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.
Learnable sparsity for vision generative models
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.