Research map / Video / Video generation
SWIFT: Sliding Window Reconstruction for Few-Shot Training-Free Generated Video Attribution
Chao Wang, Zijin Yang, Yaofei Wang, Yuang Qi, Weiming Zhang, Nenghai Yu, Kejiang Chen
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 temporal sliding-window reconstruction (SWIFT) instead of AEDR | video attribution · S-Video avg | accuracy: 73.6% -> 94.0% | Improved |
| use SWIFT instead of AEDR on LTX target | LTX-Video attribution vs EA / Real · S-Video LTX | accuracy: 97.1% -> 94.5% (EA); 95.6% -> 92.2% (Real) | Mixed |
| use MSE instead of PSNR/SSIM as loss metric (ablation) | SWIFT · Wan2.1 | accuracy: 47.8% (PSNR) -> 98.4% (MSE) | Improved |
| use MSE instead of MAE (ablation) | SWIFT · Wan2.1 | accuracy: 97.8% -> 98.4% | Improved |
| use W0/W3 corrupted window instead of W0/W1 (ablation) | SWIFT on HYV · HYV window selection | accuracy: 82.3% -> 90.7% | Improved |
| use W0/W3 instead of intuitive W0/W7 (ablation) | SWIFT on LTX · LTX window selection | accuracy: 78.6% -> 85.3% | Improved |
| increase threshold sample size from 0 (tau=1) to 200 (ablation) | SWIFT KDE threshold · avg over 5 models | avg accuracy: 85.1% -> 94.0% | Improved |
| fixed threshold tau=1 (zero-shot) (ablation) | SWIFT · per model | accuracy: HYV 90.1%, EA 89.3%, Wan2.2 98.4% vs Wan2.1 81.7%, LTX 66.1% | Mixed |
| change KDE kernel or bandwidth rule (ablation) | SWIFT · EA | avg accuracy: 97.80% -> 97.40% | No clear effect |
| shorten video to 25% of frames (ablation) | SWIFT · Wan2.2 | accuracy: 98.0% -> 94.1% | Hurt |
| apply post-processing (flip, noise, crop, compression) to test videos (ablation) | SWIFT · EA | accuracy: 97.8% -> 52.0% (Flip), 60.3% (Noise) | Hurt |
| use windowed reconstruction instead of full-video double reconstruction | SWIFT vs AEDR · runtime | runtime: 227s -> 218s (HYV); 31s -> 21s (EA) | Improved |
Closest papers by meaning
- ByTheWay: Boost Your Text-to-Video Generation Model to Higher Quality in a Training-free Way (CVPR 2025 · similarity 0.86)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.85)
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.85)
- Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis (CVPR 2024 · similarity 0.85)
- A Recipe for Scaling up Text-to-Video Generation with Text-free Videos (CVPR 2024 · similarity 0.85)
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search (NeurIPS 2025 · similarity 0.85)
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.
