmentation under limited labeled data.