Exploring Diffusion Transformer Designs via Grafting

NeurIPSOral2025

Authors: Keshigeyan Chandrasegaran, Michael Poli, Daniel Y Fu, Dongjun Kim, Lea M. Hadzic, Manling Li, Agrim Gupta, Stefano Massaroli, Azalia Mirhoseini, Juan Carlos Niebles, Stefano Ermon, Li Fei-Fei
Affiliation: Stanford University
Venue: NeurIPS 2025
Track: Oral

TL;DR

We propose grafting, a simple approach to materialize new architectures by editing pretrained diffusion transformers. It enables architectural exploration under small compute budgets.

Opening excerpt from the authors’ abstract. source

Read the paper

OpenReview

Topics

exploration transformer diffusion

← All NeurIPS 2025 Oral papers · Browse the whole archive