Research map / Video / Video generation
TTOM: Test-Time Optimization and Memorization for Compositional Video Generation
Leigang Qu, Ziyang Wang, Na Zheng, Wenjie Wang, Liqiang Nie, Tat-Seng Chua
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 |
|---|---|---|---|
| add test-time optimization with layout loss (ablation) | Wan2.1-14B · T2V-CompBench Motion | Motion score: 0.2696 -> 0.4321 | Improved |
| add parametric memory on top of TTO (ablation) | TTOM · T2V-CompBench Motion | Motion score: 0.4321 -> 0.4922 | Improved |
| add memory initialization (top-5) without TTO (ablation) | Wan2.1-14B · T2V-CompBench Motion | Motion score: 0.2696 -> 0.4754 | Improved |
| add continual TTO after memory init (top-5) (ablation) | TTOM · T2V-CompBench Motion | Motion score: 0.4754 -> 0.4846; latency 427 -> 627 | Improved |
| raise top-k memory entries fused for init from 5 to 10 (ablation) | TTOM · T2V-CompBench Motion | Motion score: 0.4754 -> 0.4437 | Hurt |
| use JSD loss instead of CE loss (ablation) | TTOM · T2V-CompBench Motion | Motion score: 0.2912 -> 0.4321 | Improved |
| use JSD loss instead of CoM loss (ablation) | TTOM · T2V-CompBench Numeracy | Numeracy score: 0.4697 -> 0.5881 | Improved |
| increase pseudo-training prompts for memory from 50 to 200 (ablation) | TTOM · T2V-CompBench Motion | Motion score: 0.3519 -> 0.4922 | Improved |
| increase optimized denoising timesteps from 1 to 5 (ablation) | TTOM · T2V-CompBench Motion | Motion score: 0.2730 -> 0.4321 | Improved |
| increase optimized timesteps from 5 to 7 (ablation) | TTOM · T2V-CompBench Motion | Motion score: 0.4321 -> 0.4296 | No clear effect |
| increase iterations per timestep from 8 to 16 (ablation) | TTOM · T2V-CompBench Motion | Motion score: 0.4321 -> 0.3547 | Hurt |
| apply TTOM to CogVideoX-5B | CogVideoX-5B · T2V-CompBench | Avg score: 0.4189 -> 0.5632 | 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.
- CogVideoX-5B — CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer (ICLR 2025) · better on 4 of 4 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
- VideoTetris: Towards Compositional Text-to-Video Generation (NeurIPS 2024 · similarity 0.89)
- PlasticMem: Adding Temporal Reasoning to Diffusions for Consistent Long Video Generation (NeurIPS 2026 · similarity 0.89)
- T2V-CompBench: A Comprehensive Benchmark for Compositional Text-to-video Generation (CVPR 2025 · similarity 0.88)
- Compositional 3D-aware Video Generation with LLM Director (NeurIPS 2024 · similarity 0.88)
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.87)
- Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis (CVPR 2024 · 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.
