Source-linked AI summary
CineEvolve: An Agentic Self-Evolution Framework for Cinematic Multi-Shot Video Generation
Jingzhe Zhang, Peixuan Zhang, Genying Li, Bing Ma, Shuchen Weng, Si Li, Boxin Shi
TL;DR
Cinematic multi-shot generation must coordinate narrative planning with visual consistency and progression, while existing methods lack narrative correction or trained policy updates. CineEvolve unifies a Director and Critic in one vision-language policy and trains them through alternating reinforcement learning. The paper reports improvements over existing approaches in narrative coherence and visual execution.
Problem
Multi-shot video generation requires narrative planning and cross-shot coherence, but existing visual methods lack narrative correction and agentic methods typically lack policy updates through training.
Source: §1 · pp. 1–2Method
CineEvolve trains a unified Director–Critic vision-language policy through alternating reinforcement learning, supported by E-GRPO and CineTraj.
Source: §1 · p. 2Results
CineEvolve significantly outperforms existing approaches in narrative coherence and visual execution.
Source: §6 · p. 9Takeaways & Limitations
CineEvolve combines iterative multi-shot synthesis and stable co-evolution using role-specific rewards and E-GRPO, with CineTraj comprising 40K samples for training.
Source: §1 · p. 2, §6 · p. 9CineEvolve’s feedback loop operates at the storyboard level, leaving subtle motion artifacts from subsequent video synthesis outside its scope.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
Cinematic multi-shot video generation requires both narrative planning and iterative refinement. Existing methods either lack narrative reasoning or rely on feed-forward workflows without policy updates. In this paper, we propose CineEvolve, an agentic self-evolution framework for cinematic multi-shot video generation that unifies a Director and a Critic within a single vision-language policy. Emulating real-world filmmaking that couples iterative direction with critical feedback, we train CineEvolve via alternating reinforcement learning, where the Director surfaces flaws to sharpen the Critic, and the updated Critic guides the Director's targeted revisions. To stabilize co-evolution, we introduce Evolutionary GRPO (E-GRPO), leveraging historical anchoring and revision safeguarding to mitigate advantage inflation. Additionally, we contribute CineTraj, a 40K cinematic trajectory dataset spanning creation, critique, and revision. Extensive experiments demonstrate that CineEvolve outperforms existing approaches. Our code will be available at https://github.com/JingZhe-8/CineEvolve.
1. Introduction
Multi-shot storytelling requires cross-shot consistency and narrative progression, yet existing visual and agentic methods lack trained policy-level feedback for correcting narrative or execution flaws. CineEvolve addresses this with a unified Director–Critic policy trained through alternating reinforcement learning, alongside E-GRPO and CineTraj.
Motivation
- Existing visual-based methods lack narrative reasoning to rectify semantic errors, while agentic methods typically use feed-forward workflows or inference-time feedback without policy updates through training.Multi-shot sequences also require cross-shot consistency and inter-shot coherence, coupling narrative planning with visual synthesis.Source: §1 · pp. 1–2
CineTraj
- CineTraj supplies trajectories spanning initial creation, defect critique, and targeted revision for training and evaluation.The dataset includes 30K samples for supervised warm-up and 10K for alternating reinforcement learning.Source: §1 · p. 2
CineEvolve
- CineEvolve unifies Director and Critic roles in one vision-language policy and co-evolves them through alternating reinforcement learning.The Director exposes failure cases to train the Critic, whose updated diagnostic feedback guides the Director’s targeted revisions.Source: §1 · p. 2
- Evolutionary GRPO (E-GRPO) uses revision safeguarding and historical anchoring to mitigate advantage inflation during co-evolution.The framework also uses critic and director rewards for closed-loop optimization.Source: §1 · p. 2
2. Related Work
Prior multi-shot methods preserve visual continuity or plan shots with tools, but do not combine narrative correction with policy-level learning. CineEvolve extends agentic self-evolution to multi-shot video generation through a jointly trained Director and Critic.
2.1. Multi-Shot Video Generation
- Visual-based methods preserve continuity but struggle to diagnose or rectify semantic narrative flaws, while agentic methods’ feed-forward execution cannot adaptively correct failures and CineCrew lacks policy-level learning.CineEvolve instead unifies Director and Critic within one policy co-evolved through reinforcement learning.Source: §2.1 · p. 2
2.2. Agentic Self-Evolution Mechanisms
- Most prior self-evolution work targets nonvisual domains, while emerging multimodal efforts largely address isolated 2D perception or static images.The authors describe CineEvolve as the first effort to extend agentic self-evolution to multi-shot video generation by coupling creative direction and diagnostic critique.Source: §2.2 · p. 2
3. Dataset
CineTraj provides curated cinematic scenes and verified trajectories connecting initial storyboard generation with critique and targeted revision. Its 40K samples support supervised warm-up and alternating reinforcement learning.
