Pith. sign in

REVIEW 5 major objections 5 minor 55 references

StorySync: Training-Free Subject Consistency in Text-to-Image Generation via Region Harmonization

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read StorySync claims that text-to-image diffusion models can generate story-consistent characters with no extra training by sharing attention only inside subject regions and harmonizing those regions across a batch of images.

desk verdict Honest, coherent training-free consistency method, but the supplied text lacks the quantitative evidence needed to verify the 'superior performance' claim. read the letter →

arxiv 2508.03735 v1 pith:PNJX6PJM submitted 2025-07-31 cs.CV cs.AI

Gopalji Gaur , Mohammadreza Zolfaghari , Thomas Brox This is my paper
classification cs.CVcs.AI
keywords training-freesubjectconsistencytext-to-imagediffusioncross-imageattentionsharingregionalfeatureharmonizationvisualstorytellingSDXLKandinsky3FLUX.1-schnell
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a text-to-image diffusion model can generate a batch of images telling one visual story with the same character in every scene, without any fine-tuning or per-subject training. The proposed method, StorySync, edits attention during generation so that subject features are shared across images only inside masked subject regions, then harmonizes fine visual details within those regions and interpolates base layouts to keep scenes diverse. If the claim holds, animators, game artists, and storyboarders could get consistent characters from off-the-shelf models like SDXL, Kandinsky 3, and FLUX.1-schnell while preserving each prompt's intent. The paper reports that this matches or beats existing training-free approaches on both subject consistency and prompt adherence.

What carries the argument

The load-bearing object is the subject mask, derived from the model's cross-attention maps, which connect each text token to image regions. At selected attention layers, StorySync replaces ordinary attention with masked cross-image attention sharing: queries attend only to keys and values inside the subject masks of every image in the batch, so the same subject's features are aligned across scenes. Regional Feature Harmonization then refines subtle details inside the masks by matching visually similar regions, and Base Layout Interpolation injects layout diversity by interpolating base layouts. These three mechanisms together carry the consistency claim.

What would settle it

Run StorySync on a story where the subject is partially occluded or blends with the background, so the internal attention maps are known to mislocate it, and measure same-subject similarity against easy-pose prompts; a sharp drop confirms the paper's stated mask-dependence limitation, while no drop would mean the method is more robust than its own limitation suggests.

Watch

Extended reading notes

Core claim

StorySync's central claim is that masked cross-image attention sharing, Regional Feature Harmonization, and Base Layout Interpolation together give zero-shot, training-free subject consistency across a story sequence without sacrificing prompt fidelity or creative diversity. At selected denoising steps, the method uses cross-attention maps to produce a subject mask per image, restricts attention to the masked subject regions across all images in the batch, and thereby aligns the subject's visual features dynamically across scenes. Regional Feature Harmonization then refines visually similar details such as eyes, fur, or clothing patterns, while Base Layout Interpolation prevents the images from collapsing into identical poses. The paper further claims the whole pipeline is model-agnostic and works with both U-Net-based models and transformer-based models without additional training.

Load-bearing premise

The load-bearing premise is that each image's subject mask, taken from the model's internal attention maps, reliably finds the same character in every scene; the paper's own limitation section says wrong masks block attention sharing and cause the character to drift.

Editorial extensions

If this is right

  • StorySync can be integrated into SDXL, Kandinsky 3, and FLUX.1-schnell with no retraining, producing consistent subjects on each architecture.
  • On the paper's comparisons, it outperforms existing training-free consistent-subject methods on both subject consistency and prompt adherence.
  • It preserves creative diversity: pose and layout variation are enforced so the character adapts to each scene instead of repeating one pose.
  • Story-generation workflows such as storyboards, animation previsualization, and game concept art can add consistent characters without per-character finetuning.
  • Because it works on both U-Net and transformer-based diffusion backbones, the mechanism appears architecture-agnostic.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the method's ceiling is set by mask accuracy; replacing cross-attention-derived masks with stronger segmentation or semantic-correspondence masks could plausibly improve consistency further.
  • Beyond the paper, the color- and texture-based region harmonization could be extended to semantic correspondences, which might handle multiple subjects and non-rigid deformations more robustly.
  • Beyond the paper, a direct testable extension is multi-subject story generation: if mask extraction can be made reliable for several entities in one image, the same masked-sharing scheme could keep each character distinct and consistent.
  • Beyond the paper, the zero-training property suggests StorySync could be composed with other plug-in controls such as layout or pose editors without retraining the base model.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes StorySync, a training-free method for subject-consistent text-to-image generation. It introduces masked cross-image attention sharing, Regional Feature Harmonization, and Base Layout Interpolation, applied at inference time to SDXL, Kandinsky 3, and FLUX.1-schnell. The abstract and conclusion claim that StorySync achieves superior subject consistency and prompt adherence compared with state-of-the-art training-free approaches, while preserving the creative abilities of the pretrained model. The supplied text contains qualitative results (Figures 5-7), ablation observations, and a limitations section, but it does not include the method equations, the mask-extraction procedure, quantitative comparison tables, or metric definitions needed to verify the central empirical claims.

