Research map / Video / Video generation
LongSCOP: Semantically Consistent Long Video Outpainting
Yu Tang, Zhiyuan You, Zhixiao Wang, Fanghua Yu, Qingyu Zhang, Xiang Yin, Chao Dong, Jinjin Gu
Read the paperCheck your idea against 137,217 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.
Proposes LongSCOP, a method for semantically consistent long video outpainting that addresses instability and semantic drift in chunk-by-chunk iterative generation.
Problem
Long video outpainting is underexplored; iterative chunk-by-chunk outpainting yields unstable or crashed long-horizon generation and semantic inconsistency across chunks.
Approach
Targets inter-chunk inconsistency in iterative long-video outpainting to keep semantics coherent across chunks; the abstract is truncated so the specific mechanism is not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add stable generation strategies (self-forcing, latent calibration, future reference) (ablation) | VACE out-painting · LVO-Bench | FVD: 225.5 -> 208.6 | Improved |
| add key reference frame retrieval (ablation) | LongSCOP · LVO-Bench | FVD: 208.6 -> 197.3 | Improved |
| add key reference frame retrieval (ablation) | LongSCOP · LVO-Bench | SSIM: 0.751 -> 0.749 | No clear effect |
| use LongSCOP on short clips | VACE · DAVIS / YouTube-VOS | FVD: 195.7 -> 197.1 (DAVIS); 45.8 -> 53.9 (YouTube-VOS) | No clear effect |
| use LongSCOP instead of VACE on short clips | VACE · DAVIS | PSNR: 18.83 -> 20.04 | Improved |
| use LongSCOP instead of best baseline on YouTube-VOS | OutDreamer · YouTube-VOS | PSNR: 20.21 -> 19.31 | Hurt |
| filter out bottom 30% of candidates by registration score before VLM (ablation) | reference retrieval · LVO-Bench | nDCG@3: 0.5612 -> 0.5815 | Improved |
| filter 80% of candidates instead of 30% (ablation) | reference retrieval · LVO-Bench | nDCG@3: 0.6115 -> 0.4805 | Hurt |
| swap Mast3R for RoMa registration with 27B VLM (ablation) | reference retrieval · LVO-Bench | Hit@3: 0.6744 -> 0.6512 | No clear effect |
| use 9B VLM instead of 27B (ablation) | reference retrieval · LVO-Bench | Hit@3: 0.6744 -> 0.6202 (Mast3R) | Hurt |
| add retrieval and bidirectional generation | VACE · inference time per 15 s video | minutes: 13.5 -> 20.1 | Hurt |
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.
- Infinite-Canvas — Infinite-Canvas: Higher-Resolution Video Outpainting with Extensive Content Generation (AAAI 2025) · better on 5 of 6 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
- Infinite-Canvas: Higher-Resolution Video Outpainting with Extensive Content Generation (AAAI 2025 · similarity 0.88)
- LongDiff: Training-Free Long Video Generation in One Go (CVPR 2025 · similarity 0.87)
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.87)
- SEINE: Short-to-Long Video Diffusion Model for Generative Transition and Prediction (ICLR 2024 · similarity 0.86)
- LVD-2M: A Long-take Video Dataset with Temporally Dense Captions (NeurIPS 2024 · similarity 0.86)
- StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation (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.
