pith. sign in

arxiv: 2607.01312 · v1 · pith:H7PAZQKVnew · submitted 2026-07-01 · 💻 cs.CV

KathaTrace: Diagnosing Semantic Trajectory Collapse in Generated Visual Narratives

Jamuna S. Murthy , Amin Karimi Monsefi , Rajiv Ramnath This is my paper

Pith reviewed 2026-07-03 21:12 UTC · model grok-4.3

classification 💻 cs.CV
keywords semantic trajectory collapsevisual narrativesstoryboard generationsemantic trajectory gapKathaTracerecoverability evaluationimage sequence generationtransition meaning
0
0 comments X

The pith

Visual narrative generators lose 23.5 points of recoverable transition meaning between scenes.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces KathaTrace as a protocol that measures semantic trajectory collapse: the loss of story transition meaning when text narratives are turned into image sequences. It creates recoverability questions under text-only, image-only, and combined conditions, then defines Semantic Trajectory Gap as the difference between text-only and image-only scores after filtering ambiguous cases. Experiments on current generators produce a consistent gap of 23.5 plus or minus 1.3, showing that visual coherence can mask the disappearance of how one scene logically follows another. The work also supplies a 25K-item benchmark drawn from classical story collections and demonstrates an actionability probe that uses the signals for post-generation repair.

Core claim

KathaTrace evaluates generated visual narratives by comparing human recoverability of transition meaning across text-only, image-only, and text-plus-image conditions on KathaBench-25K. Semantic Trajectory Gap is defined as text-only recoverability minus image-only recoverability, directly quantifying the semantic link lost during visualization. State-of-the-art generators exhibit an STG of 23.5 plus or minus 1.3, and Semantic Compass leverages these signals to improve storyboard selection.

What carries the argument

Semantic Trajectory Gap (STG), the arithmetic difference between text-only and image-only recoverability scores on filtered transition questions.

If this is right

  • Existing visual quality and coherence benchmarks miss a distinct failure mode of semantic loss between scenes.
  • Post-generation repair methods such as Semantic Compass can use KathaTrace signals to select or adjust storyboards.
  • A 25K-item benchmark enables systematic comparison of generators on transition fidelity rather than surface appearance.
  • The three-condition recoverability design isolates the visualization step as the source of meaning loss.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Training objectives for story generators may need explicit terms that penalize loss of transition semantics rather than only image-level coherence.
  • The protocol could be adapted to video or animation generators where temporal ordering carries additional narrative weight.
  • Classical story collections provide a stable, culturally diverse test bed that avoids modern copyright constraints while preserving complex transition structures.

Load-bearing premise

The recoverability questions under text-only, image-only, and text-plus-image conditions, after filtering ambiguous items, accurately and unbiasedly measure the loss of transition meaning.

What would settle it

Apply KathaTrace to a generator engineered to copy every original transition explicitly into its image sequence and observe whether the measured STG falls to near zero.

Figures

Figures reproduced from arXiv: 2607.01312 by Amin Karimi Monsefi, Jamuna S. Murthy, Rajiv Ramnath.

