PaperlayerView exact revision on arXiv

Source-linked AI summary

Diverse Video Generation with Determinantal Point Process-Guided Policy Optimization

Tahira Kazimi, Connor Dunlop, Pinar Yanardag

arXiv:2511.20647v1cs.CV

TL;DR

Text-to-video models can produce coherent, prompt-aligned videos but often repeat a narrow range of outcomes. DPP-GRPO combines DPP-based diminishing returns with groupwise policy feedback to optimize diversity across candidate sets. The authors report improved diversity while preserving fidelity and temporal coherence, and release a 30,000-prompt benchmark resource.

  • Problem

    Text-to-video models often repeat similar styles or scenarios, overlooking plausible variations for a prompt.

    Source: §1 · pp. 1–2

  • Method

    DPP-GRPO formulates diverse video generation as set-level policy optimization, combining DPP diminishing returns with GRPO groupwise feedback.

    Source: §1 · p. 2

  • Results

    The authors report consistent improvements in set diversity across standard text-to-video benchmarks and human studies while preserving fidelity and temporal coherence.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The released 30,000 prompt–variant pairs provide a benchmark resource designed specifically for diverse video generation research.

    Source: §1 · p. 2

    Because producing 1,000 Veo3 videos costs $3,000, the paper restricts its Veo3 evaluation to qualitative comparisons.

    Source: §5 · p. 5

Abstract

from arXiv · show

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.

1. Introduction

Text-to-video models can produce coherent, prompt-aligned videos yet repeatedly favor a narrow range of scenes, overlooking plausible variations. DPP-GRPO treats diversity as a set-level policy-optimization objective, while prior diversity approaches leave video-specific challenges unresolved.

1. Introduction

  • Current video generators often repeat similar styles or scenarios, overlooking other plausible variations despite producing coherent, high-fidelity videos.The introduction illustrates the pattern with repeated scenes and motions for a giraffe prompt.Source: §1 · pp. 1–2
  • Prior diversity methods face video-specific limitations, including costly test-time optimization, memory-intensive latent caches, training-set access, or architectural changes.Video diversity also spans motion, camera movement, and scene structure, dimensions not captured by visual appearance alone.Source: §1 · p. 2
  • DPP-GRPO combines GRPO groupwise optimization with a DPP diminishing-returns objective to favor semantically faithful, non-redundant prompt sets.The DPP discounts redundant samples, while GRPO supplies feedback over groups of candidates.Source: §1 · p. 2
  • DPP-GRPO is described as model-agnostic and plug-and-play, requiring no architectural changes and applying to open models and black-box APIs.The paper names Wan, CogVideoX, and Veo as examples.Source: §1 · p. 2
  • The authors report consistent gains in set diversity while preserving fidelity and temporal coherence, and release 30K prompt–variant pairs as a benchmark resource.The dataset is presented as a resource for diverse video generation research.Source: §1 · p. 2

2. Related Work

  • Prior work addresses image diversity, video preference alignment, motion, or prompt control, but the related-work section says none directly targets diversity of generated videos.The stated gap concerns diversity among video outputs, not video generation or alignment generally.Source: §2 · pp. 2–3

3. Background

The background introduces DPPs as a way to favor diverse subsets and describes GRPO as group-based reward normalization. It also notes that standard GRPO can concentrate generation on similar high-reward outputs.

Determinantal Point Processes

  • An L-ensemble DPP models subset probabilities using a positive semidefinite matrix, with the probability proportional to the determinant of the subset’s principal submatrix.The principal submatrix is indexed by the elements in the subset.Source: §3 · p. 3
  • When L entries represent similarity, the determinant acts as squared volume, increasing for diverse, linearly independent elements and shrinking for similar or redundant ones.This gives DPPs a natural preference for internally diverse subsets.Source: §3 · p. 3

Group Relative Policy Optimization

  • GRPO computes advantages by normalizing sampled-output rewards within a group and does not require a separate critic network.The method samples G responses for a query and uses group-based feedback to update the policy.Source: §3 · p. 3
  • Standard GRPO can cause diversity collapse by favoring similar high-reward outputs rather than exploring diverse valid responses.Its advantage normalization encourages higher probability for the sample with maximum reward.Source: §3 · p. 3

4. Methodology

The method trains a set-level policy to generate prompt variants that jointly span diverse cinematic modes and remain faithful to the user’s intent. DPP marginal-diversity rewards and GRPO groupwise feedback are combined with relevance rewards, then used in training and sequential generation.

