Generating coherent and communicative visual sequences, such as image sequences and videos, remains a significant challenge for current multimodal systems. Despite advances in visual quality and the integration of world knowledge, existing models still struggle to maintain logical flow, often resulting in disjointed actions, fragmented narratives, and unclear storylines. We attribute these issues to the lack of attention to visual logic, a critical yet underexplored dimension of visual sequence generation that we define as the perceptual and causal coherence among characters, actions, and scenes over time. To bridge this gap, we propose a logic-aware multi-image story visualization framework, LogiStory. The framework is built around the central innovation of explicitly modeling visual logic in story visualization. To realize this idea, we design a multi-agent system that grounds roles, extracts causal chains, and verifies story-level consistency, transforming narrative coherence from an implicit byproduct of image generation into an explicit modeling objective. This design effectively bridges structured story planning with visual generation, enhancing both narrative clarity and visual quality in story visualization. Furthermore, to evaluate the generation capacity, we construct LogicTale, a benchmark comprising richly annotated stories, emphasizing causal reasoning, and visual logic interpretability. We establish comprehensive automatic and human evaluation protocols designed to measure both visual logic and perceptual quality. Experiments demonstrate that our approach significantly improves the narrative logic of generated visual stories. This work provides a foundational step towards modeling and enforcing visual logic in general image sequence and video generation tasks.
Figures & tables
Figure 1: Comparison of the state-of-the-art multimodal models alongside our proposed approach LogiStory in the visualization of the simple story "The Crow and the Pitcher." The results highlight the challenges of visual reasoning in the process of visual sequence generation, while demonstrating the effectiveness of LogiStory.
Figure 2: Overview of LogiStory framework . Given an input story, our system first applies a multi-agent story planner to decompose the story into structured panels with detailed scripts. In the generation process, the Local Causal Monitor simulates a reader’s linear understanding by evaluating each frame for inconsistencies and generating refinement signals. Then, the Global Causal Verifier applies the causal graph to produce concrete refinement instructions to correct errors and maintain narrative flow.
Method
Visual Logic
Perceptual Quality
User study
ICons.
NCausal.
SRead.
AesthQ.
SCons.
CExpr.
ICons.
NCausal.
SRead.
App.
DeepSeek-R1 + SDXL
2.17
2.02
0.5675
0.2899
0.7747
2.13
2.33
1.86
2.13
3.0
DeepSeek-R1 + Flux
3.87
3.43
0.6872
0.3047
0.8329
4.20
3.53
3.25
3.20
3.8
DeepSeek-R1 + StoryDiff
3.07
2.06
0.5409
0.2909
0.8276
2.07
2.70
2.12
2.35
2.6
DeepSeek-R1 + ConSistory
3.23
1.98
0.5623
0.2939
0.8129
2.26
2.87
1.73
1.87
2.4
MM-StoryAgent
3.03
2.63
0.5787
0.2942
0.8023
2.54
2.67
3.04
2.43
2.6
Table 1: Performance of Automatic metric on story visualization methods. "ICons.", "NCausal.", "SRead.", "AesthQ.", "SCons.", "CExpr.", "App." refer to "Instance Consistency", "Narrative Causality", "Story Readability", "Aesthetic Quality", "Style Consistency", "Character Expressiveness", "Aesthetic Appeal". The first rank is highlighted in bold , while the second rank is underlined .
Figure 3: Quality analysis of generated stories across different methods. Representative key scenes are shown here. Complete sequences can be found in the Appendix A .
Figure 4: Ablation study on key components of LogiStory. Qualitative comparisons on representative examples demonstrate the impact of each module.
