A$^2$RD: Agentic Autoregressive Diffusion for Long Video Consistency

• Long-video synthesis must preserve entities, environments, and motion over time while ensuring narratives evolve meaningfully, challenges that remain fundamental despite strong fidelity on second-long clips.Source: §1 · p. 2
• Frame-based autoregressive models (Chen et al., 2026; Huang et al., 2025a; Yang et al., 2026a) preserve local continuity but freeze frames as conditioning, allowing errors to propagate and causing drift or repetition (Zhao et al., 2026).Source: §1 · p. 2
• Segment-based methods offer stronger narrative control but struggle with inter-segment consistency, while visual-only frame memory is insufficient to reliably track entities and narratives (Elmoghany et al., 2025).Source: §1 · p. 2
• Its Retrieve–Synthesize–Refine–Update cycle retrieves multimodal context, adaptively selects segment modes, synthesizes and self-improves boundary frames and segments, then updates memory.Source: §1 · p. 2
• A2RD achieves state-of-the-art consistency and narrative coherence in two self-improved iterations, with human-study corroboration, while scaling to ultra-long video and mitigating semantic drift and content collapse.Source: §1 · p. 2
• LVbench-C tests long-horizon consistency through cyclical entity and environment appearances, including disappearances and reappearances with optional state changes.Source: §1 · p. 2

Long-Form Video Synthesis

• Long-video methods trade local fidelity for narrative control: frame-autoregressive systems risk drift and repetition, while segment methods struggle to maintain inter-segment consistency.Extrapolation can miss details absent from the start frame, while interpolation depends on planned boundaries; uncorrected errors can propagate across segments.Source: §2 · p. 4

Test-Time Scaling for Generative Models

• Video test-time scaling has mainly optimized single-segment quality, leaving inter-segment consistency and long-range narrative progression unaddressed.Examples include RAPO's retrieval-augmented prompt refinement (Gao et al., 2025), VISTA's iterative planning and critique (Long et al., 2026), and VQQA's question-based closed-loop optimization (Song et al., 2026).Source: §2 · p. 4

Memory for LLM Agents

• A2RD's multimodal memory explicitly tracks fine-grained visual and narrative progression across modalities, addressing limits of image-based references and opaque representation memories.LLM-agent memories include text (Packer et al., 2023), hidden representations (Wang et al., 2025b), and graphs (Chhikara et al., 2025); synthesis memories include images, image–text pairs, and hidden representations.Source: §2 · p. 4

3.1. MVMem Design and Initialization

• MVMem tracks textual states, reference frames, and segment videos, retrieving relevant history and updating memory after each segment; global references are synthesized in dependency order.Textual states capture visual arcs, spatial relations, and camera states.Source: §3.1 · pp. 5–6

3.2. Agentic Auto-Regressive Generation Pipeline

• A2RD selects extrapolation for spatiotemporally continuous segments and interpolation otherwise, then retrieves context and refines boundary frames before video synthesis and memory update.Interpolation applies to multi-shot contexts spanning environments or when the next segment jumps to a new environment.Source: §3.2 · pp. 6–7

3.3. Hierarchical Boundary Frame and Video Self-Improvement

• HITS interactively self-improves synthesized frames and video segments to limit cascading temporal errors and strengthen intra- and inter-segment coherence.Frame and video verification use separate rubrics that assess consistency and basic quality.Source: §3.3 · pp. 7–8
• MAPO uses rubric-scored positive and negative refinement cases to retrieve examples, derive targeted guidelines, revise prompts, and update its case database online.A case is labeled positive when all rubric scores improve and negative when all scores worsen.Source: §3.3 · pp. 8–9

3.4. A2RD-Parallel (A2RD-Par)

• A2RD-Par synthesizes boundary frames autoregressively but generates video segments in parallel without video-level self-improvement, reducing ideal video-synthesis latency from N k_v L_V to L_V.This trades some spatial consistency for evolving environments while retaining strict character consistency and coherent stories in static environments.Source: §3.4 · p. 9

