pith. sign in

arxiv: 2606.05635 · v1 · pith:OVJTGAZ3new · submitted 2026-06-04 · 💻 cs.CV · cs.MM

ShotCrop³: Cropping Human-Centric Images into Cinematic Triple-Shot Compositions

Dehong Kong , Lina Lei , Lingtao Zheng , Chenyang Wu , Ailing Zhang , Xinran Qin , Teng Ma , Jiaqi Xu
show 5 more authors
Zhixin Wang Zhikai Chen Xuecheng Qi Renjing Pei Fan Li
This is my paper

Pith reviewed 2026-06-28 02:29 UTC · model grok-4.3

classification 💻 cs.CV cs.MM
keywords cinematic croppingtriple-shot compositionimage compositionvisual narrationmulti-shot generationreinforcement learningpseudo-labeling
0
0 comments X

The pith

ShotCrop crops one human-centric image into three cinematic shots—establishing, medium, and close-up—each with a description, reaching 2.82 times the shot localization accuracy of GPT-5.

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

The paper tries to establish a method that generates three narrative shots from a single image instead of one aesthetic crop. This would matter for workflows like poster design and visual storytelling that need multiple views to show context, subject emphasis, and detail. ShotCrop reaches this by running a model through chain-of-thought supervised fine-tuning, then semi-supervised training on high-confidence pseudo labels, and finally reinforcement learning with a custom reward. If the claim holds, automated tools could produce ready-to-use shot sets paired with descriptions from ordinary photos.

Core claim

The central claim is that Triple-Shot Compositions can be learned from limited expert data by training ShotCrop in three stages: Chain-of-Thought supervised fine-tuning to build basic shot-cropping reasoning, semi-supervised fine-tuning that retains only high-confidence pseudo labels produced by combining MLLM scoring, aesthetic assessment, and CLIP similarity, and final optimization with Group Relative Policy Optimization using a composite reward; on the resulting TSC-Bench of 1.2k expert cases the model improves average shot localization accuracy by a factor of 2.82 over GPT-5.

What carries the argument

The three-stage training pipeline of ShotCrop that first establishes reasoning via Chain-of-Thought fine-tuning, then augments data with multi-signal pseudo labels, and finally refines the policy through GRPO-S with a reward that scores localization, aesthetics, and narrative fit.

If this is right

  • Commercial poster and storyboarding tools can output ready sets of three narrative crops with descriptions from one source image.
  • The TSC-Bench supplies a fixed testbed for comparing future multi-shot composition models.
  • The same staged training pattern can be reused for other tasks that need both localization and aesthetic quality when expert labels are scarce.
  • The composite reward used in GRPO-S directly ties model updates to measurable improvements in shot positioning and narrative coverage.

Where Pith is reading between the lines

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

  • The approach could be tested on non-human subjects or full video frames to see whether the pseudo-labeling still produces reliable signals outside the human-centric case.
  • Embedding the model in consumer photo apps would let ordinary users generate multiple narrative crops automatically rather than cropping by hand.
  • Running the pseudo-label pipeline on images with unusual lighting or poses would reveal whether any of the three scoring signals introduces consistent errors.

Load-bearing premise

The pseudo-labeling strategy that merges MLLM scoring, aesthetic assessment, and CLIP similarity produces training signals that match expert judgment without systematic bias or noise.

What would settle it

Measure shot localization accuracy and description quality of ShotCrop versus GPT-5 on a fresh set of several hundred expert-annotated human-centric images; if the factor of improvement falls well below 2.82 or varies sharply by shot type, the central performance claim does not hold.

Figures

Figures reproduced from arXiv: 2606.05635 by Ailing Zhang, Chenyang Wu, Dehong Kong, Fan Li, Jiaqi Xu, Lina Lei, Lingtao Zheng, Renjing Pei, Teng Ma, Xinran Qin, Xuecheng Qi, Zhikai Chen, Zhixin Wang.

