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%.
Figures & tables
Figure 1: Visual storyboards generated by AutoStudio ( Cheng et al., 2025 ) , Gemini-3.1-pro, and CANVAS on a multi-shot museum heist story from our HardContinuityBench. Prior approaches exhibit character appearance drift, unstable scene geometry across consecutive shots, and logical inconsistencies such as artifact state changes, room layout drift across non-consecutive shots. In contrast, CANVAS preserves character identity, stable room layout, artifact state consistency, and object persistence across shots. We use similar colors in the shot descriptions highlighting that those should maintain identical appearance throughout the entire storyline.
Method
Char
BG
Future
Multi-
Training
Consist
Consist
Cond
Agent
Free
StoryGen ( Zhang et al., 2025b )
✓
✗
✗
✗
✗
StoryDiffusion ( Zhou et al., 2024 )
✓
✗
✗
✗
✗
Story2Board ( Dinkevich et al., 2025 )
✓
✗
✗
✗
✓
Story-Iter ( Mao et al., 2026 )
✓
Partial
✗
✗
✓
AutoStudio ( Cheng et al., 2025 )
✓
✗
✗
✓
✓
Table 1: Comparison of existing Story Visualization methods. Most prior methods focus on preserving character consistency across frames using reference images or attention mechanisms. However, these methods do not explicitly maintain recurring location representations or future-conditioned scene planning. Our multi-agent training-free approach tackles these challenges.
Figure 2: Overview of CANVAS Pipeline. The framework generates a sequence of visually coherent cinematic keyframes from a story script through global planning, sequential generation, and persistent visual memory.
Dataset
Method
Background Consistency
Props
Character
Char Avg
Consecutive
Non-Consecutive
PropCons
FaceCons
ClothCons
BodyCons
ViStoryBench
AutoStudio ( Cheng et al., 2025 )
3.42
3.68
3.51
3.70
3.50
3.80
3.67
( Zhuang et al., 2025 )
Story-Iter ( Mao et al., 2026 )
3.66
3.85
3.60
3.80
3.60
3.90
3.77
Story2Board ( Dinkevich et al., 2025 )
3.20
3.43
3.18
3.60
3.40
3.60
3.53
Gemini-CT ( Comanici et al., 2025 )
4.52
4.00
4.56
4.05
4.12
3.98
4.05
CANVAS (Ours)
4.83
4.83
4.88
4.25
4.52
4.12
4.30
Table 2: Comparison of storyboard generation systems under the ContinuityEval framework using gemini-2.5-Flash as Judge. Background consistency is reported separately for consecutive shot transitions and non-consecutive scene reappearances. PropCons measures identity consistency of movable objects. Character continuity evaluates identity preservation using facial appearance, clothing attributes, and hair/body cues. The final column reports the average character consistency . To avoid the risk of circularity/bias in generation and evaluation, we also compare methods using GPT5 as Judge and report results in Table 15 .
Figure 3: Human pairwise preference win-rates of CANVAS against baseline storyboard generation systems across four continuity dimensions: background, props, characters, and overall coherence. Win-rates are computed on 240 side-by-side comparison of scenes and aggregated across 15 annotators.
Method
CSD ↑
CIDS ↑
PA ↑
CM ↑
Inc ↑
Aes ↑
Cross
Self
Cross
Self
Avg
Story-Iter
0.48
0.67
0.60
0.71
3.39
57.34
14.33
4.55
Story2Board
0.23
0.45
0.35
0.39
2.92
44.44
11.93
4.56
AutoStudio
0.34
0.63
0.44
0.46
3.39
56.22
14.34
3.45
Gemini-CT
0.38
0.62
0.58
0.65
3.61
59.97
12.50
5.54
CANVAS
0.49
0.72
0.61
0.71
3.60
60.34
13.89
5.60
Table 3: Comparison on ViStoryBench-Lite. Metrics include Style Diversity (CSD), Character Identity Similarity (CIDS), Prompt Alignment (PA), OCCM (CM), Inception (Inc), and Aesthetic score (Aes).
Method
FaceCons
ClothCons
BodyCons
Char Avg
CANVAS (Full)
4.18
4.45
4.08
4.24
– Canonical Character Anchor
2.55
2.76
2.47
2.59
– Character Anchor Retrieval
4.10
4.08
3.96
4.05
– Character Memory Update
4.09
4.12
3.89
4.03
Table 4: Character-memory ablation for CANVAS . Removing canonical anchors causes the largest degradation in identity consistency, while disabling retrieval or memory updates leads to moderate drops.
Method
Consecutive
Non-Consecutive
Prop Cons
CANVAS (Full)
4.94
4.73
4.73
– Location Grouping
2.45
3.34
4.45
– Future Planning
4.76
4.11
3.55
– Background Reuse
2.12
2.44
4.54
– Prop State Update
4.34
4.33
3.76
Table 5: Ablation analysis of CANVAS backgroud and prop-state memory components. Removing location grouping and background reuse severely degrades scene consistency across shots, while disabling prop planning or state updates reduces object state consistency in the narrative.