Long Video Bench-Challenge (LVbench-C)

• LVbench-C targets three transition types: evolving character states, evolving object states, and evolving environment states.It addresses non-linear transitions that existing single- or multi-scene benchmarks neglect.Source: §4 · p. 9
• Scenarios cover 3-, 5-, and 10-minute videos, with 50 samples each for 3- and 5-minute videos and 25 for 10-minute videos.The text reports 120 scenarios overall, although the listed sample counts sum to 125.Source: §4 · p. 9
• A common construction pipeline uses screenwriter-persona prompts, a human-designed demonstration, and a state-of-the-art MLLM to generate scenarios.Generation enforces meaningful story flow, scene durations, cause-and-effect, and at least n=10 absent segments before entities or environments reappear; manual review updates constraints or demonstrations for diversity.Source: §4 · pp. 9–10
• Scenario refinement deduplicates MLLM-generated content and applies six-criterion MLLM-Judge validation, with a separate MLLM re-verification and manual subset review.The criteria check specificity, logic, naturalness, realism, repetition, and contradictions; failed scenarios are refined until passing or reaching a limited number of iterations.Source: §4 · p. 10

5.1. Settings

• Experiments cover 1-minute VBench-Long videos from 40 prompts and LVbench-C videos with 24 scenes at 3 minutes and 40 scenes at 5 minutes, using 20 samples per multi-scene setting.VBench-Long prompts are decomposed into eight continuous segments; 10-minute LVbench-C videos are evaluated in Section 6.4. Automatic metrics include semantic alignment, narrative coherence, inter-segment character, environment and motion consistency, and within-segment subject and environment consistency.Source: §5.1 · p. 10, §5.2 · p. 11

5.2. Automatic Evaluations

• A2RD scores 0.9 in narrative coherence, 20% above Naive-AR’s 0.75, and improves character consistency by 30% while reaching 0.9935 motion consistency.It also reaches 0.84 environment consistency versus VideoMemory’s 0.73; both A2RD versions achieve the highest semantic alignment. A2RD-Par scores 0.81 in narrative coherence while enabling parallel generation.Source: Table 2 · p. 11, §5.2 · p. 11
• A2RD outperforms VideoMemory on average consistency by 16% at 3 minutes and 13% at 5 minutes, scores 10% higher in narrative coherence, and maintains motion smoothness above 0.99.Baselines reach at most 0.38 character consistency and 0.40 environment consistency. LVbench-C’s multi-scene scenarios better suit baselines that enforce regular shot transitions, explaining their higher narrative scores there than on VBench-Long.Source: Table 3 · p. 12, §5.2 · p. 12

5.3. Human Evaluations

• A2RD averages 4.68/5.00 in human evaluation, above VideoMemory’s 3.93, and achieves the highest scores across all six criteria.A2RD scores 4.89 in character consistency, 4.75 in narrative coherence, 4.34 in transition smoothness versus the best baseline’s 3.34, and 4.91 in reference consistency. Seven evaluators rate 40 VBench-Long samples, with each sample scored by at least two evaluators.Source: Table 4 · p. 12, §5.3 · pp. 12–13

5.2. Automatic Evaluations

• A2RD’s average consistency rises from 0.73 to 0.78 and narrative coherence from 0.74 to 0.9 with test-time scaling, while Naive-AR’s coherence decreases.Best-of-N scaling raises Naive-AR’s average consistency from 0.61 to 0.67 and VideoMemory’s from 0.65 to 0.71; A2RD’s multi-dimensional judges select candidates across consistency and coherence.Source: §5.2 · p. 12, Figure 6 · p. 13

6.1. Ablation Studies

• Ablations show that MVMem, test-time scaling components, and adaptive segment generation all contribute to A2RD’s consistency and narrative coherence.Removing HITS lowers narrative consistency 0.90 →0.74 and character consistency 0.74 →0.68; removing global references lowers environment consistency 0.84 →0.79. Always extrapolating retains 0.83 narrative coherence but reduces consistency, while always interpolating maximizes consistency at the cost of narrative coherence; removing textual states notably harms narrative and consistency, while removing videos has less impact.Source: §6.1 · pp. 13–14