4.1. DPP-GRPO

  • DPP-GRPO optimizes sets of video prompts for both diversity and semantic alignment with the user’s query.The prompts can then be used with any text-to-video generator.Source: §4.1 · p. 4
  • The log-determinant of embedded prompts measures their semantic volume, increasing for diverse prompts and decreasing for overly similar ones.Marginal gains are computed relative to a reference set of curated variants that covers distinct semantic and cinematic modes.Source: §4.1 · p. 4
  • A relevance reward uses similarity to both the user query and reference variants to encourage semantically faithful prompts.The composite reward balances this alignment signal with the DPP diversity gain.Source: §4.1 · p. 4
  • Training first fine-tunes the policy on prompts paired with diverse variants, then uses GRPO to update it from groupwise rewards.Fine-tuning runs for up to 50 iterations; at inference, each new prompt is generated relative to the partially constructed set until K prompts are produced.Source: §4.1 · pp. 4–5
  • Inference proceeds autoregressively, while the curated dataset contains 30K variants grounded in video-level diversity and quality metrics.An architect proposes variations and a critic evaluates them using TIE, TCE, and CLIP.Source: §4.1 · p. 5

5. Experiments

Experiments report greater diversity while preserving prompt fidelity, and assess benchmark quality, efficiency, human ratings, and robustness across models and settings.

5. Experiments

  • Qualitative comparisons show variation in subjects, scenes, motion, and camera views while preserving prompt alignment.Examples include varied skateboarding maneuvers and camera movements, as well as different kayaking perspectives.Source: §5 · p. 5, Figure 4 · p. 6
  • DPP-GRPO substantially improves diversity across Wan2.1 and CogVideoX while retaining competitive VideoScore and VBench performance; CogVideoX also achieves the strongest overall VideoScore.VBench semantic consistency and appearance quality remain stable or slightly improve.Source: Table 1 · p. 7, §5 · p. 7
  • CLIP alignment improves over both base models, and VideoScore’s TVA metric also improves with DPP-GRPO.Source: §5 · p. 7
  • In the user study, DPP-GRPO scores significantly higher in diversity than all baselines and higher on average in alignment.Participants rated sets of four clips for 10 prompts on five-point scales.Source: §5 · p. 8, Table 3 · p. 8
  • DPP-GRPO achieves near-baseline efficiency, adding 0.67% overhead; this is second only to Promptist (0.51s, 0.60% overhead), while SPARKE [13], an SDXL image-based method, adds 12% overhead per image through iterative optimization.Source: §5 · p. 8, Table 2 · p. 8

5.1. Ablation Studies.

  • Across CFG values 2–12, DPP-GRPO has higher TCE and TIE than Wan, comparable CLIP fidelity, and its highest CLIP score at CFG 6.Source: Figure 6 · p. 8, §5.1 · p. 9
  • Across Wan2.1 and CogVideoX, the method improves semantic and visual diversity; the full two-term reward balances diversity with prompt fidelity.Using only relevance reduces diversity, while using only DPP diversity modestly reduces text–video consistency.Source: §5.1 · p. 8, Table 4 · p. 9
  • As generation sets grow, DPP-GRPO maintains higher TCE and TIE than Wan at every size, while Wan’s diversity plateaus earlier and CLIP alignment remains stable.Source: Figure 7 · p. 9, §5.1 · p. 9

6. Discussion

DPP-GRPO targets set-level video diversity while preserving prompt fidelity, and experiments show consistent gains in set-level diversity. Its gains remain bounded by base-model limitations in complex temporal dynamics.

