pith. sign in

arxiv: 2606.05778 · v2 · pith:YZ53GYZYnew · submitted 2026-06-04 · 💻 cs.CV

Beyond Absolute Scores: Relative Edit-induced Difference for Generalizable Image Aesthetic Assessment

Qifei Jia , Xintong Yao , Minghao Li , Yajie Chai , Qiming Lu , Baoyue Shen , Yasen Zhang , Runyu Shi
show 2 more authors
Ying Huang Yue Zhang
This is my paper

Pith reviewed 2026-06-30 11:09 UTC · model grok-4.3

classification 💻 cs.CV
keywords image aesthetic assessmentrelative learningimage editingaesthetic differencesgeneralizationranking consistencycontrollable editingCoT reasoning
0
0 comments X

The pith

Learning relative aesthetic differences from edited image pairs lets models capture comparison-based human reasoning instead of fitting absolute scores.

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

Traditional image aesthetic assessment regresses absolute mean opinion scores from single images. This paper argues that human perception depends on subconscious comparisons to implicit references, so absolute regression misses the causal factors behind aesthetic judgments and limits generalization. RED-Aes instead generates edited image pairs with controllable models, measures the induced aesthetic differences, and trains the model to predict those differences under a relative ranking consistency reward. The approach replaces score fitting with explicit learning of the visual factors that drive aesthetic change. If correct, models would transfer across image domains without needing new absolute-score annotations for each scenario.

Core claim

The paper claims that controllable image editing can simulate the comparison process in human aesthetic reasoning, and that training solely on the resulting relative differences (via a three-stage strategy and ranking consistency reward) produces models that learn generalizable aesthetic principles rather than dataset-specific score distributions.

What carries the argument

RED-Aes framework, which uses editing-based image pairs and relative ranking supervision to learn the visual factors that cause aesthetic changes.

Load-bearing premise

Controllable image editing models can generate pairs whose aesthetic differences accurately reflect the causal factors humans use when judging aesthetics.

What would settle it

A new benchmark of images whose aesthetic differences arise from factors that editing models cannot simulate, on which the relative method shows no advantage over absolute-score baselines.

Figures

Figures reproduced from arXiv: 2606.05778 by Baoyue Shen, Minghao Li, Qifei Jia, Qiming Lu, Runyu Shi, Xintong Yao, Yajie Chai, Yasen Zhang, Ying Huang, Yue Zhang.

