Pith. sign in

REVIEW 2 major objections 4 minor 44 references

Open-ended affordance reasoning — asking 'what can I edit?' — yields emotional image edits that humans prefer 88.1% of the time over fixed-taxonomy baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-07-31 23:31 UTC pith:XOCSFOZW

load-bearing objection Solid, candid paper, but the headline human-preference win doesn't yet isolate open-ended planning from the stronger editor. the 2 major comments →

arxiv 2607.23920 v1 pith:XOCSFOZW submitted 2026-07-27 cs.CV

What Can I Edit? Open-Ended Strategy Discovery and the Emotion Editability Landscape

Qing Li , Zeyu Dong , Yin Cui , Chuan Yan , Xiaojiang Peng This is my paper
classification cs.CV
keywords emotional image editingopen-ended strategy discoveryaffordance reasoningmulti-agent frameworkhuman evaluationemotion-conditional metric biaspreference-affinity landscapeconsistency-expressiveness tradeoff
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

EmoScope's central claim is that the space of emotionally meaningful edits for a given image is image-specific and far larger than any predefined factor taxonomy, so the right first question is 'what can I edit?' rather than 'how should I edit?'. The framework uses a multi-agent pipeline: a Planner reasons about the image's affordances under the target emotion to generate editing strategies from scratch; a Plan Verifier audits them across five dimensions; an Image Editor executes; and a Result Verifier checks emotional plausibility, with two refinement loops. In a large human pairwise evaluation across all eight Mikels emotion categories (4,693 valid responses), participants preferred EmoScope over EmoEdit and EmoEditor by 88.1% on average, and the advantage held across all demographic strata. The paper also shows that classifier-based metrics (Emo-A, Emo-S) have an emotion-conditional blind spot: they systematically undervalue non-stereotypical, context-grounded edits that humans prefer, with the disagreement ranging roughly from 9% (fear) to 81% (excitement) of human-loved edits being rejected by the classifier.

Core claim

EmoScope shows that emotional image editing succeeds by asking 'what can I edit?' — constructing, per image and target emotion, an editable space defined by consistency (at pixel, structural, identity, and narrative levels) and plausibility. The Planner generates strategies from emotion-conditioned reasoning about elements, e.g., seeing empty picture frames as a canvas and reusing that affordance for amusement, fear, or sadness. A hierarchy of anchors, variables, and context manages the consistency–expressiveness tradeoff. Humans preferred EmoScope over two baselines by 88.1% (4,693 responses). Classifier metrics (Emo-A, Emo-S) are stereotype matchers, with per-emotion correlation to human p

What carries the argument

The load-bearing mechanism is the affordance-discovery Planner, which generates editing strategies ab initio from emotion-conditioned reasoning about the image's elements, rather than retrieving from a library or instantiating templates. A semantic hierarchy of Anchors, Variables, and Context translates the consistency–plausibility constraints into editing boundaries. Two refinement loops (Plan-Refine and Edit-Refine) keep open-ended discovery grounded and allow fallback from ambitious to safer strategies. The plan itself, expressed at the affordance level, serves as an interactive surface for user refinement.

Load-bearing premise

The central result presumes that EmoScope's 88.1% human-preference advantage is caused by its open-ended affordance Planner rather than by its stronger generative editor relative to the baselines' own editors; the paper's Editor-only ablation tests a bare-prompt editor without the Planner, but does not test a fixed-taxonomy Planner running on the same editor with human judges.

What would settle it

Run EmoScope's full agentic pipeline but replace the open-ended Planner with one that selects from a fixed factor taxonomy, keeping the same editor and verification loops; then collect human pairwise preferences against the original EmoScope on the same image set. If the fixed-taxonomy version matches or nearly matches EmoScope's preference rate, the human-preference advantage is not due to open-ended discovery.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If open-ended affordance reasoning is the right strategy source, emotional image editing systems should replace fixed factor taxonomies with per-image affordance discovery, especially for emotions like excitement and disgust where stereotype-based edits fail.
  • Classifier-based metrics (Emo-A, Emo-S) cannot serve as standalone proxies for human judgment of emotional edits; per-emotion calibration is a prerequisite, and evaluators that reason about why an image evokes an emotion are needed.
  • The same affordance-level plan enables non-expert users to refine emotional expression at the plan level rather than the pixel level; a pilot study shows convergence in about 1.7 plan and 1.7 edit iterations.
  • The preference advantage varies systematically across content–emotion pairs, implying that adaptive intensity selection and stratified benchmarking are necessary rather than optional.

