Source-linked AI summary
SEAM: Shot Entity-Attribute Memory for Consistent Short-Drama Generation at Scale
Jiaqi Liu, Maolin Ran, Xiaoyang Lu, Jian Wang, Weiwen Liu, Jianghao Lin, Yong Yu, Weinan Zhang
TL;DR
Short-drama generation lacks persistent cross-shot state, making visual continuity a bottleneck as production scales. SEAM addresses this with a training-free, model-agnostic memory graph that retrieves and selectively injects prior state through prompt rewriting. It raises cross-episode continuity recall from 0.700 to 0.946, generalizes across six text models, and transfers directionally but not significantly to generated images.
Problem
Isolated shot prompts lack persistent state, allowing continuity drift across shots and requiring manual repair as contradictions accumulate.
Method
SEAM extracts multi-dimensional shot states into a memory graph, retrieves causally prior context, filters candidates selectively, and injects retained constraints through natural-language prompt rewriting without changing model weights or downstream backbones.
Results
SEAM raises cross-episode continuity recall from 0.700 to 0.946, remains positive across six text models, and yields directional but statistically nonsignificant gains at the generated-image layer.
Takeaways & Limitations
SEAM provides a prompt-text-layer continuity stage that is model-agnostic and deployable in a live production pipeline.
Takeaways & Limitations
Image-layer transfer is directional but not significant at n=30, and continuity recall does not repair every judged defect.
Abstract
from arXiv · showhide
Short-drama generation has grown into a large, industrialized pipeline, and as it scales from isolated shots to the episode level, visual continuity has become a critical bottleneck. Current agent frameworks generate each shot in isolation, so context drifts across shots and props, character posture, and blocking turn inconsistent. Once assembled, these small discrepancies amplify into severe visual breaks. We present SEAM (Shot Entity-Attribute Memory), a training-free, model-agnostic memory graph that repairs continuity entirely at the prompt-text layer by extracting a multi-dimensional state for every shot, retrieving only causally prior context over the resulting graph, filtering it selectively, and injecting the surviving constraints by natural-language prompt rewriting. We further release SEAM-Bench, a double-blind continuity storyboarding benchmark, on which SEAM raises cross-episode continuity recall from 0.700 to 0.946, generalizes across six mainstream text models, and yields consistent, though not yet significant, gains at the generated-image layer. Deployed as a mandatory stage in CreativeFitting's SEAM-Agent production pipeline over 201 shots, SEAM reaches a 96.5% director-acceptance rate with zero unsafe injections; a conservative counterfactual attributes at least 21.9 percentage points of that rate to its cross-episode memory.
1 Introduction
Short-drama production scales through standardized, AI-assisted workflows, but isolated shot generation loses persistent visual state and causes continuity breaks. SEAM addresses this gap at the storyboard prompt layer with a graph-based, selective memory stage that remains compatible with different generators.
- Short drama is a rapidly growing, industrialized form of vertical, multi-episode video fiction whose workflow spans writing, storyboarding, keyframe generation, and video synthesis.
- Visual continuity becomes a bottleneck because independently generated shot prompts carry no persistent state, allowing removed props such as a glove to reappear in later shots.Generators follow prompts faithfully, so contradictions accumulate across large productions and require manual editorial repair.
- Existing approaches leave the storyboard prompt layer unaddressed, although it is upstream of every generator and remains readable and editable by directors.
- SEAM represents each shot as a multi-dimensional state node connected by temporal, character, and scene relations in an explicit memory graph.
- SEAM retrieves only relevant prior state, selectively filters candidates, and naturally rewrites prompts while preserving shot grammar and director intent.
- The framework is training-free and model-agnostic, operating entirely at the prompt-text layer without modifying model weights or downstream generators.
2 Related Work
Prior work pursues visual consistency through generator-internal mechanisms, external structures, image-domain repairs, or memory systems, while evaluation remains fragmented across limited protocols and evaluator families.
- Short-drama research combines dedicated datasets, generation pipelines, role-decomposed agents, and evaluation protocols for authoring and rendering.
- Existing systems automate authoring and rendering end to end but often treat consistency as a generator or scene-planning byproduct.
- Evaluation commonly relies on human Likert ratings or a single LLM judge, leaving conclusions dependent on one evaluator family despite emerging image- and storyboard-level benchmarks.
- Visual-consistency methods include coherent multi-shot generation, memory-flow propagation, identity-preserving diffusion, storyboard or 3D anchoring, and image-domain panel repair.
- Agent memory research uses episodic memory, graph-based retrieval, associative reuse, and persistent multimodal memory to support long-horizon information access.
3 Preliminaries
The paper models a short drama as ordered episodes of structured shots and defines stateless prompt generation against which cross-shot continuity repair is evaluated. Continuity tracks persistent visual attributes of active characters, props, and scenes, with recall measuring successful repairs among judged defects.