6. Discussion

  • Experiments show consistent gains in set-level diversity with equal or better alignment and perceptual quality.The framework combines group-relative policy optimization with a DPP-based diminishing-returns term.Source: §6 · p. 9
  • DPP-GRPO encourages complementary coverage across cinematographic and semantic factors while preserving prompt fidelity.It operates in prompt space and is described as model-agnostic and plug-and-play, including for open models and black-box systems.Source: §6 · p. 9
  • The method directly yields diverse sets without post-hoc re-ranking or cherry-picking.Source: §6 · p. 9
  • The approach inherits base-model limitations in complex temporal dynamics, which are governed by temporal attention values rather than prior conditioning [32].The paper notes that difficulty with fine-grained complex actions, such as peeling, remains a constraint.Source: §6 · p. 9
  • The authors state that designing T2V generation for diversity can reduce compute-intensive trial-and-error, operational cost, and access barriers for creators or small teams.Source: §6 · p. 9

7. Additional Qualitative Examples

This section points readers to supplementary videos and reports ablations on reference-set size and reward-weight balance.

7. Additional Qualitative Examples

  • Supplementary materials contain the videos from the main paper as well as additional videos.Source: §7 · p. 12
  • Small reference sets of 5–8 examples yield the best diversity, while larger sets show diminishing returns.Source: Table 5 · p. 12
  • The reward-weight ablation varies the balance between diversity and relevance during training.Source: Table 6 · p. 12

8. Video Prompts

The paper omits full prompts from the main text because of space limitations and provides complete prompts for selected generated videos in supplementary tables.

8. Video Prompts

  • Full prompts were omitted from the main paper because of space limitations.Source: §8 · p. 12
  • Complete prompts are provided for several videos generated by the method.Source: §8 · p. 12
  • Because space limitations prevented including full prompts in the main paper, tables provide complete prompts for several videos generated by the method, as well as the system prompt.Source: §8 · p. 12

9. User Study Details

The user study presents participants with four videos generated by a given method and collects ratings for diversity and text alignment on five-point Likert scales.

9. User Study Details

  • Participants are shown four videos generated by a given method.Source: §9 · p. 12
  • The study asks users to rate both diversity and text alignment.Source: §9 · p. 12
  • Both ratings use a Likert scale from 1 to 5.Source: §9 · p. 12

10. Additional Ablations

The ablations show that modest reference sets and balanced reward weights offer useful tradeoffs between diversity and semantic alignment. Larger reference sets can weaken diversity gains, while emphasizing diversity or relevance shifts performance in opposite directions.

10.1. Reference Set Size during Training

  • Reference sets of 5–8 samples provide the best balance between diversity gains and semantic stability.Sets of 10 degrade diversity performance, while CLIP alignment remains stable overall with a mild drop for larger sets.Source: §10.1 · p. 12
  • Small multi-reference sets produce a more reliable marginal-gain signal by exposing the policy to several video-grounded modes.Diversity performance degrades with larger sets, consistent with DPP log-determinant diminishing returns and increased variance in similarity estimates.Source: §10.1 · p. 12

10.2. Ablation on λ hyperparameters

  • Increasing λdiv to 0.9 yields the highest TIE/TCE while moderately CLIP alignment; λrel = 0.9 improves CLIP but reduces diversity.The balanced (0.5, 0.5) setting provides a middle ground across all metrics.Source: §10.2 · p. 12

11. Dataset Generation

The dataset-generation process expands VBench-grounded prompts into diverse variants, then retains expansions based on video-level diversity, alignment, and quality criteria.

11. Dataset Generation

  • The dataset starts with approximately 350 prompts for each of the 7 VBench categories and includes the original VBench prompts in its evaluation set.This keeps the prompt space grounded in a widely used video-generation benchmark.Source: §11 · p. 12
  • An architect agent proposes candidate expansions and a critic evaluates the resulting videos using TCE/TIE, CLIP, and VideoScore.Only expansions meeting diversity and alignment criteria are retained, grounding textual variety in video-level behavior.Source: §11 · p. 12
  • Each curated base prompt is expanded into 10 diverse variants using an iterative, two-agent reasoning framework.The expansion instruction asks for variants that differ from references while preserving meaning.Source: §11 · pp. 12–13
  • A valid expansion can add descriptive detail while preserving the prompt’s subject, action, and setting.The example contrasts a detailed cat description with an invalid expansion that changes the number, action, and setting.Source: §11 · p. 13
Loading 2511.20647v1…