Table 6
Figure 5: Quality analysis on Complex Case. Memory Lasts: An old painter takes his student to a distant sunflower field, once shared with his younger brother before the war. As they paint, the narrative shifts briefly to a childhood flashback, two boys painting under the sunflower sky, one waving goodbye. Upon returning home, the student donates a commemorative painting to a local school, where it remains as the old painter’s lasting memory.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Quality analysis of generated stories across different methods. The Crow and the Pitcher: Under the scorching sun, across a vast, parched land, a thirsty crow flies in search of water. It eventually spots a tall glass bottle with a small amount of water at the bottom. Unable to reach it with its beak, the crow notices small pebbles scattered on the ground. One by one, it picks up the pebbles and drops them into the bottle. As the stones pile up, the water level slowly rises. At last, the water reaches the top, and the crow happily quenches its thirst.
Figure 7: Quality analysis of generated stories across different methods. Three Little Pigs: Three little pigs live their mother’s house and each build a house: one of straw, one of sticks, and one of bricks. A hungry wolf comes and blows down the straw and stick houses, causing the first two pigs to flee to the brick house. The wolf tries to blow it down but fails. He then climbs the chimney, but the pigs boil a pot of water inside, and the wolf falls in and runs away burned. All three pigs live safely together in the brick house.
Figure 8: Quality analysis of generated stories across different methods. The Pebble Bridge: A group of ants needs to cross a small stream. They drop pebbles into the water to form a bridge and successfully cross together.
Figure 9: Quality analysis of generated stories across different methods. Borrowing Arrows with Straw Boats: During a critical shortage of arrows, Zhuge Liang boldly promised to deliver 100,000 arrows within three days. To fulfill this promise, he ordered countless straw men to be placed on boats and, under the cover of thick fog, sailed toward Cao Cao’s camp. Mistaking the figures for soldiers, the enemy forces unleashed volleys of arrows at the boats. When the boats returned, they were laden with arrows, allowing Zhuge Liang to fulfill his promise cleverly and without shedding a drop of blood.
Figure 10: Quality analysis of generated stories across different methods. The Dog and His Reflection : A dog crosses a bridge with a bone in its mouth. Looking into the water, he sees his own reflection and mistakes it for another dog with a bigger bone. He snaps at the reflection, dropping his bone into the river and losing it.
Method
ICons. ↑
NCausal. ↑
SRead. ↑
AesthQ. ↑
SCons. ↑
CExpr. ↑
StoryGen
3.24
1.72
0.5366
0.2873
0.8162
1.92
StoryDiffusion
3.07
2.06
0.5409
0.2909
0.8276
2.07
Story-Adapter
3.65
2.23
0.5733
0.2952
0.8195
3.16
Ours
4.23
4.45
0.8267
0.3088
0.8572
4.32
Appendix
Table 4: Performance comparison of additional baselines on the LogicTale benchmark.
Figure 11: Comparative evaluation of different methods across difficulty levels.
Backbone
ICons. ↑
NCausal. ↑
SRead. ↑
Qwen2.5-VL-7B
3.04
3.32
0.6132
InternVL 2.5-8B
3.21
3.24
0.6328
Qwen2.5-VL-32B
3.85
3.78
0.7244
DeepSeek-V3
4.26
4.45
0.8267
Appendix
Table 5: Performance of LogiStory with different backbone models.
Method
CSD Self ↑
CIDS Self ↑
Alignment ↑
OCCM ↑
Inception ↑
Aesthetics ↑
GPT-4o
68.5
73.1
89.3
93.4
9.02
5.52
Gemini 2.0
58.6
53.7
76.1
86.9
10.12
4.91
Story-Adapter
70.0
62.6
38.8
82.0
10.36
5.81
Ours
64.6
69.4
92.2
93.4
10.16
5.88
Appendix
Table 6: Comparison on ViStoryBench-Lite.
Figure 12: Kendall’s Tau correlation between model rankings on full dataset (60 stories) and various dataset subsets across six evaluation metrics. Results indicate high ranking consistency with subsets as small as 36 stories, validating the reliability of LogicTale for evaluation purposes.