Figure 1
Figure 1. Figure 1: Given a single human-centric image, our framework outputs a cinematic triple-shot [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our method ShotCrop3 :a)Triple-Shot Compositions is a task that takes an image as input and outputs three shots; b)Construct TSC dataset by generate annotations from both experts and MLLMs; c)a three-stage training pipeline consists of CoT-SFT, Semi-SFT and GRPO-S. 3.3 Three-stage Model Training As depicted in Fig. 2c), our training methodology comprises three sequential stages, each designed t… view at source ↗
Figure 3
Figure 3. Figure 3: Details of pseudo-label generation strategy and reward function. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on TSC with baseline models. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: In the pseudo-label generation strategy, the selection rates of different models (left) and the [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison between user score and MLLM score (left) and analyzation on spearman rank [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example 1 of qualitative comparison on TSC with baseline models. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Example 2 of qualitative comparison on TSC with baseline models. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example 3 of qualitative comparison on TSC with baseline models. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example 4 of qualitative comparison on TSC with baseline models. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Example 5 of qualitative comparison on TSC with baseline models. [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Example 6 of qualitative comparison on TSC with baseline models. [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Example 7 of qualitative comparison on TSC with baseline models. [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Example 8 of qualitative comparison on TSC with baseline models. [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: System prompt of our STC-Bench. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: MLLM scorer prompt for evaluating aesthetic quality and storytelling ability. [PITH_FULL_IMAGE:figures/full_fig_p022_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Answer with CoT of example 1 and example 2. [PITH_FULL_IMAGE:figures/full_fig_p023_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Answer with CoT of example 3 and example 4. [PITH_FULL_IMAGE:figures/full_fig_p024_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Answer with CoT of example 5 and example 6. [PITH_FULL_IMAGE:figures/full_fig_p025_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Answer with CoT of example 6 and example 8. [PITH_FULL_IMAGE:figures/full_fig_p026_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: An example of our datasets. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_21.png] view at source ↗
read the original abstract

Prior work on aesthetic composition typically produces a single aesthetically pleasing crop, overlooking the narrative value of composing multiple shots from one scene. In practice, multi-shot composition is critical for downstream creative workflows: commercial posters often require multiple crops with different emphases (e.g., context, subject, and emotion/product details) to present key story beats. Therefore, we propose \textbf{Triple-Shot Compositions (TSC)}, a composition task that generates a three-shot set -- establishing, medium, and close-up -- from a single human-centric image, each paired with a brief shot description to support visual narration. To learn TSC with limited expert annotations, we introduce \textbf{ShotCrop} which undergoes a three-stage training process: it first applies Chain-of-Thought supervised fine-tuning to establish basic reasoning and aesthetic shot-cropping skills, then performs semi-supervised fine-tuning with high-confidence pseudo labels to further enhance aesthetic capability, and is finally optimized with Group Relative Policy Optimization for \textbf{ShotCrop} (GRPO-S) using a composite reward tailored for it. Specifically, our pseudo-labeling strategy combines MLLM-based scoring, aesthetic assessment, and CLIP similarity to retain high-confidence training signals. In addition, we present TSC-Bench, a benchmark of 1.2k expert-annotated test cases. Notably, ShotCrop achieves an average improvement of \textbf{2.82} times over GPT-5 in shot localization accuracy.

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 paper introduces Triple-Shot Compositions (TSC), a task to generate three cinematic shots (establishing, medium, and close-up) with descriptions from a single human-centric image. It proposes the ShotCrop model trained via a three-stage process: Chain-of-Thought supervised fine-tuning, semi-supervised fine-tuning with high-confidence pseudo labels from MLLM scoring, aesthetic assessment, and CLIP similarity, and optimization with Group Relative Policy Optimization (GRPO-S). The authors present TSC-Bench, a 1.2k expert-annotated benchmark, and claim that ShotCrop achieves an average 2.82 times improvement over GPT-5 in shot localization accuracy.

