Large generative models have reached a level of quality that would have seemed unreasonable just a few years ago. Yet most of them are still impractical to deploy: too slow for interactive use, too memory-hungry for edge devices, or misaligned after the initial training run.
My research starts from a simple observation: the bottleneck is rarely the model architecture itself. More often, it lives in the surrounding pipeline — the sampler, the decoder, the alignment procedure — and these are places where targeted algorithmic work can have outsized impact without requiring a full retraining cycle.
This is the thread that connects my work on Diffusion Models, Flow Matching, and Reinforcement Learning. Each project asks the same question in a different context: where exactly does the pipeline break down, and what is the lightest intervention that fixes it?
I plan to use this blog to share ideas, paper notes, and lessons from ongoing research. If something here is useful to you, feel free to reach out.