Figure 1
Figure 1. Figure 1: Semantic trajectory collapse. A generated storyboard may preserve characters, setting, and local coherence while changing the transition meaning recoverable from images alone. KathaTrace measures this loss using image-only recoverability and the Semantic Trajectory Gap. accessibility settings, low-literacy communication, and film previsualization, images are often the final medium and the source text may n… view at source ↗
Figure 2
Figure 2. Figure 2: KathaTrace framework. KathaTrace structures stories into scenes and transition-level recoverability questions, then tests them under text-only, image-only, and text+image evidence. It reports STG, dimension gaps, ambiguity rates, contrastive recoverability, and the optional Semantic Compass repair probe. but they do not define source-story transition targets and then test whether those targets remain recov… view at source ↗
Figure 3
Figure 3. Figure 3: Case study of a missed narrative failure. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: KATHABENCH-25K preference distributions. Human scores across perception criteria. A positive median indicates preference for the lower-STG, KathaTrace-guided, or repaired storyboard. For multiple criteria, we apply Bonferroni correction: αcorr = α C , (71) where C is the number of tested criteria. Rank-biserial correlation is reported as the ordinal effect size: rrb = W+ − W− W+ + W− , (72) where W+ and W−… view at source ↗
Figure 5
Figure 5. Figure 5: KATHABENCH-25K preference density. Violin view of human ratings [PITH_FULL_IMAGE:figures/full_fig_p051_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: KATHABENCH-25K dimension preferences. Mean preference by perception group [PITH_FULL_IMAGE:figures/full_fig_p051_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: KATHABENCH-25K ranked preferences. Mean score by criterion [PITH_FULL_IMAGE:figures/full_fig_p051_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: KATHABENCH-25K preference significance. Corrected tests with confidence intervals. for item q, and let vq be the VLM or VLM-ensemble answer. Human–VLM agreement is Acchuman = 1 |Q| X q∈Q I[vq = hq]. (73) 51 [PITH_FULL_IMAGE:figures/full_fig_p051_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: KATHABENCH-25K group preferences. Robustness view across participant groups [PITH_FULL_IMAGE:figures/full_fig_p052_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: KATHABENCH-25K human–VLM calibration. Image-only recoverability correlation. For continuous recoverability scores, Spearman rank correlation is ρsp = corr rank(R VLM image),rank(R human image )  . (74) For pairwise comparisons, agreement is PairAgr = 1 |P| X (a,b)∈P I [PITH_FULL_IMAGE:figures/full_fig_p052_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: KATHABENCH-25K Ocean examples. Representative stories from The Ocean of Story. and KathaTrace diagnostic/repair cases. These figures are illustrative and are not used as the sole basis for model ranking. K.1 Dataset Examples Figs. 11, 12, and 13 show representative visual-only KATHABENCH-25K dataset examples. Each row contains a fixed source story, a short intended meaning, and five target scenes. These e… view at source ↗
Figure 12
Figure 12. Figure 12: KATHABENCH-25K Panchatantra examples. Representative transition targets from Panchatantra stories. K.2 Qualitative Example Selection Qualitative examples are selected from the held-out KATHABENCH-25K split using fixed criteria rather than manual visual preference. We compute STG, dimension-level gaps, validity masks, ambiguity flags, and Semantic Compass repair deltas before selecting examples. Invalid ge… view at source ↗
Figure 13
Figure 13. Figure 13: KATHABENCH-25K Aesop examples. Representative transition targets from Aesop’s fables. K.4 KathaTrace Diagnostic and Repair Examples Figs. 14–18 show KATHABENCH-25K diagnostic examples covering action, causality, source-side intention, emotion, and consequence transitions. Each figure contains the source story, intended meaning, a ground-truth storyboard, a generated failure case, and a KathaTrace diagnosi… view at source ↗
Figure 14
Figure 14. Figure 14: KATHABENCH-25K action failure. Missing wedge-pulling transition in The Monkey and the Wedge. L.1 Symbolic Moral Sacrifice: King Shibi and the Dove [PITH_FULL_IMAGE:figures/full_fig_p061_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: KATHABENCH-25K causal failure. Missing trust-breaking link in The Boy Who Cried Wolf [PITH_FULL_IMAGE:figures/full_fig_p062_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: KATHABENCH-25K intention-field failure. Hidden deception in The Cat and the Birds. monkey. A visually coherent storyboard may show the monkey riding the crocodile while failing to communicate deception or danger. 62 [PITH_FULL_IMAGE:figures/full_fig_p062_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: KATHABENCH-25K emotional failure. Missing admiration-to-jealousy turn in The Fall of Sunda and Upasunda [PITH_FULL_IMAGE:figures/full_fig_p063_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: KATHABENCH-25K consequence failure. Missing greed-to-loss transition in The Goose That Laid the Golden Eggs. The intended meaning depends on two recoverable facts: the crocodile’s friendliness is false, and the monkey survives through quick verbal strategy. Without speech cues, facial cues, or explicit threat evidence, the image-only interpretation may become “friendship,” “travel,” or “river adventure.” … view at source ↗
Figure 19
Figure 19. Figure 19: KATHABENCH-25K same-story comparison. Generator-agnostic comparison on The Cat and the Birds. 65 [PITH_FULL_IMAGE:figures/full_fig_p065_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: KATHABENCH-25K hard case: symbolic sacrifice. King Shibi and the Dove [PITH_FULL_IMAGE:figures/full_fig_p066_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: KATHABENCH-25K hard case: hidden intention. The Monkey and the Crocodile [PITH_FULL_IMAGE:figures/full_fig_p066_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: KATHABENCH-25K hard case: long-range reciprocity. The Elephants and the Mice. The ambiguity rate is ρamb,m = 1 |Q| X q∈Q bm(q). (79) 66 [PITH_FULL_IMAGE:figures/full_fig_p066_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: KathaTrace beyond storyboards. Ordered narrative units for comics, illustrated pages, and video keyframes [PITH_FULL_IMAGE:figures/full_fig_p067_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Comic-style pilot: lunch-table reciprocity. [PITH_FULL_IMAGE:figures/full_fig_p067_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Comic-style pilot: spare-bicycle reciprocity. [PITH_FULL_IMAGE:figures/full_fig_p068_25.png] view at source ↗
read the original abstract