Significance. If the performance claims hold after verification of the pseudo-labeling process, this work could contribute to advancing AI-assisted cinematic composition tools for creative industries by addressing multi-shot narrative needs rather than single crops. The introduction of a specialized benchmark is a positive step, but the significance is tempered by the lack of validation for the training signals.

major comments (2)
  1. [Abstract] Abstract: The claim of a 2.82 times improvement in shot localization accuracy over GPT-5 lacks a definition of the accuracy metric, details on the evaluation protocol, error bars, statistical significance tests, or dataset split information, making it impossible to assess the reliability of the central result.
  2. [Abstract (pseudo-labeling strategy)] Abstract (pseudo-labeling strategy): The semi-supervised stage relies on high-confidence pseudo labels from MLLM-based scoring, aesthetic assessment, and CLIP similarity, but no correlation statistics or ablation studies are provided to verify that these pseudo labels align with the expert annotations in TSC-Bench, which is critical for the validity of the reported performance gain.
minor comments (1)
  1. The paper would benefit from clearer notation for the reward function in GRPO-S and more details on the composite reward tailored for ShotCrop.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract and the pseudo-labeling validation. We address each major comment below and will revise the manuscript to improve clarity and provide additional supporting analyses.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim of a 2.82 times improvement in shot localization accuracy over GPT-5 lacks a definition of the accuracy metric, details on the evaluation protocol, error bars, statistical significance tests, or dataset split information, making it impossible to assess the reliability of the central result.

    Authors: We agree that the abstract would benefit from additional context on the central claim despite space constraints. Shot localization accuracy is defined as the fraction of predicted bounding boxes with IoU > 0.5 against expert annotations on TSC-Bench. The full evaluation protocol (including the 1.2k expert-annotated cases, train/test split details, error bars from repeated runs, and statistical significance testing) appears in Sections 4.1–4.2. We will revise the abstract to include a concise definition of the metric and a pointer to the evaluation section. revision: yes

  2. Referee: [Abstract (pseudo-labeling strategy)] Abstract (pseudo-labeling strategy): The semi-supervised stage relies on high-confidence pseudo labels from MLLM-based scoring, aesthetic assessment, and CLIP similarity, but no correlation statistics or ablation studies are provided to verify that these pseudo labels align with the expert annotations in TSC-Bench, which is critical for the validity of the reported performance gain.

    Authors: The manuscript does not currently report explicit correlation statistics or ablations between the pseudo-labeling signals and TSC-Bench expert annotations. TSC-Bench serves as a held-out test set, so we will add a new analysis in the revision: correlation coefficients computed on a held-out validation subset of the training data, plus an ablation study that removes each scoring component (MLLM, aesthetic, CLIP) in turn. These additions will appear in Section 3.2 to directly address the alignment concern. revision: yes

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only; no explicit free parameters, invented entities, or non-standard axioms are stated. The method relies on standard supervised fine-tuning, pseudo-labeling, and RL-style optimization whose validity hinges on unstated assumptions about label quality.

axioms (1)
  • domain assumption MLLM-based scoring combined with aesthetic assessment and CLIP similarity reliably identifies high-quality pseudo labels for shot cropping.
    Central to the semi-supervised fine-tuning stage described in the abstract.

pith-pipeline@v0.9.1-grok · 5835 in / 1036 out tokens · 28009 ms · 2026-06-28T02:29:43.503164+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

30 extracted references · 11 canonical work pages · 6 internal anchors

  1. [1]

    Qwen3-VL Technical Report

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    arXiv preprint arXiv:2512.21675 (2025)

    Shuo Cao, Jiayang Li, Xiaohui Li, Yuandong Pu, Kaiwen Zhu, Yuanting Gao, Siqi Luo, Yi Xin, Qi Qin, Yu Zhou, et al. Unipercept: Towards unified perceptual-level image understanding across aesthetics, quality, structure, and texture.arXiv preprint arXiv:2512.21675, 2025

    work page arXiv 2025

  3. [3]

    Shikra: Unleashing Multimodal LLM's Referential Dialogue Magic

    Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm’s referential dialogue magic.arXiv preprint arXiv:2306.15195, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    Yolo-world: Real- time open-vocabulary object detection

    Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real- time open-vocabulary object detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16901–16911, 2024

