Research map / Video / Video generation
PISCES: Annotation-free Text-to-Video Post-Training via Optimal Transport-Aligned Rewards
Minh-Quan Le, Gaurav Mittal, Cheng Zhao, David Gu, Dimitris Samaras, Mei Chen
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 OT alignment of text-video embeddings (ablation) | PISCES · VBench VideoCrafter2 | Semantic score: 75.82 -> 77.63 | Improved |
| add OT alignment of text-video embeddings (ablation) | PISCES · VBench VideoCrafter2 | Quality score: 83.44 -> 83.73 | Improved |
| add OT-aligned semantic reward to quality-only (ablation) | PISCES · VBench VideoCrafter2 | Semantic score: 75.97 -> 77.63 | Improved |
| add OT-aligned quality reward to semantic-only (ablation) | PISCES · VBench VideoCrafter2 | Quality score: 82.87 -> 83.73 | Improved |
| add OT-aligned quality reward (ablation) | PISCES · VBench VideoCrafter2 | Quality score: 82.20 -> 83.77 | Improved |
| add OT-aligned semantic reward (ablation) | PISCES · VBench VideoCrafter2 | Semantic score: 73.42 -> 76.99 | Improved |
| replace L2 mapping with OT alignment (ablation) | PISCES · VBench HunyuanVideo | Spearman correlation: 0.4873 -> 0.9018 | Improved |
| replace contrastive alignment with OT (ablation) | PISCES · VBench HunyuanVideo | Semantic score: 77.00 -> 80.33 | Improved |
| use partial OT m=0.9 instead of full OT (ablation) | PISCES semantic reward · VTM accuracy | VTM accuracy (%): 86.87 -> 87.54 | Improved |
| lower partial OT mass to m=0.5 (ablation) | PISCES semantic reward · VTM accuracy | VTM accuracy (%): 81.25 -> 78.92 | Hurt |
| add spatio-temporal cost (gamma=eta=0.2) to partial OT (ablation) | PISCES semantic reward · VTM accuracy | VTM accuracy (%): 87.54 -> 89.36 | Improved |
| lower spatio-temporal weights to gamma=eta=0.1 (ablation) | PISCES semantic reward · VTM accuracy | VTM accuracy (%): 87.54 -> 87.06 | Hurt |
Closest papers by meaning
- Think, then Score: Decoupled Reasoning and Scoring for Video Reward Modeling (NeurIPS 2026 · similarity 0.86)
- Improving Video Generation with Human Feedback (NeurIPS 2025 · similarity 0.85)
- InstructVideo: Instructing Video Diffusion Models with Human Feedback (CVPR 2024 · similarity 0.85)
- Boosting Text-to-Video Generative Model with MLLMs Feedback (NeurIPS 2024 · similarity 0.85)
- Confidence-aware Reward Optimization for Fine-tuning Text-to-Image Models (ICLR 2024 · similarity 0.85)
- TLDR: Token-Level Detective Reward Model for Large Vision Language Models (ICLR 2025 · similarity 0.85)
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.