Where Pith is reading between the lines

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

  • The 'what can I edit?' principle may generalize beyond emotion to other open-ended image-editing tasks, where the bottleneck is often identifying what the image affords for the target change, not executing the change.
  • The emotion-conditional metric blind spot suggests that classifier-based evaluation in affective computing carries an implicit prior toward stereotypical visual cues; one testable extension is building a dataset of context-grounded edits and measuring classifier performance per emotion.
  • The preference-affinity landscape is relative (EmoScope vs baselines), so absolute per-image editability remains unmeasured; a Likert-scale editability survey could disentangle intrinsic difficulty from method advantage, a direction the authors themselves flag.
  • A decisive experiment would hold the editor constant and compare the open-ended Planner against a fixed-taxonomy planner under identical agentic scaffolding, directly isolating the contribution of open-ended discovery.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper introduces EmoScope, a multi-agent framework for emotion-driven image editing that reframes the task from 'how should I edit?' to 'what can I edit?'. EmoScope uses a VLM-based Planner to discover image-specific affordances, a semantic hierarchy of anchors/variables/context to navigate consistency-expressiveness tradeoffs, and dual plan/edit verification loops. The editor is Qwen-Image-Edit-2511. The main empirical claims are: (i) in a large 2AFC human study (4,693 valid responses, 1,824 pairwise questions), humans prefer EmoScope over EmoEdit and EmoEditor by 88.1% on average on emotional accuracy; (ii) EmoScope's editing strategy adapts to the target emotion rather than following a fixed template; (iii) classifier-based metrics (Emo-A, Emo-S) have emotion-conditional blind spots that undervalue non-stereotypical edits; and (iv) a content-emotion preference-affinity landscape shows systematic variation in the relative advantage. The paper also describes an interactive plan-level refinement pilot.

Significance. If the central claim is established, EmoScope provides a meaningful reframing of affective image manipulation, with a large, carefully analyzed human-preference dataset and a novel demonstration that classifier-based emotion metrics are unreliable in an emotion-dependent way. The paper is unusually transparent: it explicitly names the editor-confound limitation, provides clustered-bootstrap confidence intervals, reports demographic robustness, and separates evaluation splits clearly. The metric-blind-spot analysis (Sec. 4.4) is a valuable contribution in itself, with a 9x spread in Case A rates across emotions that should inform future AIM evaluation. The resource and protocol are strong enough that the paper deserves serious consideration even though the main causal claim needs additional support.