- A short drama is an episode-ordered set of episodes, each containing a temporally ordered sequence of shots.
- Each shot includes a location, shot grammar, visual description, dialogue, and the active characters and props.
- Characters, props, and scenes form drama-wide entity universes, while global shot order supplies cross-episode context when needed.
- The generation pipeline produces one prompt per shot for a downstream image or video generator.
- The stateless baseline makes each prompt depend only on the current shot and carries no cross-shot state.
- Continuity requires persistent entity attributes, including character appearance, prop possession, and scene layout, to remain consistent unless narrative change motivates otherwise.
- Continuity recall is defined as the number of defects successfully repaired divided by the number of defects judged in need of repair.
4 Methodology
SEAM is introduced as a Shot Entity-Attribute Memory graph for short-drama storyboarding, establishing the paper’s central memory representation.
- SEAM stands for Shot Entity-Attribute Memory.
- The framework is presented as a graph for short-drama storyboarding.
- SEAM is the paper’s named memory graph for the methodology that follows.
4.1 Overview
SEAM represents continuity as a cross-shot memory graph and delivers repairs inside a live production system. It retrieves and filters prior context, then rewrites shot descriptions while preserving shot grammar and director intent.
- SEAM extracts structured multi-dimensional visual states from episode scripts and links them into a directed memory graph for cross-shot continuity repair.The graph-based procedure turns continuity into an executable retrieval-and-injection process at the prompt-text layer.
- SEAM retrieves causally prior context, selectively filters redundant or conflicting fragments, and naturally rewrites each shot description with the surviving continuity constraints.The rewrite preserves the shot grammar and the director’s intent.
- The online mode uses LLMs, while a deterministic offline mode substitutes rules so the continuity stage can run without a reachable model.Both modes are embedded in the production system as part of the same prompt-text-layer procedure.
4.2 SEAM: A Cross-Shot Continuity Memory Graph
SEAM constructs a configurable shot-state graph and repairs continuity by retrieving only prior context, filtering candidates, and injecting the retained information into prompts. Its retrieval is causal and selective, while injection preserves the original shot structure and creative intent.
- 4.2.1 State Extraction and Graph Construction: SEAM parses each shot into a configurable multi-dimensional state and uses eight dimensions in this paper, including scene, characters, blocking, props, actions, and camera style.The extractor can operate online with a fixed LLM template or offline with deterministic rule-based parsing.
- 4.2.1 State Extraction and Graph Construction: The directed graph contains shot-state and recurring resource-entity nodes connected by temporal, character, and scene relations with continuity weights.These weighted edges support multi-hop retrieval across neighboring, character-sharing, and same-location shots.
- 4.2.2 Graph Retrieval and Selective Filtering: Causal retrieval restricts each shot’s context to the prior subgraph, preventing future state from leaking into the shot being repaired.Retrieval uses resource overlap when explicit references exist, otherwise lightweight lexical matching seeds backward graph expansion.
- 4.2.2 Graph Retrieval and Selective Filtering: SEAM filters retrieved candidates independently, discarding standalone, redundant, or scene-disconnected context and retaining only material that supplies genuine cross-shot continuity.The online mode uses an LLM rubric with an auditable justification; the offline mode uses a conservative rule.
- 4.2.3 Natural-Rewrite Injection: Natural-rewrite injection weaves retained continuity into the visual description while preserving shot grammar, visual semantics, resource references, and structural markers.A guard rejects unsafe rewrites and reverts to the previous description; mechanical appending is available as a fallback.
4.3 SEAM-Agent: SEAM Within a Multi-Agent Pipeline
SEAM-Agent places SEAM as a memory-optimizer stage between shot authoring and downstream formatting or rendering. The optimizer builds persistent episode memory, repairs the generated shotlist, and can use context from both current and earlier episodes.
- 4.3.2 SEAM as the Memory-Optimizer Stage: SEAM operates after authoring agents and before downstream loading or rendering, repairing continuity over the shotlist already generated upstream.This placement makes SEAM a standalone memory-optimizer stage rather than an authoring or generation model.
- 4.3.1 SEAM-Agent Pipeline: SEAM-Agent is a serial pipeline comprising Director, Cinematographer, SEAM Memory Optimizer, and Formatter stages.Agents communicate through shotlist artifacts that are passed between stages.
- 4.3.2 SEAM as the Memory-Optimizer Stage: The optimizer builds and persists a graph from the episode script, retrieves and filters prior context for each shot, and rewrites descriptions in place.Its output replaces the stateless shot description baseline in the pipeline.
- 4.3.2 SEAM as the Memory-Optimizer Stage: Accumulated state allows retrieved context to span earlier shots in the current episode and prior episodes, supporting cross-episode continuity repair.The prompt-text-only design is decoupled from both the authoring model and the downstream backbone.
5 Experiments
SEAM-Bench evaluates continuity repair across three dramas, six text models, multiple metric families, and generated keyframes. SEAM improves continuity recall and targeted contradiction measures consistently, while image-layer gains remain directional and deployment results are strong but bounded by documented limitations.
