PISCES: Annotation-free Text-to-Video Post-Training via Optimal Transport-Aligned Rewards

TL;DR

PISCES uses dual optimal transport-aligned rewards for annotation-free text-to-video post-training, enhancing video quality and semantic consistency.

cs.CV 🔴 Advanced 2026-02-02 13 views
Minh-Quan Le Gaurav Mittal Cheng Zhao David Gu Dimitris Samaras Mei Chen
text-to-video annotation-free optimal transport reward mechanism post-training

Key Findings

Methodology

PISCES employs a dual optimal transport (OT)-aligned rewards module for annotation-free text-to-video post-training. This method aligns text and video embeddings at both distributional and discrete token levels, providing quality and semantic rewards. The distributional OT-aligned quality reward focuses on visual quality and temporal coherence, while the discrete token-level OT-aligned semantic reward ensures semantic consistency between text and video tokens.

Key Results

  • PISCES outperforms both annotation-based and annotation-free methods on VBench, with significant improvements in quality and semantic scores.
  • In human preference studies, PISCES is preferred for visual quality and semantic consistency.
  • The dual OT-aligned rewards module is compatible with various optimization paradigms, including direct backpropagation and reinforcement learning fine-tuning.

Significance

PISCES offers an annotation-free text-to-video post-training method that addresses scalability and supervision limitations of existing methods. By aligning reward signals through optimal transport, it achieves significant advancements in video generation's visual quality and semantic consistency.

Technical Contribution

PISCES is the first to improve annotation-free reward supervision in generative post-training through the lens of optimal transport. Its dual OT-aligned rewards module aligns text and video embeddings at both distributional and discrete token levels, providing new theoretical guarantees and engineering possibilities.

Novelty

PISCES is the first method to improve annotation-free reward supervision in generative post-training through optimal transport. Its innovation lies in aligning text and video embeddings at distributional and token levels, which is a significant departure from existing methods.

Limitations

  • PISCES may experience performance degradation in complex video scenarios, especially those involving numerous dynamic elements.
  • There is room for improvement in computational complexity, particularly for large-scale datasets.

Future Work

Future work could explore PISCES's application in more video generation tasks, such as real-time video generation. Additionally, further optimization of algorithmic efficiency and scalability is a key research direction.

AI Executive Summary

Text-to-video (T2V) generation aims to synthesize videos with high visual quality and temporal consistency that are semantically aligned with input text. However, existing methods either rely on large-scale human preference annotations or operate on misaligned embeddings from pre-trained vision-language models, leading to limited scalability or suboptimal supervision. PISCES presents an annotation-free post-training algorithm that addresses these limitations via a novel Dual Optimal Transport (OT)-aligned Rewards module. This method aligns text and video embeddings at both distributional and discrete token levels, providing quality and semantic rewards. Experiments show that PISCES outperforms both annotation-based and annotation-free methods on short- and long-video generation, with significant improvements in quality and semantic scores on VBench, and human preference studies further validate its effectiveness. PISCES's dual OT-aligned rewards module is compatible with multiple optimization paradigms, including direct backpropagation and reinforcement learning fine-tuning, offering new possibilities for text-to-video generation. Despite these advancements, PISCES may experience performance degradation in complex video scenarios. Future work could explore its application in more video generation tasks and further optimize algorithmic efficiency and scalability.

Deep Analysis

Background

Text-to-video generation is a crucial task in multimedia content creation, with significant advancements in recent years due to diffusion models and vision-language models. However, existing methods either rely on large-scale human preference annotations or operate on misaligned embeddings, leading to limited scalability or suboptimal supervision. PISCES addresses these issues with an annotation-free post-training algorithm.

Core Problem

Existing text-to-video generation methods face challenges in semantic alignment and visual quality, especially in the absence of large-scale human preference annotations. Achieving high-quality text-to-video generation without annotations is a significant research problem.

Innovation

PISCES's core innovation lies in its dual optimal transport-aligned rewards module. This module aligns text and video embeddings at both distributional and discrete token levels, providing quality and semantic rewards. Unlike existing methods, PISCES does not require large-scale human preference annotations, offering better scalability.

Methodology

  • �� Uses a dual optimal transport-aligned rewards module to align text and video embeddings.
  • �� Distributional OT-aligned quality reward focuses on visual quality and temporal coherence.
  • �� Discrete token-level OT-aligned semantic reward ensures semantic consistency between text and video tokens.
  • �� Compatible with various optimization paradigms, including direct backpropagation and reinforcement learning fine-tuning.