6.2. Generalization to Other Video Diffusion Models

• A2RD improves over Naive-AR across LTX-Video and Wan 2.2, demonstrating generalization across video diffusion backbones.On LTX-Video, narrative coherence rises 0.59 →0.79 and character consistency 0.50 → 0.70. On Wan 2.2, narrative coherence rises 0.67 →0.80, character consistency 0.51 →0.69, and environment consistency 0.59 →0.78. Wan 2.2 does not support interpolation, so its experiments use always-extrapolate mode.Source: Table 5 · p. 14, §6.2 · p. 14

6.3. Consistency Analysis over Extended Horizons

• 96.7% character, 91.8% environment, and 95.0% object consistency were retained by A2RD, which outperformed baselines across all three axes on five-minute LVBench-C videos.The MLLM-Judge identifies visual differences, classifies inconsistencies as expected or unexplained, and flags obvious unexplained errors; manual verification found 80% agreement. Subtle inconsistencies may be missed by design.Source: §6.3 · pp. 14–15, Figure 7 · p. 15

6.4. Scaling to Longer-Horizon Video Generation

• On ten ten-minute LVBench-C scenarios, A2RD averages 90.5% character, 84.0% environment, and 91.5% object consistency.The evaluation uses the MLLM-Judge method developed for extended-horizon consistency analysis.Source: §6.4 · p. 15

7. Conclusions

• A2RD decouples creative synthesis from consistency enforcement using multimodal video memory, adaptive generation, and hierarchical test-time self-improvement.Memory tracks cross-modal context, adaptive generation supports narrative progression while enforcing consistency, and self-refinement of frames and segments aims to prevent error propagation.Source: §7 · pp. 15–16
• A2RD achieves reported state-of-the-art narrative coherence and visual consistency against existing baselines across 1–10-minute videos.The conclusion frames these as advances in addressing temporal consistency and narrative coherence.Source: §7 · pp. 15–16
• LVbench-C stress-tests long-horizon consistency through cyclical appearance with optional state evolutions.The benchmark is introduced alongside experiments spanning videos from one to ten minutes.Source: §7 · p. 16

Limitations

• A2RD incurs more computational overhead than the tested baselines, though its active reasoning cost is estimated at less than $0.50 per segment using Gemini 3 Flash.The estimate assumes fewer than 10K tokens per call on average and excludes generating five additional videos and five additional frames per segment; compute-matched autoregressive baseline variants are also provided.Source: Limitations · p. 16
• The study reports no human-rater agreement scores because some videos had exactly two reviewers and long-form quality judgments are subjective.Scores were averaged across raters, who were described as highly qualified.Source: Limitations · p. 16
• A2RD depends on component models with strong instruction-following and visual reasoning, while its verification rubrics may not generalize across styles, cultures, or domains.The authors encourage adapting the rubrics to domain-specific settings and identify transition smoothness and physical-environment consistency as persistent challenges for segment-based methods.Source: Limitations · p. 16

A. Terminologies and Summary of Major Implemented Functions

• Appendix A includes a glossary of key terms and a summary of the main MLLM and synthesis functions implemented in A2RD.Source: Table 6 · p. 22, Table 7 · p. 22

B.1. General Video Quality Assessments

• A2RD achieves the best Subject Consistency score and second-best Background Consistency, Aesthetic Quality, and Imaging Quality scores across six VBench metrics.The assessment covers general video quality (Huang et al., 2024).Source: Table 8 · p. 23, §B.1 · p. 23
• A2RD-Par obtains the highest scores in Background Consistency and Temporal Flickering.Source: §B.1 · p. 23
• The results indicate that A2RD does not sacrifice visual fidelity for temporal consistency and narrative coherence.Source: §B.1 · p. 23

B.2. Methodology Design Analysis: Why Does A2RD Work?