Figure 1
Figure 1. Figure 1: Relative Difference Learning for Aesthetics. (a) Human aesthetic reasoning process: Illustration of how humans rely on implicit comparison to judge aesthetics. (b) Edit-based image pairs in RED-20k: An example of source-edit pairs where edit operations induce quantifiable aesthetic shifts. (c) CoT with causal reasoning: The model produces detailed textual explanations covering visual analysis and aesthetic… view at source ↗
Figure 2
Figure 2. Figure 2: We collect 500k images from photography websites spanning diverse qual [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The RED-Aes training pipeline. Stage 1 injects aesthetic knowledge through two SFT targets: PT-O1 compares source-edit pairs to predict aesthetic dif￾ferences with CoT reasoning, while PT-O2 predicts the aesthetic impact of editing instructions without observing the edited image. Stage 2 aligns the model to output absolute aesthetic scores using only 200 anchor samples from the source pool. Stage 3 optimiz… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of RED-Aes against baselines. Each column displays the input images with Ground Truth score, and the predicted scores from each model, along with the CoT reasoning. 4.8 Case Study As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of RED-Aes against baselines. Each column displays the input images with Ground Truth score, and the predicted scores from each model, along with the CoT reasoning. 4.8 Case Study As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: PT-O1 data example in RED-20k (1/4). Each sample shows a source im￾age A paired with an edited counterpart B produced by a controllable editing model. The editing instruction targets a specific aesthetic deficiency of the source. The anno￾tation includes the consensus aesthetic difference δ and a CoT reasoning trace jointly produced by three VLM judges. The model is trained to compare the pair, generate th… view at source ↗
Figure 5
Figure 5. Figure 5: PT-O1 data example in RED-20k (1/4). Each sample shows a source im￾age A paired with an edited counterpart B produced by a controllable editing model. The editing instruction targets a specific aesthetic deficiency of the source. The anno￾tation includes the consensus aesthetic difference δ and a CoT reasoning trace jointly produced by three VLM judges. The model is trained to compare the pair, generate th… view at source ↗
Figure 6
Figure 6. Figure 6: PT-O1 data example in RED-20k (2/4). Another source-edit pair il￾lustrating a different type of aesthetic intervention (e.g., composition rebalancing or background simplification). The CoT reasoning explicitly attributes the predicted score change to the specific visual modifications introduced by the edit [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: PT-O1 data example in RED-20k (2/4). Another source-edit pair il￾lustrating a different type of aesthetic intervention (e.g., composition rebalancing or background simplification). The CoT reasoning explicitly attributes the predicted score change to the specific visual modifications introduced by the edit [PITH_FULL_IMAGE:figures/full_fig_p031_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: PT-O1 data example in RED-20k (3/4). A source-edit pair where the editing model applies a lighting or color-harmony adjustment. The annotation captures the direction and magnitude of the resulting aesthetic shift, verified by unanimous consensus among all three VLM judges [PITH_FULL_IMAGE:figures/full_fig_p029_7.png] view at source ↗
Figure 7
Figure 7. Figure 7: PT-O1 data example in RED-20k (3/4). A source-edit pair where the editing model applies a lighting or color-harmony adjustment. The annotation captures the direction and magnitude of the resulting aesthetic shift, verified by unanimous consensus among all three VLM judges [PITH_FULL_IMAGE:figures/full_fig_p032_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: PT-O1 data example in RED-20k (4/4). A source-edit pair demonstrating a semantic or subject-level edit (e.g., object removal or subject emphasis). Only pairs where all three judges unanimously agree on the direction of aesthetic change are retained, ensuring high annotation quality [PITH_FULL_IMAGE:figures/full_fig_p030_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: PT-O1 data example in RED-20k (4/4). A source-edit pair demonstrating a semantic or subject-level edit (e.g., object removal or subject emphasis). Only pairs where all three judges unanimously agree on the direction of aesthetic change are retained, ensuring high annotation quality [PITH_FULL_IMAGE:figures/full_fig_p033_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: PT-O2 data example in RED-20k. Unlike PT-O1, only the source im￾age and the editing instruction are provided as input; the edited image is withheld. The model must predict the aesthetic difference δ that would result from applying the instruction, internalizing the causal relationship between visual modifications and aesthetic changes without direct visual evidence of the outcome [PITH_FULL_IMAGE:figures/… view at source ↗
Figure 9
Figure 9. Figure 9: PT-O2 data example in RED-20k. Unlike PT-O1, only the source im￾age and the editing instruction are provided as input; the edited image is withheld. The model must predict the aesthetic difference δ that would result from applying the instruction, internalizing the causal relationship between visual modifications and aesthetic changes without direct visual evidence of the outcome [PITH_FULL_IMAGE:figures/… view at source ↗
Figure 10
Figure 10. Figure 10: RL training group example in RED-20k. Each group G consists of one source image and up to seven edited variants generated by randomly selected editing models under diverse editing instructions (e.g., style transfer, object removal, color grading, lighting adjustment, background replacement). All variants pass the three￾judge consensus filter. During Stage 3 GRPO training, the Relative Ranking Consis￾tency… view at source ↗
Figure 10
Figure 10. Figure 10: RL training group example in RED-20k. Each group G consists of one source image and up to seven edited variants generated by randomly selected editing models under diverse editing instructions (e.g., style transfer, object removal, color grading, lighting adjustment, background replacement). All variants pass the three￾judge consensus filter. During Stage 3 GRPO training, the Relative Ranking Consis￾tency… view at source ↗
read the original abstract

Traditional Image Aesthetic Assessment (IAA) methods mainly rely on regressing absolute Mean Opinion Scores (MOS). However, such a paradigm overlooks the inherently dynamic nature of human aesthetic perception, which relies on subconscious comparison against implicit visual references. Consequently, the lack of causal reasoning regarding aesthetic differences prevents models from learning generalizable aesthetic principles, thus limiting their generalization across diverse scenarios. In this work, we rethink the IAA task and propose Relative Edit-induced Difference Aesthetic learning (RED-Aes), a novel framework that leverages controllable image editing models to simulate the human aesthetic reasoning process. Instead of fitting absolute score distributions, RED-Aes explicitly learns the visual factors that drive aesthetic changes. To support this paradigm, we construct the RED-20k dataset, which comprises editing-based image pairs, quantitative aesthetic differences, and Chain-of-Thought (CoT) reasoning. Furthermore, we introduce a three-stage training strategy guided by a relative ranking consistency reward, optimizing the model solely via relative supervision. Extensive experiments demonstrate that RED-Aes achieves state-of-the-art performance on multiple public benchmarks, exhibiting superior generalization capabilities.