Significance. If the stated claims hold, the contribution is practically significant: a model-agnostic, training-free attention-editing method that maintains subject identity across scenes would be useful for storytelling, animation, and synthetic data generation. The qualitative demonstrations across three different diffusion backbones and the explicit discussion of mask-related failure modes are strengths. However, the manuscript as supplied provides no quantitative evaluation of subject consistency or prompt adherence, no baseline numbers against StoryDiffusion or ConsiStory, and no details of the proposed mechanisms beyond high-level descriptions. The central claim of superiority over state-of-the-art training-free methods is therefore currently unverified rather than established.

major comments (5)
  1. [Section 5 (Conclusion); Section 4 (Experiments)] The central empirical claim is not verifiable from the supplied text: there are no quantitative tables or metrics for subject consistency or prompt adherence, and no comparison numbers against StoryDiffusion [54] or ConsiStory [43]. The conclusion's statement of 'superior performance' in both subject consistency and prompt adherence therefore lacks supporting evidence. Please add a full quantitative evaluation, including identity-similarity metrics, prompt-fidelity metrics, and direct baseline comparisons.
  2. [Section 3 (Method); Appendix A.4] The method is not described in the supplied text: Section 3 is absent, and the three named components—masked cross-image attention sharing, Regional Feature Harmonization, and Base Layout Interpolation—are presented only by name in the Introduction. There are no equations, algorithms, mask-generation details, or threshold values. This makes the results impossible to reproduce or check. Please provide the complete method section, including how subject masks are extracted from cross-attention maps and how the harmonization objective is defined.
  3. [Section 4.4 (Limitations); Figure 7a] The load-bearing assumption that cross-attention-derived subject masks are accurate is acknowledged as fragile: Section 4.4 states that incorrect masks block attention sharing and cause inconsistency, and Figure 7a illustrates this failure. Since all cross-image attention sharing is gated by these masks, the manuscript should report mask-quality statistics, such as overlap with reference segmentations, and a sensitivity analysis for the mask threshold and the regional-feature correspondence threshold. Currently no such analysis is provided.
  4. [Section 4.3 (Ablation); Figure 6] The ablation study is qualitative only, and the text is internally inconsistent about this: the paragraph preceding Figure 6 says the components are studied 'in a quantitative manner,' but then refers to a 'qualitative ablation study in Figure 7.' The referenced quantitative ablation table and the associated numbers are not present. Please include the quantitative ablation results and clarify which figure contains them.
  5. [Section 5 (Conclusion); Figure 5] The claim that StorySync 'can be integrated with any state-of-the-art diffusion model' is supported only by qualitative examples. In particular, no details are given for how masked cross-image attention is implemented in the time-distilled transformer-based FLUX.1-schnell architecture as opposed to the U-Net architectures of SDXL and Kandinsky 3. Please specify the architecture-specific adaptations and, ideally, provide per-model quantitative results.
minor comments (5)
  1. [Abstract] The phrase 'we follow a training-free approach' is awkward; consider 'we take a training-free approach' for clarity.
  2. [Section 4.3 (Ablation)] The sentence 'we have included the results of the qualitative ablation study in Figure 7' appears to conflate qualitative and quantitative evaluation; if numerical results are intended, the wording should be corrected.
  3. [References] The reference list contains duplicate entries: [10] and [11] both cite Gal et al.’s Textual Inversion paper, and [24] and [25] both cite Intelligent Grimm with slightly different formatting. Please deduplicate and standardize the bibliography.
  4. [Figure 6] The caption refers to a red circle in the second row and to a prompt about a cat on a dirt road, but the corresponding regions are not visible in the supplied figure; please ensure the figure is legible and the annotations are clearly rendered.
  5. [Section 4.4 (Limitations)] The claim that mask and harmonization failures 'occur in only a small fraction of cases' is not quantified; please report the observed failure rate or provide the basis for this statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: StorySync's method and evaluations rest on external baselines and standard diffusion components, with no self-referential definitions or fitted-input predictions in the supplied text.