    2024

  5. [5]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with ad- vanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    Learning subject- aware cropping by outpainting professional photos

    James Hong, Lu Yuan, Michaël Gharbi, Matthew Fisher, and Kayvon Fatahalian. Learning subject- aware cropping by outpainting professional photos. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 2175–2183, 2024

    2024

  7. [7]

    Lisa: Reasoning segmentation via large language model

    Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9579–9589, 2024

    2024

  8. [8]

    Beyond image borders: Learning feature extrapolation for unbounded image composition

    Xiaoyu Liu, Ming Liu, Junyi Li, Shuai Liu, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Beyond image borders: Learning feature extrapolation for unbounded image composition. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 13023–13032, 2023

    2023

  9. [9]

    Groma: Localized visual tokenization for grounding multimodal large language models

    Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiaojuan Qi. Groma: Localized visual tokenization for grounding multimodal large language models. InEuropean Conference on Computer Vision, pages 417–435. Springer, 2024

    2024

  10. [10]

    Deepperception: Advancing r1-like cognitive visual perception in mllms for knowledge-intensive visual grounding.arXiv preprint arXiv:2503.12797, 2025

    Xinyu Ma, Ziyang Ding, Zhicong Luo, Chi Chen, Zonghao Guo, Derek F Wong, Xiaoyi Feng, and Maosong Sun. Deepperception: Advancing r1-like cognitive visual perception in mllms for knowledge-intensive visual grounding.arXiv preprint arXiv:2503.12797, 2025

    work page arXiv 2025

  11. [11]

    Image cropping under design constraints

    Takumi Nishiyasu, Wataru Shimoda, and Yoichi Sato. Image cropping under design constraints. In Proceedings of the 5th ACM International Conference on Multimedia in Asia, pages 1–7, 2023

    2023

  12. [12]

    Find beauty in the rare: Contrastive composition feature clustering for nontrivial cropping box regression

    Zhiyu Pan, Yinpeng Chen, Jiale Zhang, Hao Lu, Zhiguo Cao, and Weicai Zhong. Find beauty in the rare: Contrastive composition feature clustering for nontrivial cropping box regression. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2011–2019, 2023

    2011

  13. [13]

    Pseudo label fusion with uncertainty estimation for semi-supervised cropping box regression.IEEE Transactions on Multimedia, 26:8157–8171, 2024

    Zhiyu Pan, Jiahao Cui, Kewei Wang, Yizheng Wu, and Zhiguo Cao. Pseudo label fusion with uncertainty estimation for semi-supervised cropping box regression.IEEE Transactions on Multimedia, 26:8157–8171, 2024

    2024

  14. [14]

    Zoomer: Adaptive image focus optimization for black-box mllm

    Jiaxu Qian, Chendong Wang, Yifan Yang, Chaoyun Zhang, Huiqiang Jiang, Xufang Luo, Yu Kang, Qingwei Lin, Anlan Zhang, Shiqi Jiang, et al. Zoomer: Adaptive image focus optimization for black-box mllm. arXiv preprint arXiv:2505.00742, 2025

    work page arXiv 2025

  15. [15]

    arXiv preprint arXiv:2411.14347 (2024)

    Tianhe Ren, Yihao Chen, Qing Jiang, Zhaoyang Zeng, Yuda Xiong, Wenlong Liu, Zhengyu Ma, Junyi Shen, Yuan Gao, Xiaoke Jiang, et al. Dino-x: A unified vision model for open-world object detection and understanding.arXiv preprint arXiv:2411.14347, 2024

    work page arXiv 2024

  16. [16]

    Instructcrop: Teaching multimodal large language models to crop aesthetic images

    Xiangfei Sheng, Pangu Xie, Weidong Zou, Pengfei Chen, Tong Zhu, and Leida Li. Instructcrop: Teaching multimodal large language models to crop aesthetic images. InProceedings of the 33rd ACM International Conference on Multimedia, pages 6830–6839, 2025

    2025

  17. [17]