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 / 2 minor

Summary. The manuscript proposes RED-Aes, a framework for image aesthetic assessment that replaces absolute MOS regression with relative learning. It uses controllable image editing models to generate image pairs, quantitative aesthetic differences, and Chain-of-Thought reasoning in the new RED-20k dataset, then applies a three-stage training strategy driven by a relative ranking consistency reward. The central claim is that this yields state-of-the-art performance and superior generalization across public benchmarks by learning the visual factors that drive aesthetic changes.

Significance. If the generated edits accurately simulate the causal factors in human aesthetic judgments, the shift to relative supervision and the accompanying dataset could meaningfully improve generalization in IAA. The work supplies a concrete alternative to absolute-score fitting and introduces an explicit relative reward, which are positive contributions if the core assumption is supported.

major comments (2)
  1. [Dataset construction / RED-20k] Dataset construction: the claim that editing-model outputs simulate human aesthetic reasoning is load-bearing, yet the manuscript provides no independent check (human correlation study, ablation on edit realism, or comparison against real reference-based judgments) that the quantitative differences and CoT align with human perceptual factors rather than model-specific artifacts.
  2. [Experiments] Experiments section: the SOTA and generalization claims are asserted without reported quantitative details on editing validation, baseline comparisons, error bars, or sensitivity to post-hoc editing choices, leaving the central empirical support without visible derivation.
minor comments (2)
  1. [Training strategy] Clarify the precise mathematical definition of the relative ranking consistency reward and how it is applied across the three training stages.
  2. [Method overview] Add explicit discussion of the dependence on pre-existing editing models as an external component and any associated limitations.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address the two major comments point-by-point below, acknowledging where additional evidence or clarification is warranted.

read point-by-point responses

  1. Referee: [Dataset construction / RED-20k] Dataset construction: the claim that editing-model outputs simulate human aesthetic reasoning is load-bearing, yet the manuscript provides no independent check (human correlation study, ablation on edit realism, or comparison against real reference-based judgments) that the quantitative differences and CoT align with human perceptual factors rather than model-specific artifacts.

    Authors: We agree that the absence of an independent human validation study for the RED-20k pairs constitutes a substantive gap. The current construction relies on the controllability properties of the chosen editing models and the generated CoT to approximate aesthetic factors, but this does not substitute for direct correlation with human judgments. In the revision we will add a human correlation study on a subset of the pairs together with an ablation comparing model-generated differences against real reference-based aesthetic judgments. revision: yes

  2. Referee: [Experiments] Experiments section: the SOTA and generalization claims are asserted without reported quantitative details on editing validation, baseline comparisons, error bars, or sensitivity to post-hoc editing choices, leaving the central empirical support without visible derivation.

    Authors: The full manuscript reports benchmark results on multiple public IAA datasets with comparisons to prior methods. However, we acknowledge that explicit quantitative validation of the editing pipeline, error bars across runs, and sensitivity analysis to editing hyperparameters are not presented at the level of detail requested. We will expand the experiments section to include these elements, including tables reporting editing validation metrics and sensitivity results. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper constructs an external RED-20k dataset from controllable editing models and applies a three-stage training process with relative ranking consistency reward. Claims of SOTA generalization rest on evaluation against independent public benchmarks rather than any internal fit or self-referential loop. No equations, definitions, or self-citations reduce the reported performance or the learned aesthetic factors to the training inputs by construction. The supervision signal and evaluation data remain distinct from the model outputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Based on abstract only; the central claim rests on the domain assumption that editing models produce causally meaningful aesthetic pairs and on standard ML training choices whose details are unspecified.

free parameters (1)
  • training hyperparameters and reward scaling
    Typical ML fitting choices required for the three-stage optimization but not enumerated.
axioms (1)
  • domain assumption Controllable image editing models generate pairs whose differences reflect true aesthetic causal factors
    Invoked as the basis for simulating human reasoning and constructing RED-20k.