Dataset
#Samples
Logic Annotation
Evaluation Mode
StoryDiffusion
20
✗
Single-image
ConsiStory
20
✗
Single-image
MM-StoryAgent
100
✗
Single-image
MovieAgent
12
✗
Single-image
ViStoryBench
80
✗
Single-image
LogicTale (Ours)
60
✓
Multi-image joint
Appendix
Table 7: Comparison of LogicTale with existing story visualization benchmarks. Unlike prior datasets, LogicTale provides explicit causal logic annotations and supports multi-image joint evaluation , enabling systematic assessment of narrative coherence.
Despite recent progress in agentic multi-shot video generation, producing coherent and consistent long-form stories remains challenging. Existing agentic pipelines typically rely on textual shot plans or previously generated pixels, yet lack an explicit mechanism for propagating the consequences of story events and maintaining the video world state across shots. As a result, missing visual details may be reconstructed inaccurately, while visual drift may propagate across subsequent shots, undermining both narrative coherence and visual consistency. To address these challenges, we propose StoryEngine, a state-grounded agentic framework for video storytelling. StoryEngine establishes a separation between authoritative semantic plans and unreliable visual observations. Specifically, StoryEngine maintains a structured representation of entity placement and story-relevant states, and propagates event-induced changes to define the intended start and end states of each shot. To visually realize these states, StoryEngine constructs canonical references for recurring entities and environments, and compiles state and visual constraints into executable render plans. Meanwhile, to realize these states correctly, a bounded evaluation-guided repair loop further corrects local state inconsistencies. Together, these mechanisms preserve causal story progression and prevent local visual errors from propagating across shots. To comprehensively evaluate long-form storytelling, we construct a benchmark across diverse scenarios and visual styles, with metrics assessing storytelling quality, narrative coherence, and visual consistency. Experimental results demonstrate that StoryEngine consistently outperforms state-of-the-art methods across all evaluation dimensions, validating its effectiveness for coherent and consistent video storytelling.
Long-form visual storytelling requires maintaining continuity across shots, including consistent characters, stable environments, and smooth scene transitions. While existing generative models can produce strong individual frames, they fail to preserve such continuity, leading to appearance changes, inconsistent backgrounds, and abrupt scene shifts. We introduce CANVAS (Continuity-Aware Narratives via Visual Agentic Storyboarding), a multi-agent framework that explicitly plans visual continuity in multi-shot narratives. CANVAS enforces coherence through character continuity, persistent background anchors, and location-aware scene planning for smooth transitions within the same setting We evaluate CANVAS on two storyboard generation benchmarks ST-BENCH and ViStoryBench and introduce a new challenging benchmark HardContinuityBench for long-range narrative consistency. CANVAS consistently outperforms the best-performing baseline, improving background continuity by 21.6%, character consistency by 9.6% and props consistency by 7.6%.
Storyboard synthesis plays a crucial role in visual storytelling, aiming to generate coherent shot sequences that visually narrate cinematic events with consistent characters, scenes, and transitions. However, existing approaches are mostly adapted from text-to-image diffusion models, which struggle to maintain long-range temporal coherence, consistent character identities, and narrative flow across multiple shots. In this paper, we introduce DreamShot, a video generative model based storyboard framework that fully exploits powerful video diffusion priors for controllable multi-shot synthesis. DreamShot supports both Text-to-Shot and Reference-to-Shot generation, as well as story continuation conditioned on previous frames, enabling flexible and context-aware storyboard generation. By leveraging the spatial-temporal consistency inherent in video generative models, DreamShot produces visually and semantically coherent sequences with improved narrative fidelity and character continuity. Furthermore, DreamShot incorporates a multi-reference role conditioning module that accepts multiple character reference images and enforces identity alignment via a Role-Attention Consistency Loss, explicitly constraining attention between reference and generated roles. Extensive experiments demonstrate that DreamShot achieves superior scene coherence, role consistency, and generation efficiency compared to state-of-the-art text-to-image storyboard models, establishing a new direction toward controllable video model-driven visual storytelling.
Junjia Huang, Binbin Yang, Pengxiang Yan +6
1Sun Yat-sen University · 2ByteDance Intelligent Creation