major comments (2)
  1. [§4.1, §4.2, §4.6 (Tab. 9)] The central causal claim that open-ended affordance planning drives the human-preference advantage is not isolated at the human level. In §4.2, EmoScope (Qwen-Image-Edit-2511) is compared head-to-head against EmoEdit and EmoEditor, each with their own trained diffusion editor. Section 4.1 acknowledges this confound but the only same-editor control, A5 in Tab. 9, compares raw Qwen on a bare prompt to full EmoScope. It does not test a fixed-taxonomy planner on Qwen. The A1 ablation ('w/o Open Discovery') does restrict the planner to EmoEdit's factor taxonomy, but it is evaluated only with automated metrics (Emo-S, CLIP-I, LPIPS), which §4.4 shows have emotion-conditional blind spots vis-à-vis human judgment. Therefore the 88.1% Q1 preference could be driven by Qwen's generative instruction-following rather than by open-ended strategy discovery. A same-editor human-preference comparison (fu
  2. [§4.6 (Tab. 9) and §4.4] The ablation interpretation relies on Emo-S as a valid within-run ranking criterion, with the assertion that the metric's blind spot is 'direction-symmetric' across configurations (Sec. 4.6). No evidence is provided for this symmetry. Given Finding 3 in Sec. 4.4—per-emotion correlation between Emo-S and human preference ranges from effectively null to r=+0.51—the ranking of ablated configurations could be emotion-dependent and metric-biased. The paper should either report the ablation split by target emotion, add a small human-rated ablation subset, or soften the causal reading of the A1→Full jump (0.224→0.328) as 'the most consequential design choice.'
minor comments (4)
  1. [Table 7] The baseline label 'CLV A' appears to be a typo or formatting artifact; please make it consistent with the referenced method name.
  2. [Fig. 2 caption] Typo: 'Open-Ended Plan Discorvery' should be 'Discovery'.
  3. [Sec. 4.3 (Q4 attribution)] The per-emotion attribution differences in Fig. 4 are presented as evidence of strategy adaptivity (H2), but no significance test or effect size is reported, and the marginal distributions in Table 6 are nearly identical. The authors' cautious wording is appropriate; a statistical test on the per-emotion conditional distributions would strengthen the claim.
  4. [Sec. 4.8] The claim that the plan-feedback feature is 'functionally required' is based on 10 participants, all of whom endorsed it. This is stated as preliminary, which is fine, but the wording 'confirmed' is too strong for an n=10 pilot without a control condition.

Circularity Check

0 steps flagged

No significant circularity: all quantitative claims rest on external human judgment and standard external classifiers; no fitted parameter, no self-citation chain, and no by-construction reduction found.

full rationale

The paper's load-bearing numbers are established by external measurement rather than by construction. The 88.1% Q1 human preference (Sec. 4.2), the per-emotion Case A rates and metric-human correlations (Sec. 4.4), and the 9x8 preference-affinity landscape (Sec. 4.5) are all measured against independent human raters and the external EmoSet classifier; none is derived from the authors' own definitions. The formal editable-space definition E(I,e)={delta | C>=tau_c and P>=tau_p} (Sec. 3.1) is never inverted to produce a prediction: C and P are not operationalized computationally, no parameter is fitted to the reported preference numbers, and the semantic hierarchy is presented as design, not as a derived result. The reference list contains no self-citations (verified: no author-name overlap between the paper's authors and any reference), so self-citation patterns 3-5 do not arise. The paper also explicitly flags its own evidentiary limits, which I weighed in the verdict: the editor confound between EmoScope (Qwen-Image-Edit-2511) and the baselines' own trained editors is acknowledged in Sec. 4.1 ('A fair head-to-head ... would also require unifying the generative backbone, which we leave to future work'); the ablation (Sec. 4.6, Tab. 9) attributing the largest gain to open-ended discovery is evaluated only with Emo-S, the very metric Sec. 4.4 shows is biased against non-stereotypical edits, defended via an unverified direction-symmetry assumption; and the A1 ablation lacks a matched-editor human-preference test. These are evidence-strength / external-validity concerns (is the planner, rather than Qwen's quality, driving the human preference?), not circular reductions: the Emo-S ordering is contingent data, the disclosed bias direction works against, not for, the open-discovery conclusion, and no 'prediction' equals its input by construction. The paper is similarly careful in Sec. 4.3, explicitly declining to claim raw strategy diversity from marginals ('marginal entropies of attribution categories are too coarse'), and in Sec. 4.5, labeling the landscape a relative measure. Under the specific circularity tests required here, the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

No fitted scientific constants and no new physical entities are introduced. The two hand-authored design parameters that matter most are the intensity setting (which controls the consistency threshold for every edit) and the 8×3 sub-style menu (which quietly bounds the 'open-ended' strategy space). The remaining inputs are domain assumptions about VLM reasoning ability, the representativeness of one emotion classifier, and the validity of 2AFC emotion-preference as a fidelity measure.