Visual narratives are central to storyboards, comics, children's media, and film previsualization, where viewers understand stories from images alone. Recent generators such as StoryDiffusion produce coherent sequences, but visual coherence does not guarantee that source-story transition meaning remains recoverable. Existing benchmarks assess visual quality, content faithfulness, and scene coherence, but miss a critical failure mode: storyboards where scenes appear visually coherent while the semantic link between scenes disappears. We introduce KathaTrace, a generator-agnostic protocol for diagnosing semantic trajectory collapse, defined as the loss of transition meaning needed to understand how one scene follows another. KathaTrace evaluates transitions under three evidence conditions: text-only, image-only, and text-plus-image, and filters ambiguous items. We contribute KathaBench-25K, with 5,000 narratives from classical collections including Aesop, Panchatantra, and Kathasaritasagara, 20,000 transitions, and 28,712 recoverability questions. We define Semantic Trajectory Gap, or STG, as text-only minus image-only recoverability, measuring transition meaning lost during visualization. Human validation yields Fleiss' kappa = 0.845. Experiments across state-of-the-art generators show substantial STG of 23.5 +/- 1.3. Semantic Compass, an actionability probe, uses KathaTrace signals for post-generation repair and improves storyboard selection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces KathaTrace, a generator-agnostic protocol and KathaBench-25K dataset (5,000 classical narratives, 20,000 transitions, 28,712 questions) to diagnose semantic trajectory collapse in AI-generated visual narratives. It defines Semantic Trajectory Gap (STG) as text-only minus image-only recoverability of transition meaning, reports STG = 23.5 +/- 1.3 across SOTA generators with Fleiss' kappa = 0.845 after ambiguous-item filtering, and introduces Semantic Compass for post-generation repair using the signals.

Significance. If the recoverability questions validly isolate transition semantics, the work identifies and quantifies a previously unmeasured failure mode (semantic loss despite visual coherence) with a large-scale, human-validated benchmark and an actionable repair probe. The generator-agnostic design and dataset scale from classical sources are clear strengths.

