Generative models: efficiency and diffusion-specific alignment
Key Questions
What is FlowLong and how does it support long video generation?
FlowLong enables training-free long video generation via manifold-constrained Tweedie matching at inference time. It can also extend to audio and 3D content generation.
How does DAR improve diffusion transformers?
DAR uses timestep-adaptive residual routing to improve diffusion transformers by 2.11 FID while requiring 8.75× fewer training iterations.
What efficiency gains are reported for recent generative models?
ELIT achieves roughly 2.7× faster sampling, while new methods like CBDiffuse, DSPO, FlowLong, and DAR target diffusion-specific alignment and efficiency.
ELIT (~2.7x sampling), CBDiffuse, DSPO. New: FlowLong enables training-free long video generation via Tweedie matching, extends to audio/3D. New: DAR (timestep-adaptive residual routing) improves diffusion transformers by 2.11 FID and 8.75× fewer training iterations.
Sources (2)
Updated May 25, 2026