Figure 4: Qualitative Examples of Train Story from HardContinuityBench along with ablation showing the importance of Background and Character Memory Update in CANVAS.
Appendix figures & tables33 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Qualitative Examples of Train Story (Zoomed-in) from HardContinuityBench along with ablation showing the importance of Background and Character Memory Update in CANVAS.
Face Identity Evaluation Prompt. You are given two storyboard frames, Frame A and Frame B. The same character is expected to appear in both frames. Determine whether the faces correspond to the same individual. Focus only on the specified character. Ignore other people and background objects. Compare the following facial attributes: facial shape, eye shape and spacing, nose structure, jawline and chin structure, facial proportions, skin tone, and facial hair if present. Ignore changes caused by camera angle, facial expression, lighting conditions, or partial occlusion. Assign a score using the following rules: 5 — All visible facial features match. Facial structure, proportions, and distinctive features are consistent across both frames. 4 — All major facial features match. Minor variations appear only in rendering detail. 3 — At least half of the visible facial attributes match, but other attributes differ or are unclear. 2 — Fewer than half of the facial attributes match. 1 — Most visible facial attributes do not match. If the face of the character is not visible in one or both frames, return null .
Appendix
Table 6: Prompt used for Calculating Facial Identity Consistency.
Clothing Consistency Evaluation Prompt. You are given two storyboard frames containing the same character. Evaluate whether the clothing worn by the character is consistent across the frames. Focus only on the clothing worn by the character. Compare the following attributes: garment category (coat, jacket, shirt, dress, uniform), clothing structure (sleeve length, coat length, collar type), clothing layers (for example jacket over shirt), dominant colors, patterns or logos, and fabric appearance. Ignore differences caused only by lighting changes or cloth wrinkles. Assign a score using the following rules: 5 — Garment type, structure, color palette, and patterns are identical. 4 — Garment type and structure are identical, and color palette differs only slightly. 3 — Garment type matches but structural details or colors differ. 2 — Garment type differs but some colors or patterns match. 1 — Garment type, structure, and colors are completely different.
Appendix
Table 7: Prompt used for Calculating Clothing Similarity Consistency.
Statistic
ViStoryBench-Lite
ST-Bench
HardContinuityBench
Stories
20
30
5
Recurring Location Reappearances
Rare
Occasional
Frequent
Avg. Gap Between Location Reappearances (shots)
1.4
2.3
5.6
Character Appearance Changes
Rare
Limited
Common
Prop State Transitions (per story)
0.8
1.2
3.6
Long-Range Scene Reconstruction Required
Low
Moderate
High
Appendix
Table 8: Dataset statistics comparing HardContinuityBench with existing storyboard benchmarks. HardContinuityBench contains longer narratives, more recurring environments, larger temporal gaps between scene reappearances, and more object state transitions. These properties require models to maintain persistent world representations of characters, locations, and objects rather than relying only on short-range frame conditioning.
Background Architecture Consistency Evaluation Prompt. You are an expert evaluator for cinematic storyboard continuity. You are given two consecutive storyboard frames from the same scene in a narrative sequence. Your task is to evaluate Background Geometry Consistency (GeomCons) between the two frames. GeomCons measures whether the architectural structure of the environment remains logically coherent across shots. Your evaluation must focus only on architectural elements and determine whether the spatial layout of the environment remains consistent between the two frames. — Important Instructions Do NOT evaluate the following: - characters - character pose or movement - facial expression - lighting differences - color shifts - camera blur - small objects or movable props - decorations These elements are irrelevant to this evaluation. Your focus must be only architectural elements that define the physical structure of the environment. — Definition of Architectural Elements Architectural elements are fixed structural components of the environment that define the geometry of the scene. Examples include: - walls - doorways - doors - windows - pillars - columns - arches - staircases - floors - ceilings - railings - architectural partitions - built-in counters - permanent architectural installations These elements define the structural layout and geometry of the environment. Do NOT include movable objects such as chairs, books, tools, decorations, or equipment. — Step 1 — Identify Architectural Elements in Frame A Inspect Frame A (the earlier frame). List all architectural elements visible in the scene. Examples: - left wall - right wall - doorway - window - pillar - staircase - railing - floor boundary - ceiling beam Let the number of architectural elements be: — Step 2 — Identify Architectural Elements in Frame B Inspect Frame B (the later frame). List all architectural elements visible in the scene. Let the number of elements be: — Step 3 — Match Corresponding Architectural Elements Match architectural elements that appear in both frames. Example: | Frame A | Frame B | |——–|——–| | left wall | left wall | | doorway | doorway | | pillar | pillar | Let the number of matched elements be: — Step 4 — Detect Structural Violations Evaluate the following violations. — Appearance Violations Architectural elements appearing in Frame B that were not visible or implied in Frame A. Count the number of such cases: — Disappearance Violations Architectural elements visible in Frame A that disappear in Frame B without plausible explanation. Count the number of such cases: — Layout Violations Architectural elements that shift position or orientation in physically impossible ways. Count the number of such cases: — Step 5 — Evaluate Camera Motion Plausibility Camera movement may explain some changes. Acceptable explanations include: - camera pan - camera zoom - camera reframing - camera moving closer - camera moving sideways If the change cannot be explained by camera motion, it should be counted as a violation. — Geometry Consistency Percentage First compute the total number of violations: Total Violations = Nappear + Ndisappear + Nlayout Normalize by the total number of architectural elements: Max Elements = max( NA , NB ) Compute the Geometry Consistency Percentage: GeomCons = 100 × (1 minus Total Violations / Max Elements) Clamp the result between 0 and 100. — Score Conversion (Percentage → Rubric) Convert the percentage to a GeomCons score from 1–5. | Percentage Range | Score | Interpretation | |——————|——|—————-| | 90–100% | 5 | Perfect architectural consistency | | 75–89% | 4 | Minor geometric differences | | 55–74% | 3 | Moderate inconsistencies | | 30–54% | 2 | Major structural inconsistencies | | 0–29% | 1 | Severe spatial inconsistency |
Appendix
Table 9: Prompt for Background Geometry Consistency Evaluation (In Consecutive and Non-Consecutive Shots).