full rationale

The supplied manuscript contains no visible circular derivation. The proposed StorySync pipeline introduces masked cross-image attention sharing, Regional Feature Harmonization, and Base Layout Interpolation; these are described as new mechanisms operating on pretrained diffusion models, not as quantities defined by the evaluation metric or by the consistency outcome they are claimed to produce. No fitted parameters are introduced and then reported as predictions: the method is training-free and uses cross-attention-derived masks, and the paper's own limitation section acknowledges that incorrect masks can block attention sharing and cause inconsistency, which is an honest dependence rather than a self-justifying assumption. The comparisons are against external training-free baselines (StoryDiffusion and ConsiStory are named as prior work), and the claimed superiority is an empirical claim backed by qualitative figures; even though quantitative tables are missing from the supplied excerpt, absence of evidence is not circularity. The only self-citations that appear are to standard architecture references such as U-Net, which are not load-bearing in the sense of proving the method's uniqueness or substituting for derivation. No step reduces by construction to its own input, no citation chain is used to forbid alternatives, and no known empirical pattern is merely renamed. Therefore the circularity score is 0. Concerns about unvalidated mask accuracy or missing quantitative comparisons belong to correctness or completeness risk, not to circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim depends on the reliability of cross-attention-derived subject masks and on the assumption that inference-time attention edits preserve model behavior. No explicit numeric parameter values are provided in the extracted text, so thresholds and layer choices are treated as unreported free parameters. No new physical or conceptual entities are introduced.

free parameters (3)
  • cross-attention mask threshold = not reported in provided text
    Subject masks are derived from cross-attention maps, and binarizing them requires a threshold. Section 4.4 says inaccurate masks block attention sharing and cause inconsistency.
  • RFH region correspondence threshold = not reported in provided text
    Regional Feature Harmonization must decide which regions are visually similar; this requires a similarity threshold or matching radius. Section 4.4 states misalignment can deform the subject.
  • attention sharing layer or timestep selection = not reported in provided text
    The method likely shares attention only at selected layers or denoising steps, a design choice that trades consistency strength against prompt adherence.
assumptions (3)
  • domain assumption Cross-attention maps in pretrained diffusion models reliably localize the same subject across different prompts and images.
    Used to construct subject masks for masked cross-image attention sharing; the paper itself notes mask failures as a limitation.
  • domain assumption Inference-time modification of attention maps preserves the model's other generation capabilities and does not require retraining.
    The entire pipeline is training-free and relies on pretrained attention being safely controllable at inference time.
  • domain assumption A batch of images generated under different prompts but containing the same subject can share attention features without merging unrelated objects.
    Cross-image attention sharing assumes the subject is common across images and that other content remains independent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of StorySync: Training-Free Subject Consistency in Text-to-Image Generation via Region Harmonization." pith.science (2026). https://pith.science/paper/PNJX6PJM

@misc{pith2026250803735,
  author       = {Pith},
  title        = {Pith review of: StorySync: Training-Free Subject Consistency in Text-to-Image Generation via Region Harmonization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PNJX6PJM}},
  note         = {Machine review of arXiv:2508.03735}
}
read the original abstract

Generating a coherent sequence of images that tells a visual story, using text-to-image diffusion models, often faces the critical challenge of maintaining subject consistency across all story scenes. Existing approaches, which typically rely on fine-tuning or retraining models, are computationally expensive, time-consuming, and often interfere with the model's pre-existing capabilities. In this paper, we follow a training-free approach and propose an efficient consistent-subject-generation method. This approach works seamlessly with pre-trained diffusion models by introducing masked cross-image attention sharing to dynamically align subject features across a batch of images, and Regional Feature Harmonization to refine visually similar details for improved subject consistency. Experimental results demonstrate that our approach successfully generates visually consistent subjects across a variety of scenarios while maintaining the creative abilities of the diffusion model.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 27 canonical work pages