pith-pipeline@v0.9.1-grok · 5751 in / 1185 out tokens · 33284 ms · 2026-06-30T11:09:02.974209+00:00 · methodology

Review history (2 revisions) →

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

85 extracted references · 19 canonical work pages · 12 internal anchors

  1. [1]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops

    Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 126–135 (2017)

    2017

  2. [2]

    Qwen3-VL Technical Report

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Qwen2.5-VL Technical Report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report. ArXivabs/2502.13923(2025),https: //api.semanticscholar.org/CorpusID:276449796

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    Astronomical Journal (ISSN 0004-6256), vol

    Beers, T.C., Flynn, K., Gebhardt, K.: Measures of location and scale for velocities in clusters of galaxies-a robust approach. Astronomical Journal (ISSN 0004-6256), vol. 100, July 1990, p. 32-46.100, 32–46 (1990)

    1990

  5. [5]

    In: Noise reduction in speech processing, pp

    Benesty, J., Chen, J., Huang, Y., Cohen, I.: Pearson correlation coefficient. In: Noise reduction in speech processing, pp. 1–4. Springer (2009)

    2009

  6. [6]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18392–18402 (2023)

    2023

  7. [7]

    arXiv preprint arXiv:2512.21675 (2025)

    Cao, S., Li, J., Li, X., Pu, Y., Zhu, K., Gao, Y., Luo, S., Xin, Y., Qin, Q., Zhou, Y., 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

  8. [8]

    arXiv preprint arXiv:2507.14533 (2025)

    Cao, S., Ma, N., Li, J., Li, X., Shao, L., Zhu, K., Zhou, Y., Pu, Y., Wu, J., Wang, J., et al.: Artimuse: Fine-grained image aesthetics assessment with joint scoring and expert-level understanding. arXiv preprint arXiv:2507.14533 (2025)

    work page arXiv 2025

  9. [9]

    In: European conference on computer vision

    Datta, R., Joshi, D., Li, J., Wang, J.Z.: Studying aesthetics in photographic images using a computational approach. In: European conference on computer vision. pp. 288–301. Springer (2006)

    2006

  10. [10]

    Gao, F., Lin, Y., Shi, J., Qiao, M., Wang, N.: Aesmamba: Universal image aesthetic assessmentwithstatespacemodels.In:Proceedingsofthe32ndACMInternational Conference on Multimedia. pp. 7444–7453 (2024)

    2024

  11. [11]

    https://blog.google/innovation- and- ai/products/nano- banana- pro/(11 2025), accessed: 2026-03-19

    Google: Nano Banana Pro: Gemini 3 Pro Image model from Google DeepMind. https://blog.google/innovation- and- ai/products/nano- banana- pro/(11 2025), accessed: 2026-03-19

    2025

  12. [12]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., Zhu, Q., Xu, R., Zhang, R., Ma, S., Bi, X., et al.: Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    Seed1.5-VL Technical Report

    Guo, D., Wu, F., Zhu, F., Leng, F., Shi, G., Chen, H., Fan, H., Wang, J., Jiang, J., Wang, J., et al.: Seed1. 5-vl technical report. arXiv preprint arXiv:2505.07062 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    In: IJCAI

    He, S., Zhang, Y., Xie, R., Jiang, D., Ming, A.: Rethinking image aesthetics as- sessment: Models, datasets and benchmarks. In: IJCAI. vol. 6, p. 22 (2022)

    2022

  15. [15]

    In: European Conference on Computer Vision

    Hosu, V., Conde, M.V., Agnolucci, L., Barman, N., Zadtootaghaj, S., Timofte, R., Sun, W., Zhang, W., Cao, Y., Cao, L., et al.: Aim 2024 challenge on uhd blind photo quality assessment. In: European Conference on Computer Vision. pp. 261–286. Springer (2024)

    2024

  16. [16]

    In: Proceedings of the 32nd ACM International Conference on Multi- media

    Huang, Y., Sheng, X., Yang, Z., Yuan, Q., Duan, Z., Chen, P., Li, L., Lin, W., Shi, G.: Aesexpert: Towards multi-modality foundation model for image aesthetics Relative Edit-Induced Difference for Image Aesthetic Assessment 17 perception. In: Proceedings of the 32nd ACM International Conference on Multi- media. pp. 5911–5920 (2024)

    2024

  17. [17]