Component
Description
Evaluation Goal
Evaluate Prop Consistency (PropCons) between two storyboard frames by measuring how many props visible in one frame can be correctly recovered in the other frame with identical visual identity.
Evaluation Scope
The evaluation must rely only on visually observable information . Objects that are not visible in one of the frames must not be considered.
Visible Prop Constraint
If a prop is hidden due to camera angle, cropping, or occlusion, it must not be assumed to exist in the other frame. Only props that are clearly visible in both frames can be evaluated.
Matching Criteria
Two props are considered identical only if the following attributes match: shape, color, and texture/material appearance.
Ignored Factors
The evaluation ignores character identity, character motion, lighting differences, motion blur, and architectural structures.
Definition of Props
Props are movable non-structural objects present in the scene. Examples include chairs, tables, lamps, books, tools, decorations, statues, paintings, containers, and equipment.
Appendix
Table 10: Prop Consistency (PropCons) evaluation protocol used in ContinuityEval. The evaluator measures the recoverability of movable scene props between two frames based strictly on visible object identity.
Hair and Body Appearance Evaluation Prompt. You are given two storyboard frames containing the same character. Evaluate whether the character’s hairstyle and body-level appearance are consistent. Compare the following attributes: hairstyle, hair length, hair color, hair texture, body silhouette or build, and wearable accessories such as glasses, hats, helmets, or backpacks. Ignore differences caused by body pose, camera viewpoint, or partial occlusion. Assign a score using the following rules: 5 — Hairstyle, hair color, body silhouette, and accessories are identical. 4 — Hairstyle and body silhouette are identical, with only minor accessory or lighting differences. 3 — Either hairstyle or body silhouette matches, but the other differs. 2 — Both hairstyle and body silhouette differ but one accessory matches. 1 — Hairstyle, body silhouette, and accessories are all different.
Appendix
Table 11: Prompt used for Calculating HairStyle and Body Cues Consistency.
Dataset
Method
Background Consistency
Props
Character
Char Avg
Consecutive
Non-Consecutive
PropCons
FaceCons
ClothCons
BodyCons
ViStoryBench
AutoStudio ( Cheng et al., 2025 )
3.42
3.68
3.51
3.70
3.50
3.80
3.67
( Zhuang et al., 2025 )
Story-Iter ( Mao et al., 2026 )
3.66
3.85
3.60
3.80
3.60
3.90
3.77
Story2Board ( Dinkevich et al., 2025 )
3.20
3.43
3.18
3.60
3.40
3.60
3.53
Gemini-CT ( Comanici et al., 2025 )
4.52
4.00
4.56
4.05
4.12
3.98
4.05
CANVAS w/o QA
4.79
4.78
4.84
4.12
4.30
4.01
4.14
Appendix
Table 12: Comparison of storyboard generation systems under the ContinuityEval framework using gemini-2.5-Flash as Judge. Background consistency is reported separately for consecutive shot transitions and non-consecutive scene reappearances. PropCons measures identity consistency of movable objects. Character continuity evaluates identity preservation using facial appearance, clothing attributes, and hair/body cues. The final column reports the average character consistency . We additionally report CANVAS w/o QA , which removes the candidate verification module. Even without QA-based selection, CANVAS remains competitive with or stronger than prior baselines, indicating that the gains largely arise from the continuity planning and memory mechanisms.
Figure 6: Effect of generating N candidates on the ContinuityEval metrics.
Figure 7: Qualitative Ablation showing the importance of Prop-State Tracking in CANVAS.