  1. [54]

    Storydiffusion: Consis- tent self-attention for long-range image and video genera- tion

    Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Ji- ashi Feng, and Qibin Hou. Storydiffusion: Consis- tent self-attention for long-range image and video genera- tion. Advances in Neural Information Processing Systems , 37:110315–110340, 2024

    work page 2024

  2. [43]

    Training-free consis- tent text-to-image generation

    Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consis- tent text-to-image generation. ACM Transactions on Graph- ics (TOG), 43(4):1–18, 2024

    2024

  3. [1]

    Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H. Bermano. Domain- agnostic tuning-encoder for fast personalization of text-to- image models. SIGGRAPH Asia 2023 Conference Papers , 2023

    work page 2023

  4. [2]

    Kandinsky 3.0 technical report, 2024

    Vladimir Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Aga- fonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report, 2024

    work page 2024

  5. [3]

    The chosen one: Consistent characters in text-to- image diffusion models

    Omri Avrahami, Amir Hertz, Yael Vinker, Moab Arar, Shlomi Fruchter, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. The chosen one: Consistent characters in text-to- image diffusion models. In Special Interest Group on Com- puter Graphics and Interactive Techniques Conference Con- ference Papers ’24, SIGGRAPH ’24, page 1–12. ACM, July 2024

    work page 2024

  6. [4]

    ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers

    Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji- aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. ArXiv, abs/2211.01324, 2022

    arXiv 2022

  7. [5]

    Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and edit- ing

    Ming Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and edit- ing. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22503–22513, 2023

    work page 2023

  8. [6]

    DreamArtist: Towards Controllable One-Shot Text-to-Image Genera- tion via Positive-Negative Prompt-Tuning, Apr

    Ziyi Dong, Pengxu Wei, and Liang Lin. DreamArtist: Towards Controllable One-Shot Text-to-Image Genera- tion via Positive-Negative Prompt-Tuning, Apr. 2023. arXiv:2211.11337 [cs]

    arXiv 2023

Show all 55 references
  1. [7]

    Tam- ing transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bj ¨orn Ommer. Tam- ing transformers for high-resolution image synthesis. CoRR, abs/2012.09841, 2020

    2012 arXiv

  2. [8]

    Improved vi- sual story generation with adaptive context modeling

    Zhangyin Feng, Yuchen Ren, Xinmiao Yu, Xiaocheng Feng, Duyu Tang, Shuming Shi, and Bing Qin. Improved vi- sual story generation with adaptive context modeling. arXiv preprint arXiv:2305.16811, 2023

    2023 arXiv

  3. [9]

    Dream- sim: Learning new dimensions of human visual similar- ity using synthetic data

    Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dream- sim: Learning new dimensions of human visual similar- ity using synthetic data. arXiv preprint arXiv:2306.09344 , 2023

    2023 arXiv

  4. [11]

    Bermano, Gal Chechik, and Daniel Cohen-Or

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An Image is Worth One Word: Personalizing Text-to- Image Generation using Textual Inversion, Aug. 2022. arXiv:2208.01618 [cs]

    2022 arXiv

  5. [12]

    Bermano, Gal Chechik, and Daniel Cohen-Or

    Rinon Gal, Moab Arar, Yuval Atzmon, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based Domain Tuning for Fast Personalization of Text-to-Image Models, Mar. 2023. arXiv:2302.12228 [cs]

    2023 arXiv

  6. [13]

    Talecrafter: Interactive story visualization with multiple characters

    Yuan Gong, Youxin Pang, Xiaodong Cun, Menghan Xia, Yingqing He, Haoxin Chen, Longyue Wang, Yong Zhang, Xintao Wang, Ying Shan, et al. Talecrafter: Interactive story visualization with multiple characters. arXiv preprint arXiv:2305.18247, 2023

    2023 arXiv

  7. [14]

    Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models

    Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yun- peng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Wei Wu, Yixiao Ge, Ying Shan, and Mike Zheng Shou. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. ArXiv, abs...