major comments (2)
  1. [Abstract / recoverability question protocol] Abstract and question-design description: the headline STG claim (23.5 +/- 1.3) is interpreted as semantic trajectory collapse only if the 28,712 recoverability questions specifically probe 'transition meaning needed to understand how one scene follows another.' High Fleiss' kappa = 0.845 confirms consistency but does not establish construct validity; no evidence is given that questions avoid confounding with general story recall, visual detail extraction, or prompt alignment, and the text-plus-image condition is mentioned without reported results to rule out bias.
  2. [STG definition and results] STG definition and filtering: STG is computed directly as text-only minus image-only recoverability on filtered KathaBench-25K transitions. Without explicit reporting of exclusion criteria, inter-condition question matching, or an ablation showing that the difference survives controls for non-transition factors, the numerical result cannot be unambiguously attributed to semantic collapse rather than other comprehension differences.
minor comments (1)
  1. [Abstract] The abstract states 'experiments across state-of-the-art generators' but does not name the specific models or report per-generator breakdowns; adding this would strengthen reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive feedback. The comments identify opportunities to strengthen the presentation of construct validity and methodological details. We address each major comment below and indicate revisions that will be incorporated.

read point-by-point responses

  1. Referee: [Abstract / recoverability question protocol] Abstract and question-design description: the headline STG claim (23.5 +/- 1.3) is interpreted as semantic trajectory collapse only if the 28,712 recoverability questions specifically probe 'transition meaning needed to understand how one scene follows another.' High Fleiss' kappa = 0.845 confirms consistency but does not establish construct validity; no evidence is given that questions avoid confounding with general story recall, visual detail extraction, or prompt alignment, and the text-plus-image condition is mentioned without reported results to rule out bias.

    Authors: The recoverability questions are generated from the explicit transition semantics in the source narratives (Section 3.2), targeting the inferential link required to understand scene succession rather than isolated scene content. We agree that additional evidence for construct validity would be valuable. In the revision we will (i) report the text-plus-image condition results (recoverability >90% across items), (ii) include representative question examples that isolate transition meaning, and (iii) add a short discussion of how question phrasing was constrained to avoid general recall or visual-detail confounds. These additions will be placed in a new subsection of the methods. revision: yes

  2. Referee: [STG definition and results] STG definition and filtering: STG is computed directly as text-only minus image-only recoverability on filtered KathaBench-25K transitions. Without explicit reporting of exclusion criteria, inter-condition question matching, or an ablation showing that the difference survives controls for non-transition factors, the numerical result cannot be unambiguously attributed to semantic collapse rather than other comprehension differences.

    Authors: Exclusion criteria (low inter-annotator agreement on ambiguous items) are stated in Section 4.2, and the same question set is administered under all three evidence conditions to ensure direct matching. We acknowledge that an explicit ablation isolating transition-specific effects would further support attribution to semantic collapse. The revised manuscript will therefore include (a) a consolidated table of exclusion statistics and (b) an ablation comparing STG on transition questions versus matched non-transition (scene-detail) questions. This will be added to the results section. revision: yes

Circularity Check

0 steps flagged

No significant circularity; STG is an explicit operational definition

full rationale

The paper defines Semantic Trajectory Gap (STG) directly as text-only recoverability minus image-only recoverability on the KathaBench-25K transitions after filtering. This is presented as a measurement protocol rather than a derived quantity obtained from fitting, self-citation chains, or an ansatz that reduces to the input by construction. No equations, uniqueness theorems, or predictions are shown that collapse back to the same fitted values or self-referential definitions. The central empirical claim (STG of 23.5 +/- 1.3) is therefore an observation under the stated definition and remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

Based on abstract only; full details on assumptions unavailable. The protocol rests on human judgment reliability and representativeness of classical stories.

axioms (1)
  • domain assumption Human judgments on recoverability questions reliably measure semantic transition meaning
    Central to defining STG as the difference in recoverability.
invented entities (2)
  • Semantic Trajectory Gap (STG) no independent evidence
    purpose: Quantifies loss of transition meaning during visualization
    Newly defined as text-only minus image-only recoverability.
  • KathaTrace protocol no independent evidence
    purpose: Diagnoses semantic trajectory collapse in generated visual narratives
    New evaluation protocol with three evidence conditions.

pith-pipeline@v0.9.1-grok · 5789 in / 1317 out tokens · 35518 ms · 2026-07-03T21:12:14.517557+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 15 canonical work pages · 2 internal anchors

  1. [1]