• The analysis focuses on Adaptive Segment Generation, Frame-Video Co-Optimization, MVMem Textual States, selected MLLM-Judge criteria, and MAPO.It omits MVMem keyframes and other MLLM-Judge criteria because the paper describes them as widely studied in prior work.Source: §B.2 · p. 23

B.2.2. Why Co-Optimization in Extrapolation Mode Is Crucial (Section 3.3)

• End-frame co-optimization corrects extrapolation inconsistencies that are difficult to fix through prompt optimization alone, including missing accessories and character-identity errors.In the example, co-optimization restores a missing blue scarf and corrects a character’s face to match the reference identity.Source: Figure 9 · p. 24, §B.2.2 · p. 24

B.2.1. Adaptive Segment Generation Mechanism

• Adaptive Segment Generation selects interpolation for known environments and extrapolation when an environment is unknown, balancing consistency with natural scene progression.The paper reports that a state-of-the-art MLLM agrees with human experts on mode selection more than 85% of the time; most errors are benign misclassifications of extrapolation as interpolation.Source: §B.2.1 · p. 24

B.2.3. How A2RD Maintains Long-Distance Motion Continuity (MLLMimg

• MLLMimg retr retrieves a matching earlier shot as a begin frame, maintaining visual and motion continuity when a scene returns after a long temporal gap.A2RD retrieves the red sports car’s earlier shot from Scene 6 for Scene 12, despite retaining only segment begin and end frames directly.Source: §B.2.3 · p. 25

B.2.4. The Role of MVMem’s Textual States

• MVMem’s Textual States retrieve and enforce details omitted from otherwise comprehensive prompts, preserving entity and environment details in synthesized frames.The example includes Elena’s Racing Bib 402 and FAMILY REUNION tables and rainbow, supported alongside prior-scene reference frames and global anchors.Source: §B.2.4 · pp. 25–26

B.2.5. Why Do We Need Those MLLM-Judge Criteria?

• Physical Plausibility, Spatial Logicalness, and Narrative Progression criteria address distinct risks: awkward frames, discontinuities across segments, and consistent but meaningless storyboards.Physically implausible frames can prompt video hallucinations; a 90-degree car rotation breaks spatial continuity; without Narrative Progression, a storyboard may remain consistent but lack a meaningful story.Source: §B.2.5 · pp. 26–27

B.2.6. Memory-Augmented Prompt Optimization (MAPO)

• 6.4 to 8.3: MAPO’s example improves the average score by retrieving refinement cases, synthesizing actionable lessons, and applying them to revise a prompt.The refined prompt replaces abstract camera references with concrete anchors, specifies hand positions and tactile actions, and orders environmental elements spatially; entity and environment reference consistency and spatial logicalness each reach 10/10.Source: Figure 15 · p. 27, §B.2.6 · p. 27

Environment Inconsistencies

• Environment failures include scenes that resemble but do not faithfully reproduce the reference layout and a wooden table that abruptly appears during a coffee-serving scene.Both cases were flagged by the MLLM-Judge.Source: §B.3 · p. 28, Figure 16 · p. 28
• The two-iteration test-time self-refinement loop does not always resolve environment errors, motivating more precise physics-aware control.The authors identify advancing foundation models with physics-aware control as a direction for future work.Source: §B.3 · p. 28

Entity Inconsistencies

• An entity failure reproduces a checkered shirt in a slightly different color, and the MLLM-Judge may miss such minor inconsistencies.The authors suggest that addressing this limitation may require more accurate judges and stronger reasoning foundation models.Source: §B.3 · p. 28, Figure 16 · p. 28