    2023 arXiv

  8. [15]

    Dreamstory: Open-domain story visualiza- tion by llm-guided multi-subject consistent diffusion

    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 visualiza- tion by llm-guided multi-subject consistent diffusion. arXiv preprint arXiv:2407.12899, 2024

    2024 arXiv

  9. [16]

    Anystory: Towards unified single and multiple subject personalization in text-to-image generation

    Junjie He, Yuxiang Tuo, Binghui Chen, Chongyang Zhong, Yifeng Geng, and Liefeng Bo. Anystory: Towards unified single and multiple subject personalization in text-to-image generation. arXiv preprint arXiv:2501.09503, 2025

    2025 arXiv

  10. [17]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. arXiv preprint arxiv:2006.11239, 2020

    2006 arXiv

  11. [18]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models, Oct. 2021. arXiv:2106.09685 [cs]

    2021 arXiv

  12. [19]

    Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin C. K. Chan, and Ziwei Liu. ReVersion: Diffusion-Based Relation Inver- sion from Images, Mar. 2023. arXiv:2303.13495 [cs]

    2023 arXiv

  13. [20]

    Zero-shot generation of coherent storybook from plain text story using diffusion models

    Hyeonho Jeong, Gihyun Kwon, and Jong Chul Ye. Zero-shot generation of coherent storybook from plain text story using diffusion models. arXiv preprint arXiv:2302.03900, 2023

    2023 arXiv

  14. [21]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick. Segment anything, 2023

    2023

  15. [22]

    Multi-Concept Customization of Text-to-Image Diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-Concept Customization of Text-to-Image Diffusion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1931–1941, Vancouver, BC, Canada, June 2023. IEEE

    2023

  16. [23]

    Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to- image generation and editing. ArXiv, abs/2305.14720, 2023

    2023 arXiv

  17. [24]

    Intelligent grimm-open-ended visual storytelling via latent diffusion models

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

    2024

  18. [25]

    Intelligent grimm - open-ended visual story- telling via latent diffusion models

    Chang Liu, Haoning Wu, Yujie Zhong, Xiaoyu Zhang, and Weidi Xie. Intelligent grimm - open-ended visual story- telling via latent diffusion models. 2024 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) , pages 6190–6200, 2023

    2024

  19. [26]

    Grounding dino: Marry- ing dino with grounded pre-training for open-set object de- tection, 2024

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding dino: Marry- ing dino with grounded pre-training for open-set object de- tection, 2024

    2024

  20. [27]

    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, Fa- had Shahbaz 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. arXiv preprint arXiv:2501.13554, 2025

    2025 arXiv

  21. [28]

    Summary of chatgpt-related research and per- spective towards the future of large language models

    Yiheng Liu, Tianle Han, Siyuan Ma, Jiayue Zhang, Yuanyuan Yang, Jiaming Tian, Hao He, Antong Li, Meng- shen He, Zhengliang Liu, Zihao Wu, Lin Zhao, Dajiang Zhu, Xiang Li, Ning Qiang, Dingang Shen, Tianming Liu, and Bao Ge. Summary of chatgpt-related research and per- spective ...

    2023

  22. [29]

    A threshold selection method from gray- level histograms

    Nobuyuki Otsu. A threshold selection method from gray- level histograms. IEEE Trans. Syst. Man Cybern., 9:62–66, 1979

    1979

  23. [30]

    Synthesizing coherent story with auto-regressive la- tent diffusion models

    Xichen Pan, Pengda Qin, Yuhong Li, Hui Xue, and Wenhu Chen. Synthesizing coherent story with auto-regressive la- tent diffusion models. 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages 2908– 2918, 2022

    2024

  24. [31]

    Localizing object-level shape variations with text-to-image diffusion models

    Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch- Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22994–23004, 2023

    2023

  25. [32]

    Orthogonal adaptation for modular customization of diffusion models

    Ryan Po, Guandao Yang, Kfir Aberman, and Gordon Wet- zstein. Orthogonal adaptation for modular customization of diffusion models. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 7964–7973, 2023

    2024

  26. [33]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023

    2023 arXiv

  27. [34]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  28. [35]

    Conceptlab: Creative generation using diffusion prior constraints

    Elad Richardson, Kfir Goldberg, Yuval Alaluf, and Daniel Cohen-Or. Conceptlab: Creative generation using diffusion prior constraints. ArXiv, abs/2308.02669, 2023

    2023 arXiv

  29. [36]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. CoRR, abs/2112.10752, 2021

    2021 arXiv

  30. [37]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015

    2015 arXiv

  31. [38]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2023 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 22500–22510, 2022