    Story visualization by online text augmentation with context memory

    Daechul Ahn, Daneul Kim, Gwangmo Song, Seung Hwan Kim, Honglak Lee, Dongyeop Kang, and Jonghyun Choi. Story visualization by online text augmentation with context memory. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3125–3135, 2023

    2023

  2. [2]

    Storybench: A multifaceted benchmark for continuous story visualization

    Emanuele Bugliarello, H Hernan Moraldo, Ruben Villegas, Mohammad Babaeizadeh, Moham- mad Taghi Saffar, Han Zhang, Dumitru Erhan, Vittorio Ferrari, Pieter-Jan Kindermans, and Paul V oigtlaender. Storybench: A multifaceted benchmark for continuous story visualization. Advances in Neural Information Processing Systems, 36:78095–78125, 2023

    2023

  3. [3]

    Interleaved scene graphs for interleaved text-and-image generation assessment

    Dongping Chen, Ruoxi Chen, Shu Pu, Zhaoyi Liu, Yanru Wu, Caixi Chen, Benlin Liu, Yue Huang, Yao Wan, Pan Zhou, et al. Interleaved scene graphs for interleaved text-and-image generation assessment. InInternational Conference on Learning Representations, volume 2025, pages 74693–74756, 2025

    2025

  4. [4]

    Comm: A coherent interleaved image-text dataset for multimodal understanding and generation

    Wei Chen, Lin Li, Yongqi Yang, Bin Wen, Fan Yang, Tingting Gao, Yu Wu, and Long Chen. Comm: A coherent interleaved image-text dataset for multimodal understanding and generation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 8073–8082, 2025

    2025

  5. [5]

    Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation

    Jaemin Cho, Yushi Hu, Jason Baldridge, Roopal Garg, Peter Anderson, Ranjay Krishna, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation. InInternational conference on learning representations, volume 2024, pages 15625–15645, 2024

    2024

  6. [6]

    Story2board: a training-free approach for expressive storyboard generation.arXiv preprint arXiv:2508.09983, 2025

    David Dinkevich, Matan Levy, Omri Avrahami, Dvir Samuel, and Dani Lischinski. Story2board: a training-free approach for expressive storyboard generation.arXiv preprint arXiv:2508.09983, 2025

    work page arXiv 2025

  7. [7]

    Vista: Vi- sual storytelling using multi-modal adapters for text-to-image diffusion models

    Sibo Dong, Ismail Shaheen, Maggie Shen, Rupayan Mallick, and Sarah Adel Bargal. Vista: Vi- sual storytelling using multi-modal adapters for text-to-image diffusion models. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 12–21, 2026

    2026

  8. [8]

    Infinitystory: Unlimited video generation with world consistency and character-aware shot transitions.arXiv preprint arXiv:2603.03646, 2025

    Mohamed Elmoghany, Liangbing Zhao, Xiaoqian Shen, Subhojyoti Mukherjee, Yang Zhou, Gang Wu, Viet Dac Lai, Seunghyun Yoon, Ryan Rossi, Abdullah Rashwan, et al. Infinitystory: Unlimited video generation with world consistency and character-aware shot transitions.arXiv preprint arXiv:2603.03646, 2026

    work page arXiv 2026

  9. [9]

    Improved visual story generation with adaptive context modeling

    Zhangyin Feng, Yuchen Ren, Xinmiao Yu, Xiaocheng Feng, Duyu Tang, Shuming Shi, and Bing Qin. Improved visual story generation with adaptive context modeling. InFindings of the Association for Computational Linguistics: ACL 2023, pages 4939–4955, 2023

    2023

  10. [10]

    Vinabench: Benchmark for faithful and consistent visual narratives

    Silin Gao, Sheryl Mathew, Li Mi, Sepideh Mamooler, Mengjie Zhao, Hiromi Wakaki, Yuki Mitsufuji, Syrielle Montariol, and Antoine Bosselut. Vinabench: Benchmark for faithful and consistent visual narratives. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 2870–2879, 2025