B.4. Test-Time Self-Improvement Analysis

• After two HITS iterations, frame-level overall average rises 8.741 → 9.407 and video-level overall average rises 7.928 → 8.493; environment consistency improves 7.430 → 8.795.Spatial logicalness improves 7.083 → 8.708 and physical plausibility 8.583 → 8.792 at the frame level; video-level narrative progression rises 8.700 → 9.450 and physical plausibility 8.175 → 9.000.Source: Table 9 · p. 28, Table 10 · p. 29, §B.4 · p. 29

B.5. Estimated Latency Analysis

• Under ideal hardware assumptions, Naive-AR latency is approximately N L_M + N L_V, while Naive-Par latency is N L_M + (N + 1) L_I + L_V.The analysis assumes batch inference latency equals one call.Source: §B.5 · p. 29

B.5.1. Baseline Latency

• MovieAgent (Wu et al., 2025b) estimates ideal latency assuming parallel frame and video synthesis after three LLM-driven planning stages; VideoMemory (Zhou et al., 2026) processes shots sequentially.VideoMemory’s per-shot stages include entity reuse-or-create analysis, keyframe prompt creation and image synthesis, followed by video synthesis.Source: §B.5.1 · pp. 29–30

B.5.2. A2RD Latency

• A2RD’s estimated total latency comprises initialization, precomputation, boundary-frame processing, and video-segment synthesis; A2RD-Par parallelizes video synthesis so its latency does not scale with N.In A2RD-Par, N_interp = N and k_v = 1, with video synthesis incurring L_M + L_V.Source: §B.5.2 · p. 31
• With k_v = k_f = 3, A2RD adds roughly 12 MLLM, 2 TI2I, and 2 TI2V calls per segment versus VideoMemory, for under 7.2 minutes of computational overhead per segment.The authors characterize this latency as a trade-off for improved generative yield and autonomous scalability, noting that comparable manual refinement typically takes tens of minutes per segment.Source: §B.5.2 · p. 31

B.6. Use of AI Assistants

• The authors report using ChatGPT, Gemini, and Claude for coding and writing support, while attributing substantive content, analysis, and conclusions to the authors.Source: §B.6 · p. 31

C.1. Automatic Metrics’ Implementations

• Semantic alignment uses ViCLIP (Wang et al., 2024b), while inter-shot character and environment consistency averages DINOv3 embedding similarities across grouped scenes.YOLOv8 segments character and background regions in keyframes before DINOv3 (Siméoni et al., 2025) embeddings are compared.Source: §C.1 · pp. 31–32
• Inter-shot motion uses VBench motion_smoothness on the last 10 frames of a preceding segment and first 10 frames of its continuous successor; intra-shot consistency uses VBench metrics.The previous continuous scene is identified before comparing boundary frames.Source: §C.1 · p. 32

C.1.1. Best-of-N Baseline’s Prompts

• The image Best-of-N prompt selects one candidate holistically for scene faithfulness, visual quality, and consistency with reference entities.The prompt supplies the scene description, reference entities, and candidate images, and requests a JSON selection with a brief justification.Source: §C.1.1 · p. 32
• The video Best-of-N prompt selects one clip for prompt faithfulness, visual quality, and natural motion continuation from the previous clip.It provides the scene description, previous clip, and candidate videos, and requests a JSON selection with a brief justification.Source: §C.1.1 · p. 32

C.1.2. Prompt for Narrative Coherence Evaluation

• Narrative coherence is scored from 0 to 1 against the story context, penalizing illogical story, character, object, or environment progression and unjustified repetition.The prompt caps scores at 0.6 for repetitive actions or settings absent from the context, and at 0.2 when the video or its elements are incomprehensible.Source: §C.1.2 · pp. 32–33

C.1.3. Prompt for Grouping Scenes for Inter-Shot Metrics

• The scene-grouping prompt requests JSON groups of scene indices for shared backgrounds, characters needing consistent appearance, and objects needing consistent appearance.It asks the evaluator to analyze the video scenario and scenes for shared locations and visual consistency needs.Source: §C.1.3 · p. 33

