Meta-Diff for Few-Shot Diffusion via Task-Space Transport
October 23, 2026
Meta-Diff enables rapid adaptation of diffusion models to data-scarce target domains by learning source-target relations in a compact task space. A shared encoder and diffusion backbone allow for generation from predicted target coordinates without requiring target-specific retraining.
HOW THIS AFFECTS YOU
●
builderYou can deploy high-capacity generators in data-constrained environments like healthcare without expensive per-target fine-tuning.
●
researcherThis method provides a way to generalize across tasks by learning a recurring distribution relation.