free parameters (2)
  • Editing intensity setting (Mild/Moderate/Strong) = Moderate (default)
    Intensity maps the abstract consistency threshold τc onto the semantic hierarchy (Mild→Layer 2 only, Moderate→Layers 2–3, Strong→all layers). All benchmark runs use Moderate, so the human-preference and ablation results are conditional on this hand-picked setting, which is never varied or optimized.
  • Hand-written 8×3 sub-style taxonomy = 24 emotional sub-targets (e.g., Slapstick/Satirical/Whimsical for amusement)
    Step 3 of the Planner 'selects' among three author-authored sub-styles per Mikels category. This is a hand-authored strategy menu that partially bounds the supposedly open-ended space, and it also shapes the Q4 attribution categories used as evidence of adaptivity.
axioms (5)
  • domain assumption The VLM backbone (Gemini-3-Flash) can perform reliable emotion-conditioned affordance reasoning and generate feasible, plausible edit plans.
    Sec. 3.3: the four-step Editing Potential Discovery (deconstruction, narrative evaluation, sub-style selection, self-debate) assumes VLM accuracy on visual-emotional reasoning; the paper's own Limitations states quality "will scale with future VLM improvements."
  • ad hoc to paper The Mikels 8-emotion categories plus the three hand-defined sub-styles per emotion cover the space of relevant emotional edits.
    Sec. 3.3 Step 3: 8×3=24 targets; no external justification is given for this exact sub-style menu, and it constrains the claimed open-ended discovery.
  • domain assumption Human 2AFC preference on the Q1 'Emotion Accuracy' question measures emotional fidelity rather than visual salience or edit dramaticity.
    Sec. 4.2: the primary outcome is human preference; a possible 'wow' bias (preferring more striking edits regardless of emotional accuracy) is not controlled by matching edit magnitude or by absolute Likert ratings.
  • domain assumption The EmoSet classifier is representative of classifier-based emotion metrics.
    Sec. 4.1 and Sec. 4.4: Emo-A and Emo-S are defined by a single pretrained 8-way EmoSet classifier; the paper's blind-spot generalization claim extends beyond this one classifier without evidence from additional models.
  • ad hoc to paper Consistency can be validly discretized into Anchors/Variables/Context with monotonic relaxation as intensity increases.
    Sec. 3.2: the semantic hierarchy is an author-invented discretization of the abstract consistency constraint; no derivation or empirical validation of the ladder (Mild/Moderate/Strong) is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 19282 in / 14283 out tokens · 139249 ms · 2026-07-31T23:31:02.885417+00:00 · methodology

0 comments
read the original abstract

Emotional image editing requires more than applying affective filters or modifying predefined visual factors: an effective edit must identify what a particular image can afford for a target emotion. Existing affective image manipulation methods, including recent agentic variants, largely operate within bounded strategy spaces based on predefined factor taxonomies, knowledge libraries, or conventional editing templates, and therefore often miss image-specific, context-grounded strategies. We introduce EmoScope, a multi-agent framework that reframes the task from "how should I edit?" to "what can I edit?" EmoScope first discovers an image-specific editable space through emotion-conditioned affordance reasoning, then uses a semantic hierarchy of anchors, variables, and context to balance content consistency and emotional expressiveness before executing and verifying the edit. Because its plans are expressed as image-specific affordances rather than retrieved templates, EmoScope also exposes the editing strategy as an interactive surface for user refinement at the plan level. In a large-scale human evaluation covering all eight Mikels emotion categories, with 4,693 valid responses across 1,824 pairwise questions, participants preferred EmoScope over two competitive baselines by 88.1% on average. Attribution analysis further shows that EmoScope selects target-emotion-adaptive strategies rather than applying a uniform template. The same affordance-level plan also supports lightweight user refinement in an interactive pilot. Finally, we show that classifier-based metrics exhibit emotion-conditional blind spots toward non-stereotypical, context-grounded edits, and present a relative content-emotion preference-affinity landscape showing that EmoScope's advantage varies systematically across image-emotion combinations.