    2025

  11. [11]

    Geneval: An object-focused framework for evaluating text-to-image alignment.Advances in Neural Information Processing Systems, 36:52132–52152, 2023

    Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment.Advances in Neural Information Processing Systems, 36:52132–52152, 2023

    2023

  12. [12]

    Narrabench: A comprehensive framework for narrative benchmarking

    Sil Hamilton, Matthew Wilkens, and Andrew Piper. Narrabench: A comprehensive framework for narrative benchmarking. InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3786–3801, 2026

    2026

  13. [13]

    Dreamstory: Open-domain story visualization by llm- guided multi-subject consistent diffusion.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Huiguo He, Huan Yang, Zixi Tuo, Yuan Zhou, Qiuyue Wang, Yuhang Zhang, Zeyu Liu, Wenhao Huang, Hongyang Chao, and Jian Yin. Dreamstory: Open-domain story visualization by llm- guided multi-subject consistent diffusion.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 11

    2025

  14. [14]

    arXiv preprint arXiv:2411.04925 , year =

    Panwen Hu, Jin Jiang, Jianqi Chen, Mingfei Han, Shengcai Liao, Xiaojun Chang, and Xiaodan Liang. Storyagent: Customized storytelling video generation via multi-agent collaboration. arXiv preprint arXiv:2411.04925, 2024

    work page arXiv 2024

  15. [15]

    Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering

    Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406–20417, 2023

    2023

  16. [16]

    arXiv preprint arXiv:2512.16853 , year=

    Amita Kamath, Kai-Wei Chang, Ranjay Krishna, Luke Zettlemoyer, Yushi Hu, and Marjan Ghazvininejad. Geneval 2: Addressing benchmark drift in text-to-image evaluation.arXiv preprint arXiv:2512.16853, 2025

    work page arXiv 2025

  17. [17]

    arXiv preprint arXiv:2406.13743 , year=

    Baiqi Li, Zhiqiu Lin, Deepak Pathak, Jiayao Li, Yixin Fei, Kewen Wu, Tiffany Ling, Xide Xia, Pengchuan Zhang, Graham Neubig, et al. Genai-bench: Evaluating and improving compositional text-to-visual generation.arXiv preprint arXiv:2406.13743, 2024

    work page arXiv 2024

  18. [18]

    Narratology meets text-to-image: a survey of consistency in ai generated storybook illustrations.Artificial Intelligence Review, 2026

    Zhedong Lin, Zhongsheng Wang, Qian Liu, Xinyu Zhang, and Jiamou Liu. Narratology meets text-to-image: a survey of consistency in ai generated storybook illustrations.Artificial Intelligence Review, 2026

    2026

  19. [19]

    Intelli- gent grimm-open-ended visual storytelling via latent diffusion models

    Chang Liu, Haoning Wu, Yujie Zhong, Xiaoyun Zhang, Yanfeng Wang, and Weidi Xie. Intelli- gent grimm-open-ended visual storytelling via latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6190–6200, 2024

    2024

  20. [20]

    One-prompt-one-story: Free-lunch consistent text-to-image generation using a single prompt

    Tao Liu, Kai Wang, Senmao Li, Joost Van de Weijer, Fahad Khan, Shiqi Yang, Yaxing Wang, Jian Yang, and Ming-Ming Cheng. One-prompt-one-story: Free-lunch consistent text-to-image generation using a single prompt. InInternational Conference on Learning Representations, volume 2025, pages 24470–24497, 2025

    2025

  21. [21]

    Lay2story: extending diffusion transformers for layout-togglable story generation

    Ao Ma, Jiasong Feng, Ke Cao, Jing Wang, Yun Wang, Quanwei Zhang, and Zhanjie Zhang. Lay2story: extending diffusion transformers for layout-togglable story generation. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision, pages 16102–16111, 2025

    2025

  22. [22]

    Storydall-e: Adapting pretrained text-to- image transformers for story continuation

    Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Storydall-e: Adapting pretrained text-to- image transformers for story continuation. InEuropean conference on computer vision, pages 70–87. Springer, 2022