Method
Generation Paradigm
CharacterCons
Environment
Object State Tracking
Interactive Editing
Explicit Narrative
Long-Range
StoryGen
Autoregressive diffusion
✓
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
StoryDiffusion
Attention-based diffusion consistency
✓
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
AutoStory
Layout-guided story generation
✓
△
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
Story2Board
Training-free storyboard prompting
✓
△
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
Story-Iter
Iterative reference refinement
✓
△
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
✓
TheaterGen
LLM prompt-book + diffusion
✓
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
✓
{\color[rgb]{0.7,0,0}✗}
{\color[rgb]{0.7,0,0}✗}
Appendix
Table 13: Positioning of CANVAS relative to prior visual storytelling and multi-turn image generation frameworks. Most existing approaches enforce consistency through reference images or attention mechanisms, focusing primarily on character identity. In contrast, CANVAS explicitly models the evolving narrative world state , maintaining persistent representations of characters, environments, and object states across shots to enable long-range story continuity. △ indicates partial or implicit support.
Model
Background
Character
Props
Gemini-2.5-Flash
0.71
0.68
0.67
Gemini-3-Pro-Image
0.65
0.61
0.62
GPT-5
0.70
0.66
0.68
Appendix
Table 14: Pearson correlation between human expert ratings and VLM-based autoraters across three evaluation dimensions. Scores are computed on 50 randomly sampled storyboard transitions from ViStoryBench and ST-Bench.
Figure 8: Streamlit-Based Interface to Collect Human Preference on the Stories (Shown Side-by-Side) Generated by CANVAS and Other Baselines.
Method
Background Consistency
Props
Character Consistency
Char Avg
Consec
Non-Consec
PropCons
FaceCons
ClothCons
BodyCons
AutoStudio ( Cheng et al., 2025 )
3.28
3.48
3.36
3.58
3.38
3.68
3.55
Story-Iter ( Mao et al., 2026 )
3.58
3.73
3.48
3.78
3.52
3.82
3.71
Story2Board ( Dinkevich et al., 2025 )
3.12
3.32
3.02
3.50
3.28
3.52
3.43
Gemini-CT ( Comanici et al., 2025 )
4.30
3.80
4.20
3.92
3.67
3.88
3.82
CANVAS (Ours)
4.60
4.60
4.52
4.11
4.39
4.02
4.18
Appendix
Table 15: ContinuityEval results using GPT-5 as the judge (To avoid the risk of circularity/bias of generation and evaluation using same model). Using GPT-5 as an independent evaluator, CANVAS consistently improves continuity across all dimensions compared to the strongest baseline (Gemini-CT). Background continuity improves from 4.30 → 4.60 (+7.0%) for consecutive scenes and from 3.80 → 4.60 (+21.1%) for non-consecutive scene reappearances, indicating substantially stronger long-range environment consistency. Prop persistence also improves from 4.20 → 4.52 (+7.6%), suggesting that CANVAS better maintains object states across shots. Character consistency also improves substantially, with the average character score increasing from 3.82 → 4.18 (+9.4%), indicating more stable identity preservation across shots. Overall, the results show that CANVAS provides the largest gains in non-consecutive background consistency, highlighting its ability to maintain coherent environments across distant shots—one of the most challenging aspects of long-range visual storytelling.
Method
CSD ↑
CIDS ↑
Prompt Alignment (PA) ↑
CM ↑
Inc ↑
Aes ↑
CP ↑
Cross
Self
Cross
Self
Scene
Shot
CI
IA
Avg
Story-Iter
0.48
0.67
0.60
0.71
3.50
3.67
3.31
3.11
3.39
57.34
14.33
4.55
0.30
Story2Board
0.23
0.45
0.35
0.39
3.12
2.34
3.0
3.23
2.92
44.44
11.93
4.56
0.24
AutoStudio
0.34
0.63
0.44
0.46
3.34
3.55
3.56
3.11
3.39
56.22
14.34
3.45
0.235
Gemini-CT
0.38
0.62
0.58
0.65
3.46
3.82
3.96
3.20
3.61
59.97
12.50
5.54
0.244
CANVAS
0.49
0.72
0.61
0.71
3.50
3.75
3.23
3.9
3.60
60.34
14.89
5.60
0.312
Appendix
Table 16: Comparison on ViStoryBench-Lite across story-image and multimodal generation methods. Metrics include Cross-Shot Diversity (CSD), Cross-Image Dependency Score (CIDS), Prompt Alignment (PA), Consistency Metric (CM), Inconsistency (Inc), Aesthetic score (Aes), and Character Preservation (CP). Gemini-CTo results are taken directly from the benchmark evaluation, and CANVAS uses the same evaluation protocol.
