Diverse Video Generation with Determinantal Point Process-Guided Policy Optimization

TL;DR

Text-to-video diffusion models often produce low-diversity outputs. The authors introduce DPP-GRPO, a set-level policy optimization that combines a Determinantal Point Process (DPP) based diminishing-returns objective with Group Relative Policy Optimization (GRPO) to explicitly reward diverse yet faithful video generations in prompt space. The framework uses a DPP marginal gain Delta(p_i|R_q) = log det(L_phi(R_q U {p_i})) − log det(L_phi(R_q)) and a relevance term to balance diversity with prompt fidelity, trained via a two-stage process and applied to open models without architectural changes; a 30K-prompt dataset supports evaluation. Results on Wan2.1 and CogVideoX show consistent improvements in set-level diversity across VBench, VideoScore, and human studies while maintaining alignment and perceptual quality, and a 30K-prompt diversity benchmark is released. This approach enables plug-and-play diversification for text-to-video systems, reducing the need for costly post-hoc tuning and enabling broader creative exploration across diverse cinematic factors.

Abstract

While recent text-to-video (T2V) diffusion models have achieved impressive quality and prompt alignment, they often produce low-diversity outputs when sampling multiple videos from a single text prompt. We tackle this challenge by formulating it as a set-level policy optimization problem, with the goal of training a policy that can cover the diverse range of plausible outcomes for a given prompt. To address this, we introduce DPP-GRPO, a novel framework for diverse video generation that combines Determinantal Point Processes (DPPs) and Group Relative Policy Optimization (GRPO) theories to enforce explicit reward on diverse generations. Our objective turns diversity into an explicit signal by imposing diminishing returns on redundant samples (via DPP) while supplies groupwise feedback over candidate sets (via GRPO). Our framework is plug-and-play and model-agnostic, and encourages diverse generations across visual appearance, camera motions, and scene structure without sacrificing prompt fidelity or perceptual quality. We implement our method on WAN and CogVideoX, and show that our method consistently improves video diversity on state-of-the-art benchmarks such as VBench, VideoScore, and human preference studies. Moreover, we release our code and a new benchmark dataset of 30,000 diverse prompts to support future research.

Figures (9)

• Figure 1: Given an input prompt, DPP-GRPO enables generation of diverse sets of videos spanning cinematic factors such as camera motion or scene layout. We formulate diversity as set-level policy optimization with a DPP-based diminishing-returns term to suppress similar videos as the set grows. For brevity, we display only the first frame of each clip; see the supplement for frame-by-frame comparisons and videos. We note that our results are uncurated and shown in the exact order produced by our method.
• Figure 2: Frame-level examples illustrating motion, subject, and camera diversity achieved by DPP-GRPO. Given the same input prompt, our method generates videos that vary in subject appearance, environment, and camera movement while maintaining semantic consistency with the original prompt.
• Figure 3: Framework Overview. The model generates a group of $G$ candidates $(C_1,\ldots,C_G)$ for each prompt. Each candidate is scored by a composite reward combining DPP marginal gain $\Delta(S \cup C_i)$ and relevance $R_{\text{rel}}(C_i)$. Groupwise normalization produces advantages $(A_i)$, which update the policy under the DPP--GRPO objective.
• Figure 4: Qualitative Comparison.DPP-GRPO diversifies several cinematic factors such as subject, scene, motion, and camera-view diversity while preserving the prompt alignment and achieves more diverse and semantically faithful videos compared to baselines. (a) For clarity, we provide the first frames of each video (b) Detailed frame-by-frame comparisons of the same videos are given (please kindly zoom-in for details). Please visit our SM for more comparisons and high-quality videos.
• Figure 5: Visual comparison of T2V generations from various base models, with and without our method. For each baseline (CogVideoX, Wan, VEO), we show two generated videos and their representative frames. Our method enhances the diversity and quality of the generated videos across integrated models.

Theorems & Definitions (1)

• Definition 3.1