Research map / Video / Video generation
Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformer
Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, Devi Parikh
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 |
|---|---|---|---|
| condition on class labels | TATS-base · UCF-101 | IS: 57.63 to 79.28 | Improved |
| condition on class labels | TATS-base · UCF-101 | FVD: 420 -> 332 | Improved |
| use replicate padding instead of zero padding in VQGAN (ablation) | time-agnostic VQGAN · Padding type study (UCF-101 VQGAN) | equivariance score: 0.75 | Improved |
| use replicate padding instead of reflect or circular padding (ablation) | time-agnostic VQGAN · Padding type study (UCF-101 VQGAN) | equivariance score | Improved |
| remove all paddings and add real frames to the input (ablation) | VQGAN · Padding type study (UCF-101 VQGAN) | equivariance score: equivariance score = 1 | Improved |
| replace zero padding with replicate padding for sliding-window generation (ablation) | vanilla video VQGAN · Sky/Taichi/UCF-101 long videos | FVD change over 1024 frames | Improved |
| add hierarchical autoregressive plus interpolation transformer (ablation) | TATS-base · UCF-101, Sky, Taichi long videos | FVD change over 1024 frames | Improved |
| add hierarchical transformer (ablation) | TATS-base · UCF-101 long videos | CCS and ICS coherence | Improved |
| use hierarchical transformer instead of single autoregressive transformer | TATS · Generating one 1024-frame video | generation time: 30 min -> 7.5 min | Improved |
| use TATS instead of DIGAN in human evaluation on Sky | TATS · Sky human eval vs DIGAN | preference for fewer artifacts: 67%; 355 vs. 175 | Improved |
| use TATS-base instead of DIGAN on Sky Time-lapse | TATS-base · Sky Time-lapse (16 frames) | FVD: 114.6 -> 132.56 | Hurt |
| use TATS-base instead of DIGAN on Taichi-HD | TATS-base · Taichi-HD (16 frames) | FVD: 128.1 -> 94.60 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- CCVS — CCVS: Context-aware Controllable Video Synthesis (NeurIPS 2021) · better on 3 of 3 shared results
- Sound2Sight — Sound2Sight: Generating Visual Dynamics from Sound and Context (ECCV 2020) · better on 2 of 2 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
- Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis (CVPR 2024 · similarity 0.88)
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.88)
- LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity (CVPR 2025 · similarity 0.88)
- VSTAR: Generative Temporal Nursing for Longer Dynamic Video Synthesis (ICLR 2025 · similarity 0.88)
- LVD-2M: A Long-take Video Dataset with Temporally Dense Captions (NeurIPS 2024 · similarity 0.88)
- LongDiff: Training-Free Long Video Generation in One Go (CVPR 2025 · similarity 0.87)
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.