Method
SF
NC
VQ
CC
PLC
CT
CD
OQ
Avg
StoryIter ( Mao et al., 2026 ) + Veo-3.1
3.20
2.60
3.40
1.80
4.40
2.80
4.00
2.80
3.13
AutoStudio ( Cheng et al., 2025 ) + Veo-3.1
3.00
2.80
3.20
2.00
3.50
3.00
3.50
3.20
3.03
Gemini-CT ( Zhang et al., 2025a ) + Veo-3.1
3.00
2.20
3.60
2.20
3.80
3.60
3.80
4.40
3.33
Story2Board ( Dinkevich et al., 2025 ) + Veo-3.1
3.60
2.20
1.80
1.00
3.60
2.20
3.40
3.20
2.63
CANVAS (Ours) + Veo-3.1
3.40
4.00
3.80
4.00
3.80
4.00
4.40
4.80
4.03
Appendix
Table 17: Comparison of the videos generated by each of the baselines and CANVAS paired with a standard I2V model across eight evaluation metrics proposed by FilMaster ( Huang et al., 2025 ) using Gemini-2.5-Flash as the judge. We evaluate videos generated using the Veo-3.1-preview Image-to-Video (I2V) model, where the first frame is initialized with the keyframe produced by each storyboard generation baseline. To ensure fair comparison, we keep the video prompt identical across methods, varying only the keyframe input. Each prompt–keyframe pair is used to generate short clips, which are then concatenated to produce a final video of approximately 1.5 minutes. Metrics are abbreviated as follows: SF (Script Faithfulness), NC (Narrative Coherence), VQ (Visual Quality), CC (Character Consistency), PLC (Physical Law Compliance), CT (Cinematic Techniques), CD (Compelling Degree), and OQ (Overall Quality). CANVAS achieves the strongest overall performance, particularly in narrative coherence, character consistency, cinematic techniques, and compelling storytelling, demonstrating the benefits of continuity-aware generation that maintains consistent characters, environments, and object states across shots.
Method
SF
NC
VQ
CC
PL
CT
CD
OQ
Story2Board ( Dinkevich et al., 2025 ) + Veo-3.1
2.25
1.75
1.75
1.00
3.00
2.75
2.75
2.25
Story-Iter ( Mao et al., 2026 ) + Veo-3.1
3.75
2.25
3.75
2.00
3.75
4.50
4.25
4.25
Gemini-CT + + Veo-3.1
3.75
2.25
4.25
2.50
3.50
4.25
4.00
4.00
CANVAS (Ours) + Veo-3.1
4.00
3.00
4.75
4.00
3.25
5.00
4.75
4.75
Appendix
Table 18: Comparison of the videos generated by each of the baselines and CANVAS paired with a standard I2V model across eight evaluation metrics proposed by FilMaster ( Huang et al., 2025 ) using Gemini-2.5-Pro as the judge. We evaluate videos generated using the Veo-3.1-preview Image-to-Video (I2V) model, where the first frame is initialized with the keyframe produced by each storyboard generation baseline. To ensure fair comparison, we keep the video prompt identical across methods, varying only the keyframe input. Each prompt–keyframe pair is used to generate short clips, which are then concatenated to produce a final video of approximately 1.5 minutes. Metrics are abbreviated as follows: SF (Script Faithfulness), NC (Narrative Coherence), VQ (Visual Quality), CC (Character Consistency), PLC (Physical Law Compliance), CT (Cinematic Techniques), CD (Compelling Degree), and OQ (Overall Quality). CANVAS achieves the strongest overall performance, particularly in narrative coherence, character consistency, cinematic techniques, and compelling storytelling, demonstrating the benefits of continuity-aware generation that maintains consistent characters, environments, and object states across shots.
Method
Model
CSD ↑
CIDS ↑
PA ↑
CM ↑
Inc ↑
Aes ↑
CP
Cross
Self
Cross
Self
Scene
Shot
CI
IA
Avg.
Copy-Paste Baseline
-
0.735
0.770
0.911
0.993
0.62
2.08
0.87
1.55
1.28
92.76
5.46
4.39
0.550
Story Image Method
StoryGen ( Liu et al., 2024 )
SD1.5
0.405
0.562
0.405
0.591
1.05
2.32
1.43
1.75
1.64
52.98
7.15
4.09
0.277
StoryGen ( Liu et al., 2024 )
SD1.5
0.396
0.551
0.396
0.602
1.18
2.25
1.49
1.50
1.60
52.53
7.67
4.09
0.224
StoryGen ( Liu et al., 2024 )
SD1.5
0.316
0.617
0.316
0.610
0.89
2.41
1.39
2.05
1.68
40.13
6.25
3.86
0.240
Appendix
Table 19: Benchmark comparison across story image and multimodal large models on ViStoryBench-Lite ( Zhuang et al., 2025 ) . We copy the table from the paper to report numbers of the existing baselines (Only Story-Image and Open-source Models excluding the StoryVideo methods and Commercial Platforms) and compare CANVAS using the same evaluation suite released by the paper.
Figure 9: Qualitative comparison of cross-shot continuity with and without background Planning. The first row shows outputs from CANVAS with structured planning , which performs location canonicalization and prop forecasting before shot generation. The second row shows results without background planning . While both depict the same narrative interaction, in the first shot, the fireplace remains occluded and also the toy was present at the background. Hence, the next shot just looks like a zoomed-in version of Santa, maintaining higher narrative continuity.