Trajectory construction
- The teacher Critic diagnoses shot-level defect locations, textual issues, and severity ratings by comparing storyboards with source scenes and shot descriptions.The dataset also provides critique, initial-creation, and targeted-revision instances for distinct training objectives.Source: §3 · p. 3
Scene curation and initial generation
- The curation pipeline removes duplicates, segments and groups shots, and filters scene quality before teacher-led planning and storyboard generation.The teacher Director converts annotated stories and shot descriptions into production plans and initial storyboards, which then enter critique, revision, and re-evaluation.Source: Figure 2 · p. 3, §3 · p. 3
Trajectory construction
- CineTraj retains closed-loop demonstrations only when teacher-guided re-evaluation finds that severity-weighted defects strictly decrease after revision.Each trajectory connects initial generation, defect diagnosis, a targeted revision plan, and a revised storyboard.Source: §3 · p. 3
Data Summary
- CineTraj contains 30,000 verified trajectories for supervised warm-up and 10,000 trajectories for alternating reinforcement learning.It balances critique, initial creation, and targeted revision at a 4:2:4 sample ratio, and includes an independent 100-sequence benchmark excluded from training.Source: §3 · p. 3
4. Method
CineEvolve trains a shared Director–Critic policy through supervised warm-up and alternating reinforcement learning, then uses the Critic’s feedback for iterative storyboard revision. E-GRPO adds safeguards and historical anchoring to stabilize those alternating updates.
4. Method
- The method frames co-evolution as alternating Critic optimization, Director optimization, and E-GRPO stabilization to promote mutual improvement.The Critic is optimized using critic rewards, while the Director uses multimodal director rewards.Source: §4 · p. 4
4.1. Training and Inference Paradigm
- Supervised warm-up on CineTraj demonstrations precedes alternating Critic and Director reinforcement-learning updates, followed by an inference loop of generation, critique, and targeted revision.At test time, the policy can revise for up to 10 steps before the approved storyboard conditions video rendering.Source: §4.1 · p. 4
4.2. Critic Policy Optimization
- The Critic diagnoses defective shots, describes their flaws, and rates severity, while composite rewards assess format compliance, semantic fidelity, and severity calibration.Its replay pool is refreshed with half retained and half newly generated Director storyboards so diagnostics track evolving failure patterns.Source: §4.2 · p. 4
4.3. Director Policy Optimization
- The Director creates and revises storyboards using a two-stage schedule and rewards that balance executable tool calls with visual quality and narrative coherence.A frozen Critic snapshot supplies diagnoses and self-evaluation rewards; an external VLM scores coherence against a source-scene anchor excluded from policy inputs.Source: §4.3 · p. 5
4.4. Evolutionary GRPO
- E-GRPO combines revision safeguarding and historical anchoring to mitigate inflated advantages and stabilize alternating policy updates.Edits adding Critic-verified severe defects are capped at the parent storyboard reward, and revisions no better than their parent receive no positive learning signal; historical rewards also anchor comparisons.Source: §4.4 · pp. 5–6
5. Experiments
CineEvolve is evaluated with benchmark metrics, qualitative comparisons, human preferences, and ablations. It consistently outperforms existing approaches across eight metrics and receives the highest preference share in all four user-study dimensions.
5.1. Training Details
- Training uses two supervised warm-up epochs followed by five rounds of alternating reinforcement learning, with 64 optimization steps per role per round.The experiments use 16 NVIDIA A100 GPUs.Source: §5.1 · p. 6
5.2. Evaluation Metrics
- Evaluation covers visual quality and alignment, within-shot consistency, and cross-shot coherence using eight metrics.Cross-shot coherence metrics assess entity consistency and transition fidelity as well as attributes and narrative progression.Source: §5.2 · p. 6
5.3. Comparison with Existing Methods
- CineEvolve consistently outperforms existing multi-shot generation approaches across eight metrics.The quantitative comparison is presented in Table 1, alongside ablation results.Source: Table 1 · p. 7, §5.3 · p. 8
- CineEvolve receives the highest preference share in all four dimensions among seven methods, based on 25 stories rated by 20 volunteers.The dimensions are Visual Quality, Text Alignment, Cross-shot Consistency, and Inter-shot Coherence.Source: §5.3 · p. 8, Table 2 · p. 8
5.4. Ablation Studies
- Ablations associate removing alternating RL, freezing the Critic, separating policies, or replacing E-GRPO with degraded performance on distinct metrics and storyboard defects.The variants show chronological reversals and drift, weaker cross-shot consistency, lower transition coherence, or identity and state flickering, respectively.Source: Figure 5 · p. 8, §5.4 · p. 9
6. Conclusion
CineEvolve unifies Director and Critic roles in a vision-language policy and uses alternating reinforcement learning for iterative multi-shot synthesis. The paper reports improved narrative coherence and visual execution over existing approaches.
6. Conclusion
- CineEvolve unifies Director and Critic roles in one vision-language policy and uses alternating reinforcement learning for iterative multi-shot synthesis.Role-specific rewards couple diagnostic fidelity with self-evaluative feedback.Source: §6 · p. 9
- E-GRPO calibrates policy optimization to sustain stable co-evolution.The conclusion also identifies role-specific rewards as part of this training design.Source: §6 · p. 9
- CineTraj comprises 40K samples for cinematic multi-shot video generation.Source: §6 · p. 9
- Experiments demonstrate that CineEvolve outperforms existing approaches in narrative coherence and visual execution.Source: §6 · p. 9
Limitations
CineEvolve’s feedback loop primarily diagnoses and revises storyboards, leaving subtle motion artifacts from later video synthesis outside its current scope.
Limitations
- Defect diagnosis and targeted revision primarily operate at the storyboard level.Source: Limitations · p. 9
- Subtle motion artifacts arising during subsequent video synthesis remain outside the current feedback loop.Source: Limitations · p. 9
- The authors expect these residual artifacts to diminish naturally as foundational video generation backbones advance.Source: Limitations · p. 9