A 5B-parameter video diffusion transformer is made deployable on mobile hardware via recurrence distillation, learnable head pruning, step distillation, and decoder optimization, achieving 83.79 VBench at 20s latency.
Sana-sprint: One-step diffusion with continuous-time consistency distillation
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
MobileWan: Closing the Quality Gap for Mobile Video Diffusion
A 5B-parameter video diffusion transformer is made deployable on mobile hardware via recurrence distillation, learnable head pruning, step distillation, and decoder optimization, achieving 83.79 VBench at 20s latency.