Figure 10: The figure compares storyboards generated by recent training-free methods—Story2Board, AutoStudio, Story-Iter, Gemini-CT—and CANVAS for a multi-shot narrative set across a robotics lab and hallway. Each row shows a method and each column shows a shot, illustrating how well character identity, environments, and object states are maintained across the sequence. Existing methods produce visually plausible frames but suffer from identity drift, layout inconsistencies, and unstable object states, especially when scenes reappear after transitions. In contrast, CANVAS preserves consistent character appearance, stable environments, and coherent object interactions throughout the narrative by using global continuity planning and world-state memory tracking. Overall, the figure shows that explicitly modeling characters, locations, and object states enables CANVAS to maintain strong long-range visual coherence across multi-shot storyboards.
Prompt: Location Clustering for Background Continuity
You are a narrative continuity planner for visual storytelling. Your goal is to identify and cluster recurring locations across a sequence of storyboard shots.
Instructions: 1. Identify the physical environment in which each shot occurs. 2. Assign a unique location_id to each distinct location. 3. Cluster shots that occur in the same environment under the same location_id . 4. If a shot returns to a previously seen environment, mark it as location_reappearance . 5. If a shot directly continues the previous scene without changing environment, mark it as previous_frame_continuation . 6. If the location appears for the first time, mark it as fresh_location . Return STRICT JSON in the following format: { "locations": { "museum_gallery": [ "Scene_1_Shot_1", "Scene_1_Shot_2", "Scene_3_Shot_1" ], "security_room": [ "Scene_2_Shot_1" ] }, "shots": { "Scene_1_Shot_1": { "location_id": "museum_gallery", "continuity_mode": "fresh_location" } } } Constraints • Maintain consistent location identities across shots. • Do not create duplicate location IDs for the same environment. • When a location reappears later in the story, reuse the original location_id .
Appendix
Table 20: Prompt used for clustering shots into recurring locations to maintain background continuity across the storyboard.
Prompt: Future Prop and Geometry Aware Background Planning
You are a story continuity planner responsible for maintaining consistent environments across storyboard shots. Your task is to reason about how the background scene should evolve based on the current and future states of important props and architectural geometry of the background location.
Instructions Given: • The description of the current shot. • The planned metadata for the shot (characters, location, prop states). • The previously known world state of props and environments. Plan the background scene so that it remains logically consistent with both the current story state and upcoming events. Reason about the following: 1. Which props should remain visible in the background environment. 2. Which props should disappear because they were taken, destroyed, or moved. 3. Which props are likely to appear in the background due to upcoming story events. 4. Which props are currently carried by characters and therefore should not remain in the environment. 5. Which environmental objects must persist across shots to maintain scene identity (e.g., furniture, structures, display cases). 6. Whether the camera framing hides some props even though they still exist in the environment. Return STRICT JSON in the following format: { "background_props": [ "display_case", "museum_sign", "security_camera" ], "must_appear": [ "golden_artifact" ], "must_not_appear": [ "artifact_case_glass" ], "carried_props": { "thief": ["golden_artifact"] }, "reasoning": "The artifact is carried away by the thief, so it should not remain in the gallery background. The broken display case should remain visible." } Constraints • Background elements should remain consistent with previously established environments. • If a prop was present earlier in the same location and nothing removed it, it should persist. • Props carried by characters must not remain in the background. • The planned background should support future story events.
Appendix
Table 21: Prompt used for future prop-aware background planning to ensure that scene environments evolve consistently with object state transitions.
Prompt: Character Appearance State Planning (Single Character)
You are a narrative continuity planner responsible for tracking the visual appearance of a single character across a sequence of storyboard shots. Your goal is to determine the appearance state of the character in each shot and ensure that the appearance evolves consistently with the story.
Input • The name of the target character. • A sequence of storyboard shot descriptions. Instructions 1. For each shot, determine whether the target character appears in the scene. 2. If the character appears, determine the character’s visual appearance state (e.g., clothing, uniform, disguise). 3. If the appearance does not change, reuse the same appearance state identifier as the previous shot. 4. If the story explicitly changes the character’s clothing or visual style, assign a new appearance state. 5. If the character does not appear in a shot, mark the state as not_present . Return STRICT JSON in the following format: { "character": "Ethan", "appearance_timeline": { "Scene_1_Shot_1": "blue_curator_jacket", "Scene_1_Shot_2": "blue_curator_jacket", "Scene_1_Shot_3": "blue_curator_jacket", "Scene_2_Shot_1": "black_security_jacket", "Scene_2_Shot_2": "not_present", "Scene_3_Shot_1": "black_security_jacket" } } Constraints • Maintain consistent appearance across shots unless the story explicitly indicates a change. • Reuse appearance identifiers whenever the clothing or visual style remains the same. • Ensure that later reappearances match the most recent appearance state.
Appendix
Table 22: Prompt used during Character Appearance Planning.
Prompt: Prop State Update Planning (Single Object)
You are a narrative continuity planner responsible for tracking the state of a single prop (object) across a sequence of storyboard shots. Your goal is to determine how the object’s state evolves throughout the story and ensure logical consistency with the described events.