    Joint probability distribution regression for image cropping

    Tengfei Shi, Chenglizhao Chen, Yuanbo He, Wenfeng Song, and Aimin Hao. Joint probability distribution regression for image cropping. In2023 IEEE International Conference on Image Processing (ICIP), pages 990–994. IEEE, 2023. 10

    2023

  18. [18]

    OpenAI GPT-5 System Card

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaugh- lin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. Openai gpt-5 system card.arXiv preprint arXiv:2601.03267, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    Spatial-semantic collaborative cropping for user generated content

    Yukun Su, Yiwen Cao, Jingliang Deng, Fengyun Rao, and Qingyao Wu. Spatial-semantic collaborative cropping for user generated content. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4988–4997, 2024

    2024

  20. [20]

    Image cropping with spatial-aware feature and rank consistency

    Chao Wang, Li Niu, Bo Zhang, and Liqing Zhang. Image cropping with spatial-aware feature and rank consistency. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10052–10061, 2023

    2023

  21. [21]

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency.arXiv preprint arXiv:2508.18265, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    Good view hunting: Learning photo composition from dense view pairs

    Zijun Wei, Jianming Zhang, Xiaohui Shen, Zhe Lin, Radomir Mech, Minh Hoai, and Dimitris Samaras. Good view hunting: Learning photo composition from dense view pairs. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5437–5446, 2018

    2018

  23. [23]

    Aescrop: Aesthetic-driven cropping guided by composition

    Yen-Hong Wong and Lai-Kuan Wong. Aescrop: Aesthetic-driven cropping guided by composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3906–3913, 2025

    2025

  24. [24]

    Focusing on your subject: Deep subject-aware image composition recommendation networks.Computational Visual Media, 9(1):87–107, 2023

    Guo-Ye Yang, Wen-Yang Zhou, Yun Cai, Song-Hai Zhang, and Fang-Lue Zhang. Focusing on your subject: Deep subject-aware image composition recommendation networks.Computational Visual Media, 9(1):87–107, 2023

    2023

  25. [25]

    PhotoFramer: Multi-modal Image Composition Instruction

    Zhiyuan You, Ke Wang, He Zhang, Xin Cai, Jinjin Gu, Tianfan Xue, Chao Dong, and Zhoutong Zhang. Photoframer: Multi-modal image composition instruction.arXiv preprint arXiv:2512.00993, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [26]

    Aesthetic image cropping meets vlp: Enhancing good while reducing bad.Journal of Visual Communication and Image Representation, 105:104316, 2024

    Quan Yuan, Leida Li, and Pengfei Chen. Aesthetic image cropping meets vlp: Enhancing good while reducing bad.Journal of Visual Communication and Image Representation, 105:104316, 2024

    2024

  27. [27]

    Reliable and efficient image cropping: A grid anchor based approach

    Hui Zeng, Lida Li, Zisheng Cao, and Lei Zhang. Reliable and efficient image cropping: A grid anchor based approach. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5949–5957, 2019

    2019

  28. [28]

    Human-centric image cropping with partition-aware and content-preserving features

    Bo Zhang, Li Niu, Xing Zhao, and Liqing Zhang. Human-centric image cropping with partition-aware and content-preserving features. InEuropean Conference on Computer Vision, pages 181–197. Springer, 2022

    2022

  29. [29]

    Pro- crop: Learning aesthetic image cropping from professional compositions.arXiv preprint arXiv:2505.22490, 2025

    Ke Zhang, Tianyu Ding, Jiachen Jiang, Tianyi Chen, Ilya Zharkov, Vishal M Patel, and Luming Liang. Pro- crop: Learning aesthetic image cropping from professional compositions.arXiv preprint arXiv:2505.22490, 2025

    work page arXiv 2025

  30. [30]

    sequence of shots,

    Zhihang Zhong, Mingxi Cheng, Zhirong Wu, Yuhui Yuan, Yinqiang Zheng, Ji Li, Han Hu, Stephen Lin, Yoichi Sato, and Imari Sato. Clipcrop: conditioned cropping driven by vision-language model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 294–304, 2023. 11 A Qualitative Comparison Figures 9-16 present additional qualitativ...

    2023