Figures

Figures reproduced from arXiv: 2607.23920 by Chuan Yan, Qing Li, Xiaojiang Peng, Yin Cui, Zeyu Dong.

Figure 1
Figure 1. Figure 1: A paradigm shift in emotional image editing. Top: on Amusement, EmoEdit [37] produces image-agnostic stereotypes (scattered balls from a predefined factor taxonomy) while EmoEd￾itor [15] pursues the emotion so aggressively that the subject is lost. Bottom: EmoScope reframes the problem as open-ended dis￾covery in a per-image editable space: it identifies the two empty picture frames as themselves a canvas,… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of EmoScope. Given an input image and a target emotion, the Planner discovers the editable space through emotion [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The semantic hierarchy defines editing boundaries, not strategies; higher intensity unlocks deeper layers (Mild = Layer 2 only; Moderate = L2–3; Strong = full L1–3). Both rows tar￾get amusement. Row 1 (street scene, narrative-heavy affordance) stays Whimsical throughout: a balloon-carrying corgi (Mild) → hanging lanterns + chalk footprints + a small creature (Moderate) → warm storybook palette + giant bubb… view at source ↗
Figure 4
Figure 4. Figure 4: Q4 attribution distribution (%) for EmoScope’s winning [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: One Case A example (Excitement). Source, baseline’s edit (Emo-A ✓, classifier accepts), our edit (Emo-A ×, classifier rejects); human raters unanimously prefer ours (3/3). The base￾line instantiates a stereotype template (pasted action-sport figures); ours produces a context-grounded interpretation (an Eiffel Tower fireworks night) that the classifier fails to read. This case is repre￾sentative of the broa… view at source ↗
Figure 5
Figure 5. Figure 5: Per-emotion metric–human correlation: Pearson [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Content–emotion preference-affinity landscape ( [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Cross-source stability (Fear only). Three image regimes (curated photograph, in-the-wild snapshot, iconic paint￾ing); per source, EmoScope produces a context-grounded Fear edit. The examples compactly illustrate the broader cross-source pattern across the Mikels emotion set. regime, each edited toward Fear—the same agent, prompts, and hyperparameters across all three. Even at this single￾emotion glance the… view at source ↗

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

44 extracted references · 6 linked inside Pith

  1. [1]

    Artemis: Affective language for visual art

    Panos Achlioptas, Maks Ovsjanikov, Kilichbek Haydarov, Mohamed Elhoseiny, and Leonidas J Guibas. Artemis: Affective language for visual art. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11569–11579, 2021

    2021

  2. [2]

    AIM-Bench: Benchmarking and improv- ing affective image manipulation via fine-grained hierarchi- cal control.arXiv preprint arXiv:2604.10454, 2026

    Shi Chen, Xuecheng Wu, Heli Sun, Yunyun Shi, Xinyi Yin, Fengjian Xue, Jinheng Xie, Dingkang Yang, Hao Wang, Junxiao Xue, et al. AIM-Bench: Benchmarking and improv- ing affective image manipulation via fine-grained hierarchi- cal control.arXiv preprint arXiv:2604.10454, 2026

    Pith/arXiv arXiv 2026

  3. [3]

    Emoticrafter: Text-to-emotional-image generation based on valence-arousal model

    Shengqi Dang, Yi He, Long Ling, Ziqing Qian, Nanxuan Zhao, and Nan Cao. Emoticrafter: Text-to-emotional-image generation based on valence-arousal model. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion (ICCV), pages 15218–15228, 2025

    2025

  4. [4]

    An argument for basic emotions.Cognition & emotion, 6(3-4):169–200, 1992

    Paul Ekman. An argument for basic emotions.Cognition & emotion, 6(3-4):169–200, 1992

    1992

  5. [5]

    Language-driven artistic style transfer

    Tsu-Jui Fu, Xin Eric Wang, and William Yang Wang. Language-driven artistic style transfer. InEuropean Con- ference on Computer Vision, pages 717–734. Springer, 2022

    2022

  6. [6]

