Research map / Video / Video generation
NoisEasier: Test-Time Noise Optimization for Text-to-Video Generation
Yujiang Pu, Yu Kong
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 NoisEasier test-time noise optimization | T2V-Turbo (MS) · VBench | Multiple Objects: 54.50 -> 76.62 | Improved |
| add NoisEasier | T2V-Turbo (VC2) · VBench | Spatial Relation: 39.95 -> 52.61 | Improved |
| add NoisEasier | T2V-Turbo (MS) · VBench | Motion Smoothness: 95.51 -> 95.29 | Hurt |
| add NoisEasier | AnimateLCM · T2V-CompBench | Consistent Attribute Binding: 0.6746 -> 0.7807 | Improved |
| add NoisEasier | T2V-Turbo (VC2) · T2V-CompBench | Dynamic Attribute: 0.0113 -> 0.0097 | Hurt |
| optimize full trajectory (z_T and perturbations) instead of initial noise only (ablation) | NoisEasier · VBench Overall Consistency | score: T2V-Turbo MS 29.67 -> 32.05 | Improved |
| optimize full trajectory instead of initial only (ablation) | NoisEasier · VBench Dynamic Degree | score: AnimateLCM 30.83 -> 30.56; MS 66.94 -> 71.39 | Mixed |
| add ImageReward only vs combined rewards (ablation) | NoisEasier · VBench Dynamic Degree | score: 70.83 -> 61.11 | Hurt |
| combine all rewards instead of single reward (ablation) | NoisEasier · VBench Overall Consistency | score: 36.21 (ViCLIP only) vs 32.05 combined; Dynamic 63.89 -> 71.39 | Improved |
| add negative-aware reward calibration (ablation) | NoisEasier · VBench Color | score: 93.29 -> 94.38 | Improved |
| use random POS-perturbation negatives instead of LLM negatives (ablation) | NoisEasier · VBench Spatial Relation | score: 47.95 -> 46.39 (vs no negs) | Hurt |
| use noise optimization instead of best-of-n sampling | NoisEasier · VBench Multiple Objects | score: 80.06 (BoS) vs 76.62 | Hurt |
Closest papers by meaning
- T2V-Turbo: Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback (NeurIPS 2024 · similarity 0.89)
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search (NeurIPS 2025 · similarity 0.89)
- FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling (ICLR 2024 · similarity 0.89)
- VideoTetris: Towards Compositional Text-to-Video Generation (NeurIPS 2024 · similarity 0.88)
- LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity (CVPR 2025 · similarity 0.88)
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.88)
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.
