Research map / Video / Video generation
Free-Lunch Long Video Generation via Layer-Adaptive O.O.D Correction
Jiahao Tian, Chenxi Song, Wei Cheng, Chi Zhang
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 |
|---|---|---|---|
| apply layer-adaptive VRPR / TSA+VRPR instead of uniform TSA+VRPR on all layers (ablation) | FreeLOC on Wan2.1-T2V-1.3B · VBench 4x, Aesthetic Quality | Aesthetic Quality: 56.34 -> 61.21 | Improved |
| layer-adaptive selection by probing instead of random layer assignment (ablation) | FreeLOC · VBench 4x, Aesthetic Quality | Aesthetic Quality: 54.44 -> 61.21 | Improved |
| add TSA to direct sampling at 4x length (ablation) | Wan2.1-T2V-1.3B direct sampling · VBench 4x, Dynamic Degree | Dynamic Degree: 4.32 -> 37.01 | Improved |
| add TSA to direct sampling at 4x length (ablation) | Wan2.1-T2V-1.3B direct sampling · VBench 4x, Subject Consistency | Subject Consistency: 98.50 -> 97.41 | Hurt |
| add VRPR alone to direct sampling at 4x length (ablation) | Wan2.1-T2V-1.3B direct sampling · VBench 4x, Aesthetic Quality | Aesthetic Quality: 49.43 -> 54.13 | Improved |
| add VRPR alone to direct sampling at 4x length (ablation) | Wan2.1-T2V-1.3B direct sampling · VBench 4x, Subject Consistency | Subject Consistency: 98.50 -> 98.42 | No clear effect |
| replace clipping with VRPR for relative position re-encoding (ablation) | FreeLOC · VBench 4x, Dynamic Degree | Dynamic Degree: 22.21 -> 36.27 | Improved |
| replace grouping with VRPR for relative position re-encoding (ablation) | FreeLOC · VBench 4x, Aesthetic Quality | Aesthetic Quality: 51.32 -> 61.21 | Improved |
| replace sliding window attention with TSA (ablation) | FreeLOC · VBench 4x, Aesthetic Quality | Aesthetic Quality: 54.92 -> 61.21 | Improved |
| replace sliding window attention with TSA (ablation) | FreeLOC · VBench 4x, Dynamic Degree | Dynamic Degree: 37.81 -> 36.27 | Hurt |
| replace LongDiff selected-frame attention with TSA (ablation) | FreeLOC · VBench 4x, Dynamic Degree | Dynamic Degree: 28.53 -> 36.27 | Improved |
| FreeLOC vs sliding window baseline at 4x length | Wan2.1-T2V-1.3B · VBench 4x, Subject Consistency | Subject Consistency: 96.15 -> 98.44 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FreeNoise — FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling (ICLR 2024) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.90)
- FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling (ICLR 2024 · similarity 0.88)
- LongDiff: Training-Free Long Video Generation in One Go (CVPR 2025 · similarity 0.88)
- FreePCA: Integrating Consistency Information across Long-short Frames in Training-free Long Video Generation via Principal Component Analysis (CVPR 2025 · similarity 0.87)
- Ouroboros-Diffusion: Exploring Consistent Content Generation in Tuning-free Long Video Diffusion (AAAI 2025 · similarity 0.86)
- VSTAR: Generative Temporal Nursing for Longer Dynamic Video Synthesis (ICLR 2025 · 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.