- Experimental setup: SEAM-Bench covers three produced dramas totaling 68 episodes and uses scripts, reference images, and held-out expert storyboards for evaluation.The benchmark releases scripts and reference images, while professional director storyboards remain restricted and are used for evaluation and unblinding.
- Experimental setup: Six heterogeneous text models produce paired uninjected and memory-injected shotlists, isolating memory injection as the varying factor.The paired CSVs are column- and shot-aligned, enabling within-model contrasts.
- Q1: continuity repair: 0.946 continuity recall versus 0.700 for episode-local retrieval shows stronger cross-episode repair on his-toyboy.SEAM repaired 35 of 37 judged defects, while the control repaired 7 of 10; the denominators differ because cross-episode retrieval exposes more candidate defects.
- Q2: model-agnosticism: Recall remains high across all 18 drama–backbone cells, while targeted contradiction probability falls from 0.29–0.35 to 0.10–0.17.Blind prompt-layer gains are positive in 17 of 18 model–dimension entries, with glm-5.1 prop continuity the sole exception.
- Q3: image-layer transfer: Generated-image gains are directional but nonsignificant at n=30, with improvements on character appearance (+3.7) and prop continuity (+3.3).Scene layout and frame quality remain flat, and stochastic text-to-image generation attenuates transfer.
- Online deployment: 96.5% director acceptance with zero unsafe injections was achieved in deployment, while cross-episode memory contributed a conservative lower-bound Δ of 21.9 percentage points.SEAM injected into 27.4% of shots; the residual misses involved incomplete rather than contradictory state tracking.
- Limitations: The evidence is limited by nonsignificant image-layer transfer, a Kendall’s τ decline from 0.446 to 0.412, near-0.20 camera-angle accuracy, and reliance on one judge family.Fallback modes are also unevaluated.
6 Conclusion
SEAM is presented as a training-free, model-agnostic memory graph for repairing cross-shot and cross-episode continuity in prompt text. It improves continuity recall across six text models, transfers directionally to images, and is deployed in production.
- SEAM repairs cross-shot and cross-episode continuity in prompt text using a training-free, model-agnostic memory graph.
- Continuity recall rises from 0.700 to 0.946 with SEAM.
- SEAM remains positive across six text models, with recall spanning 0.731–0.992.
- Image-layer transfer is directional, while future work targets video-layer evaluation.
- SEAM is released and runs as a mandatory SEAM-Agent stage in production.
A Metric Definitions
The paper defines reference-based, reference-free, alignment, and targeted continuity metrics for evaluating storyboard consistency. These metrics quantify repair success, cinematographic agreement, entity carry-over, semantic coherence, and image–prompt fidelity.
- Continuity metrics: Continuity recall measures the fraction of judged continuity defects that are successfully repaired.A defect is recorded when an entity’s visual state contradicts a later state without narrative motivation.
- Reference-based metrics: Reference-based metrics compare AI and director storyboards through field labels, distributions, and soft alignment.Fields include shot size, camera angle, and camera movement; alignment uses Hungarian matching with a similarity threshold.
- Reference-free metrics: Reference-free metrics quantify entity carry-over, adjacent-shot coherence, and contradiction across prior shots.The suite includes entity out-degree, adjacent cosine similarity, and NLI contradiction measures.
- Targeted repair metrics: The targeted contradiction probe evaluates flagged repairs by comparing graph-expected entity states with shot descriptions before and after injection.Figure 5 covers 2,946 shot-level before/after pairs.
- Image-layer metrics: Image-layer metrics use DINOv2 similarities, DreamSim distance, and CLIP-T to assess identity, scene reference, self-consistency, and image–prompt fidelity.Higher DINOv2 similarity is better, lower DreamSim distance indicates more stable appearance, and CLIP-T measures visual-description alignment.
B Online Deployment Architecture
SEAM operates as the mandatory memory-optimization stage inside CreativeFitting’s multi-agent storyboarding pipeline. The system preserves director and editor control through repository write-back, feedback loops, and override paths.
- Pipeline flow: CreativeFitting’s pipeline sends scripts through Director, Cinematographer, SEAM Memory Optimizer, and Formatter/Loader agents before rendering.
- Pipeline flow: The Director Agent decomposes scripts into shots and fixes pacing and narrative intent.
- Pipeline flow: The Cinematographer Agent assigns shot parameters and resolves entities against the shared Asset Repository.Unmatched items remain text-only rather than being forced onto an incorrect asset.
- Memory optimization: SEAM runs as a mandatory stage that extracts shot state, retrieves prior graph context, and filters continuity information.
- Human oversight: Directors and editors can provide manual re-edit feedback and override any pipeline stage.The repaired storyboard is written back to the VFS Storyboard Repository, while telemetry tracks production scale, repair rate, and reliability.