C.2. Human Metrics’ Scoring Guidelines

• Participants watch a randomly sampled, anonymized subset of videos from all methods, each comprising multiple 8-second segments.The videos are presented in randomized order.Source: §C.2 · p. 33
• For each video, participants review reference images and scene descriptions, then rate it on six criteria using a 5-point Likert scale.The study is expected to take approximately 30 minutes.Source: §C.2 · pp. 33–34

Evaluation Criteria

• LVbench-C evaluates character, object, and environment consistency, transition smoothness, narrative coherence, and reference-image consistency (N/A when no reference images are provided).Source: Evaluation Criteria · p. 34

D. LVbench-C: Examples

• In the 3-minute, 24-scene example, a Dutch oven appears in scenes 1–5, disappears in 6–15, and reappears in 16–20.Source: Figure 17 · p. 35
• In the 5-minute, 40-scene example, characters appear in scenes 1–4, disappear for scenes 5–14, then reappear with evolved states in scenes 15–40.Source: Figure 18 · p. 36
• The lighthouse narrative moves from a morning routine in the lantern room to supply delivery and a storm warning at the base, then storm preparation and the assistant’s staircase climb.Source: D. LVbench-C: Examples · p. 37
• Later scenes show rain lashing the lantern room, a mud-wet lighthouse base, and a night storm as the keeper struggles to balance while the tower sways.Source: D. LVbench-C: Examples · p. 37
• In the 10-minute, 80-scene example, the lantern room, lighthouse base, and spiral staircase recur across non-consecutive scenes as their states evolve through a storm.Source: Figure 19 · p. 38

E.1.1. MVMem Initialization: Reasoning, Equation (1)

• The initialization prompt identifies key moving characters and objects, describes their original appearance, and asks what other reference each depends on for visual consistency.The entity list is refined against the storyline and returned as a JSON array.Source: §E.1.1 · p. 39

E.1.2. MVMem Initialization: Environment Synthesis

• Environment references are synthesized as one cohesive image containing the described elements, without humans unless explicitly specified.Source: §E.1.2 · p. 40

E.1.3. MVMem Initialization: Entity Synthesis

• Entity references are generated as centered, high-fidelity images on a clean white background, with faces excluded for objects unless explicitly required.Source: §E.1.3 · p. 40

E.2.1. Prompt for Adaptive Segment Generation Mode (All Scenes), Equation

• Scenes are grouped when location, temporal continuity, and action flow align; framing changes split static scenes, while location changes and time jumps are hard boundaries.Framing or zoom changes remain continuous in moving environments.Source: §E.2.1 · p. 40

E.2.2. Prompt to Obtain 𝒱rel (All Scenes), Equation (3)

• For each scene, the prompt selects at most one contiguous previous scene, requiring the same location, continuous time, and directly flowing action; if several qualify, it chooses the most recent.Source: §E.2.2 · p. 41

E.2.3. Prompt to Obtain 𝒮rel (All Scenes), Equation (3)

• The visual-consistency prompt reasons over earlier scenes for recurring objects, characters, and spatial layouts, then selects up to ten relevant scenes per category.Selections are restricted to lower-index scenes and prioritize recent, visually relevant material.Source: §E.2.3 · pp. 41–42

E.2.4. Prompt to Retrieve Relevant Global References (All Scenes), Equation (4)

• Global-reference retrieval identifies which named character, object, or environment anchors are present or relevant in each scene.Source: §E.2.4 · p. 42

E.2.5. Prompt to Retrieve the Best End-of-Shot Frame for Scene Continuation, MLLMimg; E.2.6. Prompt to Generate Frame Prompts, MLLMimg

• Frame continuation uses subject, motion, and environment to choose the best end-of-shot starting frame; frame prompts are first planned across scenes, then refined with memory.The prompts preserve specified scene content and visual style, maintain character and environment consistency, and preserve camera angle and lighting for continuations; first appearances establish detailed references for later scenes.Source: §E.2.5 · p. 42, §E.2.6 · pp. 42–43