Experiments

Experiments were conducted using VideoCrafter2 and HunyuanVideo datasets, representing short- and long-video settings. Evaluation metrics included quality and semantic scores on VBench and human preference studies. Results show that PISCES outperforms existing methods across all metrics.

Results

PISCES achieves significant improvements in quality and semantic scores on VBench, outperforming both annotation-based and annotation-free methods. Human preference studies show higher preference for PISCES in terms of visual quality and semantic consistency.

Applications

PISCES can be applied in multimedia content creation, robotics, and accessibility. Its annotation-free nature offers advantages in large-scale applications.

Limitations & Outlook

PISCES may experience performance degradation in complex video scenarios, especially those involving numerous dynamic elements. Future work could explore its application in more video generation tasks and further optimize algorithmic efficiency and scalability.

Plain Language Accessible to non-experts

Imagine you're in a kitchen cooking a meal. You have a recipe (text) and need to make a delicious dish (video) based on it. PISCES is like a smart kitchen assistant that helps you create the dish without needing detailed steps. It uses a method called 'optimal transport' to ensure each ingredient (keyword in the text) finds its right place in the dish (corresponding video segment). This way, even without detailed cooking instructions (annotations), you can make a delicious dish (high-quality video).

ELI14 Explained like you're 14

Hey there! Imagine you're playing a game where you need to create a virtual world (video) based on a mission description (text). PISCES is like your super helper that helps you create this world perfectly without needing detailed instructions. It uses something called 'optimal transport' magic to make sure every mission detail finds its right place in the world. So even without detailed mission instructions (annotations), you can create an awesome virtual world! Isn't that cool?

Glossary

Optimal Transport

A mathematical method for finding the best match between two probability distributions.

Used to align text and video embeddings.

Distributional Alignment

Aligning data from different modalities at the distribution level.

Ensures overall visual quality between text and video.

Discrete Token Alignment

Aligning data from different modalities at the token level.

Ensures semantic consistency between text and video.

Reward Mechanism

A feedback signal used in machine learning to guide model learning.

Enhances video generation quality and semantic consistency.

Post-Training

Further optimization conducted after initial model training.

Improves text-to-video generation performance.

Open Questions Unanswered questions from this research

  • 1 How to improve PISCES's performance in complex video scenarios without increasing computational complexity?
  • 2 What is the potential of PISCES in real-time video generation?
  • 3 How to further optimize PISCES's computational efficiency for large-scale datasets?

Applications

Immediate Applications

Multimedia Content Creation

PISCES can be used to generate high-quality multimedia content for films, advertisements, and more.

Long-term Vision

Real-Time Video Generation

PISCES has potential for real-time video generation, requiring solutions for computational efficiency and dynamic scenario adaptation.

Abstract

Text-to-video (T2V) generation aims to synthesize videos with high visual quality and temporal consistency that are semantically aligned with input text. Reward-based post-training has emerged as a promising direction to improve the quality and semantic alignment of generated videos. However, recent methods either rely on large-scale human preference annotations or operate on misaligned embeddings from pre-trained vision-language models, leading to limited scalability or suboptimal supervision. We present $\texttt{PISCES}$, an annotation-free post-training algorithm that addresses these limitations via a novel Dual Optimal Transport (OT)-aligned Rewards module. To align reward signals with human judgment, $\texttt{PISCES}$ uses OT to bridge text and video embeddings at both distributional and discrete token levels, enabling reward supervision to fulfill two objectives: (i) a Distributional OT-aligned Quality Reward that captures overall visual quality and temporal coherence; and (ii) a Discrete Token-level OT-aligned Semantic Reward that enforces semantic, spatio-temporal correspondence between text and video tokens. To our knowledge, $\texttt{PISCES}$ is the first to improve annotation-free reward supervision in generative post-training through the lens of OT. Experiments on both short- and long-video generation show that $\texttt{PISCES}$ outperforms both annotation-based and annotation-free methods on VBench across Quality and Semantic scores, with human preference studies further validating its effectiveness. We show that the Dual OT-aligned Rewards module is compatible with multiple optimization paradigms, including direct backpropagation and reinforcement learning fine-tuning. Project page: https://roar-ai.github.io/pisces

cs.CV