Reproduce “PISCES: Annotation-free Text-to-Video Post-Training via Optimal Transport-Aligned Rewards” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Minh-Quan Le, Gaurav Mittal, Cheng Zhao, David Gu, Dimitris Samaras, Mei Chen
PISCES introduces an annotation-free post-training algorithm for Text-to-Video generation that leverages Optimal Transport to align text and video embeddings. This method creates dual rewards for global quality and fine-grained semantic fidelity, enabling T2V models to achieve performance comparable to or exceeding human-annotation-based approaches in visual quality, temporal consistency, and semantic alignment.
Key insights
- Misalignment within VLM text and video embeddings is a primary bottleneck for effective annotation-free T2V post-training, leading to suboptimal reward signals.
- Optimal Transport (OT) offers a principled mathematical framework for achieving robust text-video alignment at both distributional and token levels, which can be effectively leveraged for reward modeling.
- Combining a distributional OT-aligned quality reward with a discrete token-level OT-aligned semantic reward provides complementary supervision, allowing for comprehensive improvements in T2V generation across visual quality, temporal consistency, and semantic fidelity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.