E.2.6. Prompt to Generate Frame Prompts, MLLMimg

• Frame-prompt refinement evaluates alternative camera angles against spatial relations, then preserves continuity by changing only elements needed for physical plausibility and consistency.Prompts specify visible and hidden entities, relative orientations, camera position, entity positions, and depth layers; spatial directions are anchored to entities or the environment rather than the camera.Source: §E.2.6 · p. 44

E.2.7. Prompt to Synthesize Frames, TI2I in Equation (5)

• Frame synthesis uses scene prompts and ordered reference images to generate an image with strict visual consistency while advancing the narrative.The prompt maps reference images by index and asks the model to follow the scenes for narrative progression.Source: §E.2.7 · p. 45

E.2.8. Prompt to Generate Video Prompts, MLLMpgen-vid in Equation (6)

• Video prompts condition each segment on scene descriptions, video-state memory, and the previous prompt, enforcing entity, environment, motion, and camera-direction continuity.They start from the beginning frame and either end at a specified ending frame or conclude naturally to fulfill the scene description; multiple actions or camera transitions can be divided into temporal segments.Source: §E.2.8 · p. 45

E.2.9. Prompt to Synthesize Video Segments, TI2V in Equation (6)

• The generated video prompt is used to synthesize the video segment.Source: §E.2.9 · p. 46

E.3.1. Prompt to Extract Frame States, MLLMimg

• Frame-state extraction records entity identity and observable state, including positions and orientations relative to other entities or environmental features rather than the frame.It identifies characters, objects, and environments and represents spatial relations in a scene graph.Source: §E.3.1 · pp. 46–47

E.3.2. Prompt to Extract Video States, MLLMvid

• Video-state extraction supplements known begin-frame states with new entities and revealed identity details, and describes entity motions and camera dynamics across the segment.Source: §E.3.2 · p. 47
• The MLLMvid prompt extracts video-state motions and describes camera behavior at the beginning, middle, and end.Source: §E.3.2 · p. 48

E.3.3. Prompt to Judge Frame - Consistency over Images (Entity, Environment, Narrative), Equation (7)

• Frame judging scores entity appearance, environment layout, and narrative progression against mapped reference images and the frame prompt.Source: §E.3.3 · p. 48

E.3.4. Prompt to Judge Frame - Consistency over Images (Spatial Logicalness), Equation

• Spatial-logicalness judging treats anchor images as canonical layouts, requiring the current frame to match unless its prompt describes changes.Source: §E.3.4 · p. 49

E.3.5. Prompt to Judge Frame - Textual States, Equation (7)

• Text-state judging compares objects, characters, and environments with relevant prior scenes and the storyline, flagging unjustified static-attribute discrepancies but allowing natural progression.Source: §E.3.5 · p. 49

E.3.6. Prompt to Judge Frame - Basic Quality, Equation (7)

• Basic frame-quality judging scores instruction following and physical plausibility, requiring prompt details and a physically plausible unified moment.Source: §E.3.6 · p. 50

E.3.7. Prompt to Judge Video - Inter Consistency, Equation

• Inter-scene judging checks contiguous-scene continuity in entity appearance and motion, environment layout and lighting, and camera movement.Source: §E.3.7 · p. 51

E.3.8. Prompt to Judge Video - Intra Consistency and Basic Video Quality, Equation

• Intra-video judging scores instruction following, physical plausibility, narrative progression, fit to the provided end frame, and character consistency.Source: §E.3.8 · p. 52
• The video judge compares object and environment identities with their recorded changes without watching the video, then returns seven 1–10 criterion scores in JSON.The criteria are instruction following, physical plausibility, logical progression, frame fit, character consistency, object consistency, and environment consistency.Source: §E.3.8 · p. 52

E.4.1. Prompt to Reason Over Feedback (Frame Prompt)