    GPT-4o System Card

    Hurst, A., Lerer, A., Goucher, A.P., Perelman, A., Ramesh, A., Clark, A., Os- trow, A., Welihinda, A., Hayes, A., Radford, A., et al.: Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [18]

    Quality Press (1993)

    Iglewicz, B., Hoaglin, D.C.: Volume 16: how to detect and handle outliers. Quality Press (1993)

    1993

  19. [19]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5148–5157 (2021)

    2021

  20. [20]

    In: European conference on computer vision

    Kong, S., Shen, X., Lin, Z., Mech, R., Fowlkes, C.: Photo aesthetics ranking net- work with attributes and content adaptation. In: European conference on computer vision. pp. 662–679. Springer (2016)

    2016

  21. [21]

    Dataset available from https://github

    Krasin, I., Duerig, T., Alldrin, N., Ferrari, V., Abu-El-Haija, S., Kuznetsova, A., Rom, H., Uijlings, J., Popov, S., Veit, A., et al.: Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github. com/openimages2(3), 18 (2017)

    2017

  22. [22]

    In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Ku, M., Jiang, D., Wei, C., Yue, X., Chen, W.: Viescore: Towards explainable metrics for conditional image synthesis evaluation. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 12268–12290 (2024)

    2024

  23. [23]

    FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space

    Labs, B.F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dock- horn, T., English, J., English, Z., Esser, P., et al.: Flux. 1 kontext: Flow match- ing for in-context image generation and editing in latent space. arXiv preprint arXiv:2506.15742 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    British journal of psychology95(4), 489–508 (2004)

    Leder, H., Belke, B., Oeberst, A., Augustin, D.: A model of aesthetic appreciation and aesthetic judgments. British journal of psychology95(4), 489–508 (2004)

    2004

  25. [25]

    Journal of experimental social psychology49(4), 764–766 (2013)

    Leys, C., Ley, C., Klein, O., Bernard, P., Licata, L.: Detecting outliers: Do not use standard deviation around the mean, use absolute deviation around the median. Journal of experimental social psychology49(4), 764–766 (2013)

    2013

  26. [26]

    arXiv preprint arXiv:2503.22679 (2025)

    Li, W., Zhang, X., Zhao, S., Zhang, Y., Li, J., Zhang, L., Zhang, J.: Q-insight: Understanding image quality via visual reinforcement learning. arXiv preprint arXiv:2503.22679 (2025)

    work page arXiv 2025

  27. [27]

    arXiv preprint arXiv:2509.21871 (2025)

    Liu, B., Hu, Y., Jin, S., Dou, S., Shi, G., Shao, J., Gui, T., Huang, X.: Unlocking the essence of beauty: Advanced aesthetic reasoning with relative-absolute policy optimization. arXiv preprint arXiv:2509.21871 (2025)

    work page arXiv 2025

  28. [28]

    Advances in neural information processing systems36, 34892–34916 (2023)

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems36, 34892–34916 (2023)

    2023

  29. [29]

    In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Oct 2017)

    Liu, X., van de Weijer, J., Bagdanov, A.D.: Rankiqa: Learning from rankings for no-reference image quality assessment. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Oct 2017)

    2017

  30. [30]

    In: 2011 international con- ference on computer vision

    Marchesotti, L., Perronnin, F., Larlus, D., Csurka, G.: Assessing the aesthetic quality of photographs using generic image descriptors. In: 2011 international con- ference on computer vision. pp. 1784–1791. IEEE (2011)

    2011

  31. [31]

    In: 2012 IEEE conference on computer vision and pattern recognition

    Murray, N., Marchesotti, L., Perronnin, F.: Ava: A large-scale database for aes- thetic visual analysis. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 2408–2415. IEEE (2012)

    2012

  32. [32]

    Reber, R., Schwarz, N., Winkielman, P.: Processing fluency and aesthetic plea- sure: Is beauty in the perceiver’s processing experience? Personality and social psychology review8(4), 364–382 (2004) 18 Q. Jia, X. Yao et al

    2004

  33. [33]

    In: Proceedings of the IEEE international conference on computer vision

    Ren, J., Shen, X., Lin, Z., Mech, R., Foran, D.J.: Personalized image aesthetics. In: Proceedings of the IEEE international conference on computer vision. pp. 638–647 (2017)

    2017

  34. [34]

    Advances in neural information processing systems35, 25278–25294 (2022)

