SPATIALALIGN: Aligning Dynamic Spatial Relationships in Video Generation
This work improves the accuracy of spatial relationships in generated videos, which is an incremental improvement for users of text-to-video generation.
This paper addresses the problem of text-to-video generators often ignoring spatial constraints in generated videos. They introduce SPATIALALIGN, a self-improvement framework using zeroth-order regularized Direct Preference Optimization to fine-tune T2V models, resulting in significantly improved alignment with dynamic spatial relationships compared to baselines.
Most text-to-video (T2V) generators prioritize aesthetic quality, but often ignoring the spatial constraints in the generated videos. In this work, we present SPATIALALIGN, a self-improvement framework that enhances T2V models capabilities to depict Dynamic Spatial Relationships (DSR) specified in text prompts. We present a zeroth-order regularized Direct Preference Optimization (DPO) to fine-tune T2V models towards better alignment with DSR. Specifically, we design DSR-SCORE, a geometry-based metric that quantitatively measures the alignment between generated videos and the specified DSRs in prompts, which is a step forward from prior works that rely on VLM for evaluation. We also conduct a dataset of text-video pairs with diverse DSRs to facilitate the study. Extensive experiments demonstrate that our fine-tuned model significantly out performs the baseline in spatial relationships. The code will be released in Link. Project page: https://fengming001ntu.github.io/SpatialAlign/