• For failed frame prompts, the reasoning prompt compares the scene description, original prompt, and validation feedback, identifies discrepancies and problematic wording, and asks what to clarify or simplify.It also instructs the model to analyze prompt content rather than generation quality and to choose edit for minor issues or regenerate for major failures.Source: §E.4.1 · pp. 53–54
• For frame-prompt spatial errors, the instructions add explicit entity positions relative to the environment or other entities, use reference-frame camera angles when the specified angle is impossible, and avoid hard-to-control facing directions.Examples include keeping an entity at a stated location or on a specified side of a reference.Source: §E.4.1 · pp. 53–54

E.4.2. Prompt to Reason Over Feedback (Video Prompt)

• For failed video prompts, the discrepancy analysis checks character and object appearance, environment layout and lighting, and whether motions, directions, and transitions match the scene description.The prompt’s refinement guidance adds consistency constraints for identities and environment, clarifies motion flow and sequencing, and removes overly specific or conflicting constraints.Source: §E.4.2 · p. 54
• Video-prompt revisions clarify stable character, object, and environment attributes and action sequencing, while simplifying excessive detail, impossible motion constraints, redundancy, and conflicting descriptions.The guidance prioritizes character, object, and environment consistency before motion issues.Source: §E.4.2 · p. 55

E.4.3. Prompt to Extract Contrastive Patterns (Frame Prompt)

• Frame-prompt pattern extraction compares successful and failed refinements and requests 8–12 concrete, briefly exemplified patterns focused on structure, specificity, terminology, spatial logic, and recurring failures.The patterns are intended to capture how prompt changes led to success or failure.Source: §E.4.3 · pp. 55–56

E.4.4. Prompt to Extract Contrastive Patterns (Video Prompt)

• Video-prompt pattern extraction likewise analyzes successful and failed refinements to identify key patterns.The prompt uses positive and negative refinement samples as its inputs.Source: §E.4.4 · p. 56
• The video-prompt pattern extractor requests actionable guidelines covering structure, motion, continuity, specificity, conflicts, vocabulary, and failure patterns.Each pattern should be returned as a separate concise item.Source: §E.4.4 · p. 56

E.4.5. Prompt to Refine Frame Prompt

• Frame-prompt refinement preserves the original unless analysis finds issues, then changes only problematic parts and uses environment-relative descriptions for spatial fixes.The instructions caution against camera-relative facing directions as difficult to control.Source: §E.4.5 · pp. 56–57

E.4.6. Prompt to Refine Video Prompt

• Video-prompt refinement preserves the initial state of every entity and defaults to a static camera, splitting necessary camera changes into time-bounded segments.It otherwise retains the original prompt except for targeted changes identified by analysis.Source: §E.4.6 · p. 57

E.5.1. Prompt to Verify Character Consistency Drift, Equation

• Character-consistency checks compare reference and current frames for identity and appearance, flagging only clear changes such as a different person, face, body type, or species.Minor outfit, accessory, lighting, or pose variations should not count as violations.Source: §E.5.1 · p. 58

E.5.2. Prompt to Verify Object Consistency Drift, Equation

• Object-consistency checks compare key-object identity, shape, color, and size, flagging clear appearance changes but not minor position, wear, lighting, or occlusion differences.Source: §E.5.2 · p. 58

E.5.3. Prompt to Verify Environment Consistency Drift, Equation

• Environment checks require the same physical space, flagging clear differences in architecture, layout, or location but not camera-angle changes, story events, or minor lighting shifts.Source: §E.5.3 · p. 59

F.2. Baseline Examples

• A2RD storyboard examples cover LVBench-C videos of 3, 5, and 10 minutes, with 24, 40, and 79 scenes, respectively.Source: Figure 20 · p. 60, Figure 21 · p. 60, Figure 22 · p. 61
• Baseline storyboards for the 3-minute Scuba Diver scenario are described as accumulating character, object, and environment inconsistencies that degrade narrative coherence and visual continuity.The examples include Naive-AR, Naive-Par, MovieAgent, ViMax, and VideoMemory.Source: Figure 23 · p. 61, Figure 24 · p. 62, Figure 25 · p. 62