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large- scale dataset for training next generation image-text models. Advances in neural information processing systems35, 25278–25294 (2022)

    2022

  35. [35]

    Bmj349(2014)

    Sedgwick, P.: Spearman’s rank correlation coefficient. Bmj349(2014)

    2014

  36. [36]

    Seedream 4.0: Toward Next-generation Multimodal Image Generation

    Seedream, T., Chen, Y., Gao, Y., Gong, L., Guo, M., Guo, Q., Guo, Z., Hou, X., Huang, W., Huang, Y., et al.: Seedream 4.0: Toward next-generation multimodal image generation. arXiv preprint arXiv:2509.20427 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  37. [37]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  38. [38]

    OpenAI GPT-5 System Card

    Singh, A., Fry, A., Perelman, A., Tart, A., Ganesh, A., El-Kishky, A., McLaughlin, A., Low, A., Ostrow, A., Ananthram, A., et al.: Openai gpt-5 system card. arXiv preprint arXiv:2601.03267 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    IEEE transactions on image processing27(8), 3998–4011 (2018)

    Talebi, H., Milanfar, P.: Nima: Neural image assessment. IEEE transactions on image processing27(8), 3998–4011 (2018)

    2018

  40. [40]

    LongCat-Image Technical Report

    Team, M.L., Ma, H., Tan, H., Huang, J., Wu, J., He, J.Y., Gao, L., Xiao, S., Wei, X., Ma, X., et al.: Longcat-image technical report. arXiv preprint arXiv:2512.07584 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  41. [41]

    Unsplash: Unsplash lite dataset 1.3.0.https://unsplash.com/data(2020)

    2020

  42. [42]

    Qwen-Image Technical Report

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.m., Bai, S., Xu, X., Chen, Y., et al.: Qwen-image technical report. arXiv preprint arXiv:2508.02324 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  43. [43]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wu, H., Zhang, Z., Zhang, E., Chen, C., Liao, L., Wang, A., Xu, K., Li, C., Hou, J., Zhai, G., et al.: Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 25490–25500 (2024)

    2024

  44. [44]

    Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels

    Wu, H., Zhang, Z., Zhang, W., Chen, C., Liao, L., Li, C., Gao, Y., Wang, A., Zhang, E., Sun, W., et al.: Q-align: Teaching lmms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  45. [45]

    arXiv preprint arXiv:2505.14460 (2025)

    Wu, T., Zou, J., Liang, J., Zhang, L., Ma, K.: Visualquality-r1: Reasoning-induced image quality assessment via reinforcement learning to rank. arXiv preprint arXiv:2505.14460 (2025)

    work page arXiv 2025

  46. [46]

    Advances in Neural Information Processing Systems36, 15903–15935 (2023)

    Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagere- ward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems36, 15903–15935 (2023)

    2023

  47. [47]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yang, Y., Xu, L., Li, L., Qie, N., Li, Y., Zhang, P., Guo, Y.: Personalized im- age aesthetics assessment with rich attributes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19861–19869 (2022)

    2022

  48. [48]

    arXiv e-prints pp

    You, Z., Gu, J., Li, Z., Cai, X., Zhu, K., Dong, C., Xue, T.: Descriptive image quality assessment in the wild. arXiv e-prints pp. arXiv–2405 (2024)

    2024

  49. [49]

    Advances in Neural Information Pro- cessing Systems36, 31428–31449 (2023)

    Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Pro- cessing Systems36, 31428–31449 (2023)

    2023

  50. [50]

    Advances in Neural Information Processing Systems37, 3058–3093 (2024) Relative Edit-Induced Difference for Image Aesthetic Assessment 19

    Zhao,H., Ma,X.,Chen,L.,Si,S.,Wu,R.,An,K.,Yu,P.,Zhang,M.,Li,Q.,Chang, B.: Ultraedit: Instruction-based fine-grained image editing at scale. Advances in Neural Information Processing Systems37, 3058–3093 (2024) Relative Edit-Induced Difference for Image Aesthetic Assessment 19

    2024

  51. [51]

    arXiv preprint arXiv:2404.09619 (2024) A Multi-expert Voting Protocol We employ a multi-expert voting protocol to produce consensus aesthetic scores for the source poolO

