Research map / Video / Video generation
S²Flow: Towards Fast and Authentic Training-Free High-Resolution Video Generation
Chaoqun Wang, Shaobo Min, Xu Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LR-guided adaptive windows instead of fixed uniform windows at 50% sparsity (ablation) | LRG-WinAttn · VBench window sparsity ablation | generation success: Fail to Generate -> 0.7389 | Improved |
| use 50% LR-guided window instead of full attention (ablation) | S2Flow · VBench window sparsity ablation | mean score / speedup: 0.7416 -> 0.7389, 1x -> 4x | Mixed |
| reduce window sparsity from 50% to 25% (ablation) | LRG-WinAttn · VBench window sparsity ablation | mean score: 0.7389 -> 0.6797 | Hurt |
| increase window sparsity from 50% to 75% (ablation) | LRG-WinAttn · VBench window sparsity ablation | mean score: 0.7389 -> 0.7404, 4x -> 1.8x | No clear effect |
| use second-order instead of first-order solver (ablation) | LRI-HiSolver · VBench solver order ablation | mean score: 0.7319 -> 0.7389 | Improved |
| use third-order instead of second-order solver (ablation) | LRI-HiSolver · VBench solver order ablation | mean score: 0.7389 -> 0.7390 | No clear effect |
| upsample at step 30 instead of step 40 (ablation) | S2Flow · VBench upsampling step ablation | mean score: 0.6080 -> 0.7389 | Improved |
| upsample at step 30 instead of step 20 (ablation) | S2Flow · VBench upsampling step ablation | mean score: 0.7332 -> 0.7389 | Improved |
| upsample at step 30 instead of step 10 (ablation) | S2Flow · VBench upsampling step ablation | mean score: 0.6939 -> 0.7389 | Improved |
| apply S2Flow to HunyuanVideo at 2560x1536 | HunyuanVideo · VBench (473 prompts, 2560x1536) | inference time / mean score: 61 -> 15 minutes, 0.6963 -> 0.7389 | Improved |
| S2Flow vs MACS | S2Flow · VBench (473 prompts, 2560x1536) | mean score: 0.7366 -> 0.7389 | Improved |
| S2Flow transformer-level speed at 2560x1536 | S2Flow · Runtime per 10 denoising steps | seconds: 733.0s -> 96.4s | Improved |
Closest papers by meaning
- HiFlow: Training-free High-Resolution Image Generation with Flow-Aligned Guidance (NeurIPS 2025 · similarity 0.88)
- Boosting Flow-based Generative Super-Resolution Models via Learned Prior (CVPR 2024 · similarity 0.87)
- FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video Synthesis (CVPR 2024 · similarity 0.87)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.86)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.86)
- StreamFlow: Streamlined Multi-Frame Optical Flow Estimation for Video Sequences (NeurIPS 2024 · similarity 0.86)
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.