Input • The name of the target prop (object). • A sequence of storyboard shot descriptions. Instructions 1. For each shot, determine whether the target prop appears in the scene. 2. If the prop appears, determine its state (e.g., intact, broken, inside container, carried by character, missing). 3. If the prop is carried by a character, record which character carries it. 4. If the prop is removed from the scene (e.g., stolen, destroyed, moved), update its state accordingly. 5. If the prop continues unchanged across shots, reuse the same state identifier. 6. If the prop does not appear in a shot but still exists in the story world, mark it as not_visible . Return STRICT JSON in the following format: { "prop": "golden_artifact", "state_timeline": { "Scene_1_Shot_1": { "state": "inside_display_case", "carrier": null }, "Scene_1_Shot_2": { "state": "inside_display_case", "carrier": null }, "Scene_1_Shot_3": { "state": "carried", "carrier": "masked_thief" }, "Scene_2_Shot_1": { "state": "not_visible", "carrier": "masked_thief" }, "Scene_3_Shot_1": { "state": "missing_from_gallery", "carrier": "masked_thief" } } } Constraints • Maintain consistent prop states across shots unless the story explicitly changes them. • If a prop is carried by a character, it should not remain in the previous location. • If a prop disappears from the environment, update its state to reflect the narrative event. • When the prop reappears later, ensure its state matches the last known world state.
Appendix
Table 23: Prompt used during global planning to determine the state timeline for a single prop across shots.
Prompt: Previous-Frame Continuation Decision
You are a visual continuity planner for storyboard generation. Your task is to decide whether the current shot is a direct continuation of the previous shot, such that the previous frame should be reused as a visual anchor for generation.
Input • The description of the previous shot. • The description of the current shot. • The location assignment of both shots. • The characters present in both shots. • The prop states in both shots. Instructions Determine whether the current shot should be treated as previous_frame_continuation , location_reappearance , or fresh_location . Mark the current shot as previous_frame_continuation only if the current shot depends on preserving the spatial structure of the immediately previous frame. Reason about the following: 1. Whether the previous shot and current shot occur in the same physical environment. 2. Whether the current shot is an immediate temporal continuation rather than a later revisit. 3. Whether the spatial arrangement of the scene should remain consistent across the cut. 4. Whether important background geometry must be preserved, such as walls, doors, furniture, display cases, tables, platforms, or room layout. 5. Whether character positions, object placements, or scene composition depend on the previous frame. 6. Whether the current shot is a close-up, zoom-in, alternate camera angle, or tighter crop of the same ongoing scene. 7. Whether there are state changes in the scene that still require maintaining the same base spatial environment. Use the following rules: • Choose previous_frame_continuation if the scene is still unfolding in the same environment and preserving the exact spatial setup from the previous frame is important. • Choose location_reappearance if the shot returns to a known location after an intervening scene or temporal gap, where the same location identity should be preserved but not necessarily the exact previous composition. • Choose fresh_location if the shot occurs in a new environment not previously shown. • Even if some props or characters change state, still choose previous_frame_continuation when the underlying scene geometry should remain continuous. Return STRICT JSON in the following format: { "continuity_mode": "previous_frame_continuation", "same_location": true, "requires_spatial_continuity": true, "spatial_dependencies": [ "display_case remains in same position", "same museum gallery layout", "artifact remains centered under spotlight" ], "reasoning": "..." } Constraints • Base the decision on scene continuity, not only on shared characters or shared location name. • Prioritize whether preserving spatial arrangement from the immediately previous frame will improve visual continuity. • If the current shot is compositionally dependent on the previous frame, prefer previous_frame_continuation .
Appendix
Table 24: Prompt used to decide whether the current shot should reuse the previous frame as a continuation anchor.
pt=P(st)
Appendix
Algorithm 2 Candidate Anchor Retrieval and Prompt Construction
Prompt: Candidate Frame Generation
You are a cinematic storyboard generator. Your task is to generate a candidate frame for the current shot while maintaining visual continuity with previously generated frames and respecting the global story plan.
Input • Shot description st • Retrieved visual anchors At (may include previous frame, character anchors, background anchors, and prop anchors) • Character appearance states from the global plan • Prop states and prop reasoning for the current shot • Location identifier and background reasoning Instructions Generate a visually coherent frame for the current shot by following these rules: 1. Respect the shot description Ensure the generated frame faithfully depicts the actions, characters, and environment described in the current shot. 2. Preserve character identity Characters must maintain the appearance state specified by the global plan. If character anchors are provided, match the visual identity (face, clothing, body appearance) shown in those anchors. 3. Maintain spatial continuity If the previous frame is provided as an anchor, treat the current shot as a continuation of the same scene. Preserve the spatial layout of the environment, including background structure, object placement, and character positions. 4. Use background anchors If background anchors are provided, maintain the same environment geometry (e.g., walls, furniture, display cases, room layout). 5. Respect prop states Ensure that props appear according to the planned prop states: • Props that must appear should be clearly visible. • Props carried by characters should move with the corresponding character. • Props that were removed, destroyed, or taken away should not remain in the environment. 6. Maintain cinematic coherence The frame should resemble a realistic cinematic shot with consistent lighting, perspective, and scene composition. Output Generate a high-quality candidate frame that satisfies the shot description and respects all provided anchors and continuity constraints.