    Zhou, Z., Wang, Q., Lin, B., Su, Y., Chen, R., Tao, X., Zheng, A., Yuan, L., Wan, P., Zhang, D.: Uniaa: A unified multi-modal image aesthetic assessment baseline and benchmark. arXiv preprint arXiv:2404.09619 (2024) A Multi-expert Voting Protocol We employ a multi-expert voting protocol to produce consensus aesthetic scores for the source poolO. GivenNexp...

    work page arXiv 2024

  52. [52]

    Analyze the image from an aesthetic perspective and identify the key factors that affect its visual quality. Focus on high−level aesthetic aspects such as: − Composition and visual balance − Subject prominence and depth relationships − Semantic clarity and narrative coherence − Mood, emotion, and visual tension Avoid limiting the analysis to low−level par...

  53. [53]

    Based on your analysis, assign an overall aesthetic score to the image. − Score range: 0 to 10 − Rounded to two decimal places − No explanation or text allowed Output only the numeric score inside the following tags: <answer1> X.XX </answer1> ==================== Part 3: Aesthetic−Oriented Image Editing Proposals (Outside Tags) ====================

  54. [54]

    Propose image editing strategies that directly address the aesthetic issues identified in Part 1. 22 Q. Jia, X. Yao et al. The goal is to improve the image by correcting or mitigating these specific deficiencies

  55. [55]

    When proposing editing strategies, follow these principles: − Prioritize structural, semantic, and perceptual edits over simple parameter tuning − Treat image editing as explicit action decisions (e.g., add, remove, replace, restructure, emphasize, suppress) − Encourage diversity and avoid redundancy − Propose new, reasonable editing operations if necessary

  56. [56]

    ==================== Part 4: Generate a Single Executable Editing Instruction ====================

    You may consider (but are not limited to) the following image editing action space: − Object−level editing: add, remove, replace, move, scale, rearrange objects − Attribute−level editing: modify color, material, shape, pose, or expression − Background and scene editing: background replacement, scene transformation, environment restructuring − Composition ...

  57. [57]

    score" : [...],

    Based on the editing strategies proposed in Part 3, generate ONE concise , executable image editing instruction. The instruction must: − Describe only the editing action − Contain no explanation, analysis, or evaluation − Be directly usable by an image editing model or system Output only this single−sentence editing instruction inside the following tags: ...

  58. [58]

    The instruction must describe an image EDIT operation (modifying an existing image, not generating a new image from scratch)

  59. [59]

    The instruction must clearly and substantially involve the given edit category

  60. [60]

    The edit can be free−form and creative, including adding or removing objects, changing layout or composition, or altering style and environment

  61. [61]

    Output rules: − Output ONLY the instruction text

    The edit must be feasible for modern image editing models. Output rules: − Output ONLY the instruction text. − No explanations. − No formatting. Training Stage Prompts.This section details the prompt templates designed forthequestionpartofthetrainingdatausedinourmethod.Thesetemplatesare employed across different stages to construct the Q&A pairs for model...

  62. [62]

    A strict consensus mechanism was adopted, requiringunanimous agreementamong all three judges to filter out ambiguous or noisy pairs

    as judges to determine the direction of aesthetic change (δ). A strict consensus mechanism was adopted, requiringunanimous agreementamong all three judges to filter out ambiguous or noisy pairs. Tab. 10 reports the pairwise agreement rates calculated on the candidate poolpriorto consensus filtering. The moderate Relative Edit-Induced Difference for Image ...

    work page arXiv

  63. [63]

    The overhead structure dominates the top half, which may feel heavy or obstructive

    Composition and framing: - Image A is framed from under a large structure (hangar or bridge), which creates a dramatic, slightly constrained perspective. The overhead structure dominates the top half, which may feel heavy or obstructive. - Image B removes the overhead structure, offering a cleaner, more open composition. The aircraft is centered and fully...

