DeCoPrune cuts video diffusion memory tokens by 85% with training-free pruning
XingangP · x · 2026-10-12
DeCoPrune is a training-free token pruning method for autoregressive video diffusion models.
- Key insight: denoising difficulty reveals what's worth remembering — keep tokens that are hard to predict from context, prune the redundant ones, with no extra forward passes.
- Results on LingBot World v2: 85.43% fewer historical KV tokens, 4.14× faster continuation generation, and long-range visual recall close to FullKV.
- The team also releases CMBench, a benchmark for long-range visual memory in video generation.
Paper, code, and dataset are all publicly available.
Related event: DeCoPrune: Training-Free Pruning Speeds Video Generation 4x(2 posts)→
More from Multimodal
- Scale AI CEO Alexandr Wang: Muse Wonder video quality feels on par with Opus 5.5 — alexandr_wang · 2026-10-13
- Thrixel, founded by UChicago prof behind MeshCNN, generates part-based 3D assets for AI agents — RanaHanocka · 2026-10-13
- BudgetPix: adaptive tokenization matches full-compute pixel diffusion with 25% of the compute — _akhaliq · 2026-10-13
- Muse Kitchen launches under Alexandr Wang's Muse, with research blog and subscription access — ren_hongyu · 2026-10-13
- Meta Muse Wonder: pricing, benchmarks and limits rounded up — mawcopolow · 2026-10-13
- 'He Might Lose His Memory Tomorrow': a moving AI-generated anime short — finalbosu · 2026-10-13