    2022

  23. [23]

    Story-iter: A training-free iterative paradigm for long story visualization

    Jiawei Mao, Xiaoke Huang, Yunfei Xie, Yuanqi Chang, Mude Hui, Bingjie Xu, Zeyu Zheng, Zirui Wang, Cihang Xie, and Yuyin Zhou. Story-iter: A training-free iterative paradigm for long story visualization. InThe Fourteenth International Conference on Learning Representations

  24. [24]

    Logistory: A logic-aware framework for multi-image story visualization.arXiv preprint arXiv:2603.28082, 2026

    Chutian Meng, Fan Ma, Chi Zhang, Jiaxu Miao, Yi Yang, and Yueting Zhuang. Logistory: A logic-aware framework for multi-image story visualization.arXiv preprint arXiv:2603.28082, 2026

    work page arXiv 2026

  25. [25]

    Make-a-story: Visual memory conditioned consistent story generation

    Tanzila Rahman, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, Shweta Mahajan, and Leonid Sigal. Make-a-story: Visual memory conditioned consistent story generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2493–2502, 2023

    2023

  26. [26]

    Redistory: Region-disentangled diffusion for consistent visual story generation.arXiv preprint arXiv:2602.01303, 2026

    Ayushman Sarkar, Zhenyu Yu, Chu Chen, Wei Tang, Kangning Cui, and Mohd Yamani Idna Idris. Redistory: Region-disentangled diffusion for consistent visual story generation.arXiv preprint arXiv:2602.01303, 2026

    work page arXiv 2026

  27. [27]

    Storygpt-v: Large language models as consistent story visualizers

    Xiaoqian Shen and Mohamed Elhoseiny. Storygpt-v: Large language models as consistent story visualizers. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 13273–13283, 2025

    2025

  28. [28]

    Animaker: Multi-agent animated storytelling with mcts-driven clip generation

    Haoyuan Shi, Yunxin Li, Xinyu Chen, Longyue Wang, Baotian Hu, and Min Zhang. Animaker: Multi-agent animated storytelling with mcts-driven clip generation. InProceedings of the SIGGRAPH Asia 2025 Conference Papers, pages 1–11, 2025. 12

    2025

  29. [29]

    Msvbench: Towards human-level evaluation of multi-shot video generation

    Haoyuan Shi, Yunxin Li, Nanhao Deng, Zhenran Xu, Xinyu Chen, Longyue Wang, Baotian Hu, and Min Zhang. Msvbench: Towards human-level evaluation of multi-shot video generation. arXiv preprint arXiv:2602.23969, 2026

    work page arXiv 2026

  30. [30]

    Storybooth: Training-free multi-subject consistency for improved visual storytelling

    Jaskirat Singh, Junshen K Chen, Jonas Kohler, and Michael Cohen. Storybooth: Training-free multi-subject consistency for improved visual storytelling. InInternational Conference on Learning Representations, volume 2025, pages 48678–48690, 2025

    2025

  31. [31]

    AttriStory: Fine-grained Attribute Realization for Visual Storytelling with Diffusion Models

    Manogna Sreenivas, Rohit Kumar, and Soma Biswas. Attristory: Fine-grained attribute realiza- tion for visual storytelling with diffusion models.arXiv preprint arXiv:2605.20777, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  32. [32]

    Storyimager: A unified and efficient framework for coherent story visualization and completion

    Ming Tao, Bing-Kun Bao, Hao Tang, Yaowei Wang, and Changsheng Xu. Storyimager: A unified and efficient framework for coherent story visualization and completion. InEuropean Conference on Computer Vision, pages 479–495. Springer, 2024

    2024

  33. [33]

    arXiv preprint arXiv:2503.05242 , year =

    Xuenan Xu, Jiahao Mei, Chenliang Li, Yuning Wu, Ming Yan, Shaopeng Lai, Ji Zhang, and Mengyue Wu. Mm-storyagent: Immersive narrated storybook video generation with a multi- agent paradigm across text, image and audio.arXiv preprint arXiv:2503.05242, 2025

    work page arXiv 2025

  34. [34]

