Research map / Video / Video generation
Training-Free Efficient Video Generation via Dynamic Token Carving
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Jenga is a training-free DiT video-generation inference pipeline combining block-wise dynamic attention carving (3D space-filling-curve token blocks with top-k and probability-based selection) with progressive multi-resolution denoising and a text-attention amplifier, achieving up to 8.83x speedup on HunyuanVideo with negligible VBench loss.
Problem
Video diffusion transformers are too slow for practical use due to quadratic self-attention cost over very long token sequences and the multi-step nature of diffusion sampling.
Approach
Combine dynamic block-sparse attention over space-filling-curve-ordered 3D latent blocks (attention carving) with a progressive resolution schedule that denoises early steps at low resolution, plus a text-attention bias to keep field of view stable at low resolution; all applied without retraining.
Key results
- Jenga-Flash achieves 8.83x DiT speedup on HunyuanVideo (1625s to 184s) with VBench 82.73% vs 82.74% baseline.
- Jenga on HunyuanVideo-I2V gives 4.43x speedup (1499s to 338s) with VBench 87.75% vs 87.49%.
- Jenga on Wan2.1-1.3B gives 6.52x speedup (115s to 17s) with VBench drop of 0.76 points (83.28% to 82.52%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine dynamic attention carving, progressive resolution, and timestep skipping | HunyuanVideo · HunyuanVideo | DiT time: 1625s -> 184s | Helped |
| use Jenga-Flash instead of the baseline | HunyuanVideo · HunyuanVideo | VBench: 82.74% -> 82.73% | No clear effect |
| replace full attention with dynamic block-wise Attention Carving | HunyuanVideo · HunyuanVideo | DiT time: 1625s -> 748s | Helped |
| remove the condition mask (ablation) | Jenga · HunyuanVideo | VBench: 83.07% -> 81.82% | Hurt |
| remove the adjacency mask (ablation) | Jenga · HunyuanVideo | VBench: 83.07% -> 82.42% | Hurt |
| remove the importance mask (ablation) | Jenga · HunyuanVideo | VBench: 83.07% -> 77.41% | Hurt |
| replace SFC with hwt linear partitioning (ablation) | Jenga · HunyuanVideo | VBench: 83.07% -> 82.82% | Hurt |
| increase the resolution schedule from 2 stages to 3 stages (ablation) | Jenga · HunyuanVideo | VBench: 83.07% -> 80.53% | Hurt |
| increase the low-resolution step ratio from 50% to 90% (ablation) | Jenga · HunyuanVideo | VBench: 83.07% -> 78.74% | Hurt |
| increase the cutoff probability from 0.3 to 0.9 (ablation) | Jenga · HunyuanVideo | DiT time: 225s -> 330s | Hurt |
| set text-amplifier bias factor rho to 1.5 rather than 0.5 (ablation) | Jenga · HunyuanVideo | VBench: 83.07% -> 82.80% | Hurt |
| apply Jenga without ProRes | HunyuanVideo-I2V · HunyuanVideo-I2V | VBench: 87.49% -> 87.75% | Helped |
Surprising findings
- Jenga-Base and Jenga-Turbo score higher than the HunyuanVideo baseline on VBench (e.g., semantic score 72.84% to 77.48%).
- Removing the importance mask (keeping only condition and adjacency masks) causes a large drop in VBench (77.41% vs 83.07%).
Limitations
- Latent resizing during stage transitions can produce boundary artifacts; quality degrades in static scenes or with clear boundaries.
- Block partitioning is static SFC and not semantically adaptive.
- Wan2.1 and I2V results require per-model parameter tuning (e.g., cutoff p=0.9 for Wan2.1).
Experimental setup
Datasets: VBench VBench-I2V Inter4K Sora prompts
Benchmarks: VBench VBench-I2V
Metrics: VBench total VBench-Q VBench-S CLIP-score DiT latency speedup PFLOPs FVD user preference win rate
Baselines: CLEAR MInference SVG TeaCache Bottleneck Sampling STA SpargeAttn AdaSpa
Builds on: MInference MoBA FlashAttention TeaCache Bottleneck Sampling
Closest papers by meaning
- MonarchRT: Efficient Attention for Real-Time Video Generation (NeurIPS 2026 · similarity 0.90)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.89)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.89)
- Towards Precise Scaling Laws for Video Diffusion Transformers (CVPR 2025 · similarity 0.89)
- Parameterized Stripe Attention for Efficient Video Generation (NeurIPS 2026 · similarity 0.89)
- Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis (CVPR 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.