Appendix
Table 25: Prompt used to generate candidate storyboard frames using retrieved anchors and global continuity constraints.
pt=P(st)
Appendix
Algorithm 3 VLM-Based QA Candidate Selection with Scoring Prompt
Prompt: VLM-Based Continuity Scoring Prompt
You are a visual continuity evaluator. Your task is to evaluate a candidate storyboard frame and assign scores based on how well it satisfies the global continuity plan and the current shot description.
Input • Current shot description st • Candidate image It • Character appearance states from the global plan • Prop states for the current shot • Location identifier for the scene • Previous frame It−1 (if continuation) Evaluation Criteria Evaluate the image along the following dimensions and assign a score from 1 to 5. 1. Shot Alignment (1–5) How well does the image depict the action described in the shot? 2. Character Appearance Consistency (1–5) Do characters match the planned appearance states? 3. Background Continuity (1–5) Does the environment match the planned location and maintain spatial continuity with the previous frame? 4. Prop State Correctness (1–5) Do props appear with the correct state according to the global plan? Return STRICT JSON { "shot_alignment": 5, "character_consistency": 4, "background_continuity": 5, "prop_state_correctness": 4, "overall_score": 4.5, "reasoning": "The frame correctly depicts the museum gallery..." } Scoring Guidelines • 5 = perfect match with the plan • 4 = minor inconsistencies • 3 = moderate inconsistencies • 2 = major inconsistencies • 1 = completely incorrect
Appendix
Table 26: Prompt used by the VLM to score candidate storyboard frames during QA-based candidate selection (Algorithm 3 ).
Mc(c,C(c,t))←anchorc
Appendix
Algorithm 4 Visual Memory Update with Anchor Extraction
Prompt: Character Anchor Extraction
You are extracting visual anchors for characters from a storyboard frame. Input • Frame image It • Shot description st • Character appearance states C(c,t) Task Identify characters visible in the frame and extract a clean visual anchor image that captures their current appearance (face, clothing, hairstyle, and overall identity). Return STRICT JSON { "characters":[ { "name":"Ethan", "appearance_state":"trench_coat", "anchor":"character_anchor_image" }, { "name":"Lena", "appearance_state":"blue_dress", "anchor":"character_anchor_image" } ] } Guidelines • Extract anchors only for characters clearly visible. • Anchors should capture the full appearance identity.
Appendix
Table 27: Prompt for extracting character appearance anchors for updating character memory.
Prompt: Location Background Anchor Extraction
You are extracting the visual environment anchor for a location from a storyboard frame. Input • Frame image It • Shot description st • Location identity L(st) Task Identify the environment corresponding to the location and extract a clean background anchor image that represents the spatial layout of the scene. Return STRICT JSON { "location":"museum_gallery", "anchor":"background_anchor_image" } Guidelines • The anchor should capture the environment layout. • Avoid including large foreground characters when possible.
Appendix
Table 28: Prompt for extracting background anchors associated with locations.
Prompt: Prop Anchor Extraction
You are extracting visual anchors for important objects from a storyboard frame. Input • Frame image It • Shot description st • Object state descriptions O(o,t) Task Identify objects that appear in the frame and extract visual anchors representing their current state. Return STRICT JSON { "objects":[ { "name":"golden_artifact", "state":"inside_display_case", "anchor":"prop_anchor_image" } ] } Guidelines • Extract anchors only for visible objects. • Ensure the object state matches the visual evidence.
Appendix
Table 29: Prompt used to extract prop anchors and update object-state memory.
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
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.
Recent diffusion and autoregressive models have substantially improved text-to-video generation, yet producing coherent long-form story videos with consistent characters and aligned audio remains challenging. Existing methods often suffer from narrative drift, unstable character identity, weak cross-scene continuity, and audio-visual mismatch over extended sequences. We propose DramaAgent, a hierarchical, agentic, and model-agnostic framework for long-form text-to-video-and-audio generation. Rather than improving the underlying video backbone itself, DramaAgent introduces an upper-level control layer that decomposes generation into story planning, persistent character conditioning, scene-wise synthesis, and reflection-guided targeted repair. The framework maintains reusable story and character states across scenes, diagnoses failures such as identity drift, missing scene semantics, temporal discontinuity, and cross-modal mismatch, and repairs problematic clips in a stage-specific manner. Experiments across multiple video generation backbones show that DramaAgent improves long-horizon coherence, character consistency, narrative fidelity, and scene-level audio-visual consistency over direct generation and strong baselines. These results suggest that hierarchical agentic control is a practical direction for controllable long-form audiovisual generation. Code: https://github.com/AIGeeksGroup/DramaAgent. Website: https://aigeeksgroup.github.io/DramaAgent.