    Seed-story: Multimodal long story generation with large language model

    Shuai Yang, Yuying Ge, Yang Li, Yukang Chen, Yixiao Ge, Ying Shan, and Ying-Cong Chen. Seed-story: Multimodal long story generation with large language model. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1850–1860, 2025

    2025

  35. [35]

    Openstory++: A large-scale dataset and benchmark for instance-aware open-domain visual storytelling.arXiv preprint arXiv:2408.03695, 2024

    Zilyu Ye, Jinxiu Liu, Ruotian Peng, Jinjin Cao, Zhiyang Chen, Yiyang Zhang, Ziwei Xuan, Mingyuan Zhou, Xiaoqian Shen, Mohamed Elhoseiny, et al. Openstory++: A large-scale dataset and benchmark for instance-aware open-domain visual storytelling.arXiv preprint arXiv:2408.03695, 2024

    work page arXiv 2024

  36. [36]

    MuSS: A Large-Scale Dataset and Cinematic Narrative Benchmark for Multi-Shot Subject-to-Video Generation

    Haojie Zhang, Di Wu, Bingyan Liu, Linjie Zhong, Yuancheng Wei, Xingsong Ye, Nanqing Liu, and Yaling Liang. Muss: A large-scale dataset and cinematic narrative benchmark for multi-shot subject-to-video generation.arXiv preprint arXiv:2604.23789, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  37. [37]

    arXiv preprint arXiv:2512.19539 , year=

    Kaiwen Zhang, Liming Jiang, Angtian Wang, Jacob Zhiyuan Fang, Tiancheng Zhi, Qing Yan, Hao Kang, Xin Lu, and Xingang Pan. Storymem: Multi-shot long video storytelling with memory.arXiv preprint arXiv:2512.19539, 2025

    work page arXiv 2025

  38. [38]

    Contextualstory: Consistent visual storytelling with spatially- enhanced and storyline context

    Sixiao Zheng and Yanwei Fu. Contextualstory: Consistent visual storytelling with spatially- enhanced and storyline context. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 10617–10625, 2025

    2025

  39. [39]

    Videomemory: Toward consis- tent video generation via memory integration.arXiv preprint arXiv:2601.03655, 2026

    Jinsong Zhou, Yihua Du, Xinli Xu, Luozhou Wang, Zijie Zhuang, Yehang Zhang, Shuaibo Li, Xiaojun Hu, Bolan Su, and Ying-cong Chen. Videomemory: Toward consistent video generation via memory integration.arXiv preprint arXiv:2601.03655, 2026

    work page arXiv 2026

  40. [40]

    Opening: A comprehensive benchmark for judging open-ended interleaved image-text generation

    Pengfei Zhou, Xiaopeng Peng, Jiajun Song, Chuanhao Li, Zhaopan Xu, Yue Yang, Ziyao Guo, Hao Zhang, Yuqi Lin, Yefei He, et al. Opening: A comprehensive benchmark for judging open-ended interleaved image-text generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 56–66, 2025

    2025

  41. [41]

    Storydiffusion: Consistent self-attention for long-range image and video generation.Advances in Neural Information Processing Systems, 37:110315–110340, 2024

    Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation.Advances in Neural Information Processing Systems, 37:110315–110340, 2024

    2024

  42. [42]

    human_evidence_accepted

    Cailin Zhuang, Ailin Huang, Yaoqi Hu, Jingwei Wu, Wei Cheng, Jiaqi Liao, Hongyuan Wang, Xinyao Liao, Weiwei Cai, Hengyuan Xu, et al. Vistorybench: Comprehensive benchmark suite for story visualization.arXiv preprint arXiv:2505.24862, 2025. 13 Appendix Appendix Contents A Dataset Construction and Human Validation Details 17 A.1 Construction Overview . . . ...

    work page arXiv 2025