    2023

  32. [39]

    Instant- booth: Personalized text-to-image generation without test- time finetuning

    Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instant- booth: Personalized text-to-image generation without test- time finetuning. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 8543–8552, 2023

    2024

  33. [40]

    Make-a-storyboard: A general framework for storyboard with disentangled and merged control

    Sitong Su, Litao Guo, Lianli Gao, Hengtao Shen, and Jingkuan Song. Make-a-storyboard: A general framework for storyboard with disentangled and merged control. ArXiv, abs/2312.07549, 2023

    2023 arXiv

  34. [41]

    Emergent correspondence from image diffusion

    Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. ArXiv, abs/2306.03881, 2023

    2023 arXiv

  35. [42]

    Add-it: Training-free object in- sertion in images with pretrained diffusion models

    Yoad Tewel, Rinon Gal, Dvir Samuel, Yuval Atzmon, Lior Wolf, and Gal Chechik. Add-it: Training-free object in- sertion in images with pretrained diffusion models. arXiv preprint arXiv:2411.07232, 2024

    2024 arXiv

  36. [44]

    Oneactor: Consistent character generation via cluster- conditioned guidance

    Jiahao Wang, Caixia Yan, Haonan Lin, Weizhan Zhang, Mengmeng Wang, Tieliang Gong, Guang Dai, and Hao Sun. Oneactor: Consistent character generation via cluster- conditioned guidance. arXiv preprint arXiv:2404.10267 , 2024

    2024 arXiv

  37. [45]

    Spotactor: Training-free layout- controlled consistent image generation

    Jiahao Wang, Caixia Yan, Weizhan Zhang, Haonan Lin, Mengmeng Wang, Guang Dai, Tieliang Gong, Hao Sun, and Jingdong Wang. Spotactor: Training-free layout- controlled consistent image generation. arXiv preprint arXiv:2409.04801, 2024

    2024 arXiv

  38. [46]

    Characterfactory: Sampling consis- tent characters with gans for diffusion models.arXiv preprint arXiv:2404.15677, 2024

    Qinghe Wang, Baolu Li, Xiaomin Li, Bing Cao, Liqian Ma, Huchuan Lu, and Xu Jia. Characterfactory: Sampling consis- tent characters with gans for diffusion models.arXiv preprint arXiv:2404.15677, 2024

    2024 arXiv

  39. [47]

    Elite: Encoding visual concepts into textual embeddings for customized text-to-image gener- ation

    Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image gener- ation. 2023 IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 15897–15907, 2023

    2023

  40. [48]

    Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation

    Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference...

    2023

  41. [49]

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

    Shuai Yang, Yuying Ge, Yang Li, Yukang Chen, Yixiao Ge, Ying Shan, and Yingcong Chen. Seed-story: Multimodal long story generation with large language model. arXiv preprint arXiv:2407.08683, 2024

    2024 arXiv

  42. [50]

    Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models. arXiv preprint arXiv:2308.06721, 2023

    2023 arXiv

  43. [51]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018

    2018

  44. [52]

    Inversion-based Style Transfer with Diffusion Models

    Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based Style Transfer with Diffusion Models. In 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR) , pages 10146–10156, Vancouver, BC, Canada, Ju...

    2023

  45. [53]

    Contextualstory: Consistent visual storytelling with spatially-enhanced and storyline con- text

    Sixiao Zheng and Yanwei Fu. Contextualstory: Consistent visual storytelling with spatially-enhanced and storyline con- text. arXiv preprint arXiv:2407.09774, 2024

    2024 arXiv

  46. [55]

    Storymaker: Towards holistic consistent characters in text-to-image generation

    Zhengguang Zhou, Jing Li, Huaxia Li, Nemo Chen, and Xu Tang. Storymaker: Towards holistic consistent characters in text-to-image generation. arXiv preprint arXiv:2409.12576, 2024

    2024 arXiv

  47. [56]

    DomainStudio: Fine-Tuning Diffusion Models for Domain- Driven Image Generation using Limited Data, Aug

    Jingyuan Zhu, Huimin Ma, Jiansheng Chen, and Jian Yuan. DomainStudio: Fine-Tuning Diffusion Models for Domain- Driven Image Generation using Limited Data, Aug. 2023. arXiv:2306.14153 [cs]

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.