  64. [64]

    While dramatic, this creates harsh transitions and potentially lost detail in shadows

    Lighting and exposure: - Image A has strong contrast due to the shadowed foreground and bright sky. While dramatic, this creates harsh transitions and potentially lost detail in shadows. - Image B has more even lighting. The sky is brighter and more evenly exposed, with softer shadows on the aircraft. This enhances detail visibility and reduces harshness

  65. [65]

    However, Image B’s sky appears more vibrant and saturated without being oversaturated, enhancing the overall color harmony and visual richness

    Color quality: - Both images have similar color palettes (silver aircraft, blue sky, white clouds). However, Image B’s sky appears more vibrant and saturated without being oversaturated, enhancing the overall color harmony and visual richness

  66. [66]

    No loss of detail is observed

    Sharpness and clarity: - Both images are sharp, but Image B’s removal of the overhead structure eliminates potential visual clutter and allows the viewer’s eye to focus more directly on the aircraft. No loss of detail is observed

  67. [67]

    The subject is clearer and more immediately recognizable

    Subject emphasis and visual clarity: - Image B places the aircraft more prominently in the frame, with no competing overhead elements. The subject is clearer and more immediately recognizable

  68. [68]

    - Image B’s background is cleaner, with unobstructed sky and distant buildings that provide context without competing for attention

    Background cleanliness and distraction level: - Image A’s overhead structure is a significant distraction, pulling attention away from the aircraft and creating a visually busy upper frame. - Image B’s background is cleaner, with unobstructed sky and distant buildings that provide context without competing for attention

  69. [69]

    - Image B feels more open, airy, and celebratory of the aircraft’s design, enhancing its visual appeal and emotional resonance

    Emotional impact and visual appeal: - Image A feels more industrial and moody due to the shadowed framing. - Image B feels more open, airy, and celebratory of the aircraft’s design, enhancing its visual appeal and emotional resonance

  70. [70]

    The removal of the distracting overhead structure and the improved lighting balance contribute to a more cohesive and professional aesthetic

    Overall coherence and professional quality:- Image B presents a more polished, gallery-ready image. The removal of the distracting overhead structure and the improved lighting balance contribute to a more cohesive and professional aesthetic. Conclusion: Image B improves upon Image A in nearly every aesthetic dimension: composition, lighting, color vibranc...

  71. [71]

    No significant change

    Composition and framing: Both images are nearly identical in framing and subject placement. No significant change

  72. [72]

    The shadows are deeper, which adds drama but may slightly reduce detail in dark areas

    Lighting and exposure: Image B exhibits slightly higher contrast and more pronounced highlights on the subjects’ faces and clothing. The shadows are deeper, which adds drama but may slightly reduce detail in dark areas. This is a subtle enhancement in mood but a slight regression in tonal balance

  73. [73]

    However, the overall color temperature feels marginally cooler, which may reduce the warmth and intimacy conveyed

    Color quality: Image B has marginally more saturated colors — the purple shirt and red lipstick appear slightly richer. However, the overall color temperature feels marginally cooler, which may reduce the warmth and intimacy conveyed. Net effect is a slight improvement in vibrancy but a slight loss in tonal harmony

  74. [74]

    There is no noticeable increase in noise or artifacts

    Sharpness and clarity: Image B appears marginally sharper, especially in the hair strands and fabric textures. There is no noticeable increase in noise or artifacts. This is a clear improvement

  75. [75]

    The emotional connection between the couple feels more intense due to the heightened visual clarity

    Subject emphasis and visual clarity: The increased contrast and sharpness in Image B make the subjects “pop” more against the background. The emotional connection between the couple feels more intense due to the heightened visual clarity. This is a positive change

  76. [76]

    No change

    Background cleanliness and distraction level: The background (Ferris wheel) is equally clean and blurred in both images. No change

  77. [77]

    The subjects’ expressions and pose feel more impactful

    Emotional impact and visual appeal: The slight increase in contrast and sharpness in Image B enhances the romantic, cinematic mood. The subjects’ expressions and pose feel more impactful. This is a clear improvement in emotional resonance

  78. [78]

    finished

    Overall coherence and professional quality: Image B feels slightly more polished — the enhancements in sharpness and contrast contribute to a more “finished” look. There is no loss in coherence; if anything, it’s slightly more refined. Holistic assessment: While Image B has a minor regression in tonal harmony due to increased contrast and cooler temperatu...

  79. [79]

    The foreground cocktail, which is the focal point in Image A, loses texture and detail in Image B

    Sharpness and clarity: Image B is significantly blurrier than Image A. The foreground cocktail, which is the focal point in Image A, loses texture and detail in Image B. This is a major regression

  80. [80]

    In Image B, the blur reduces its visual dominance and makes the scene feel less intentional or polished

    Subject emphasis: In Image A, the cocktail is crisp and draws the eye immediately. In Image B, the blur reduces its visual dominance and makes the scene feel less intentional or polished

Showing first 80 references.