    Evaluating image editing with llms: A comprehensive benchmark and intermediate-layer probing approach.arXiv preprint arXiv:2603.19775, 2026

    Shiqi Gao, Zitong Xu, Kang Fu, Huiyu Duan, Xiongkuo Min, Guangtao Zhai, et al. Evaluating image editing with llms: A comprehensive benchmark and intermediate-layer probing approach.arXiv preprint arXiv:2603.19775, 2026

    arXiv 2026

  7. [7]

    Regressor-guided generative image editing balances user emotions to reduce time spent online.arXiv preprint arXiv:2501.12289, 2025

    Christoph Gebhardt, Robin Willardt, Seyedmorteza Sa- dat, Chih-Wei Ning, Andreas Brombach, Jie Song, Otmar Hilliges, and Christian Holz. Regressor-guided generative image editing balances user emotions to reduce time spent online.arXiv preprint arXiv:2501.12289, 2025

    arXiv 2025

  8. [8]

    GraPE: A generate-plan-edit framework for compositional t2i synthe- sis.arXiv preprint arXiv:2412.06089, 2024

    Ashish Goswami, Satyam Kumar Modi, Santhosh Rishi Deshineni, Harman Singh, Parag Singla, et al. GraPE: A generate-plan-edit framework for compositional t2i synthe- sis.arXiv preprint arXiv:2412.06089, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Emoscene: A dual-space dataset for controllable affective image generation.arXiv preprint arXiv:2604.00933, 2026

    Li He, Longtai Zhang, Wenqiang Zhang, Yan Wang, and Lizhe Qi. Emoscene: A dual-space dataset for controllable affective image generation.arXiv preprint arXiv:2604.00933, 2026

    arXiv 2026

  10. [10]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. InAdvances in Neural Information Processing Sys- tems. Curran Associates, Inc., 2017

    2017

  11. [11]

    Editing on the generative manifold: A theoretical and empirical study of general diffusion-based image editing trade-offs.arXiv preprint arXiv:2603.29736, 2026

    Yi Hu, Leying Yi, Emily Davis, and Finn Carter. Editing on the generative manifold: A theoretical and empirical study of general diffusion-based image editing trade-offs.arXiv preprint arXiv:2603.29736, 2026

    arXiv 2026

  12. [12]

    Emotion-Director: Bridging affective shortcut in emotion-oriented image generation.arXiv preprint arXiv:2512.19479, 2025

    Guoli Jia, Junyao Hu, Xinwei Long, Kai Tian, Kaiyan Zhang, KaiKai Zhao, Ning Ding, and Bowen Zhou. Emotion-Director: Bridging affective shortcut in emotion-oriented image generation.arXiv preprint arXiv:2512.19479, 2025

    arXiv 2025

  13. [13]

    EmoFeedback 2: Reinforcement of contin- uous emotional image generation via lvlm-based reward and textual feedback.arXiv preprint arXiv:2511.19982, 2025

    Jingyang Jia, Kai Shu, Gang Yang, Long Xing, Xun Chen, and Aiping Liu. EmoFeedback 2: Reinforcement of contin- uous emotional image generation via lvlm-based reward and textual feedback.arXiv preprint arXiv:2511.19982, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Early timestep zero-shot candidate selection for instruction-guided image editing

    Joowon Kim, Ziseok Lee, Donghyeon Cho, Sanghyun Jo, Yeonsung Jung, Kyungsu Kim, and Eunho Yang. Early timestep zero-shot candidate selection for instruction-guided image editing. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV), pages 18844–18854, 2025

    2025

  15. [15]

    Make me happier: Evoking emotions through im- age diffusion models

    Qing Lin, Jingfeng Zhang, Yew-Soon Ong, and Mengmi Zhang. Make me happier: Evoking emotions through im- age diffusion models. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pages 16367–16376, 2025

    2025

  16. [16]

    Emoagent: A multi-agent framework for diverse affective image manipulation.IEEE Transactions on Affec- tive Computing, pages 1–18, 2026

    Qi Mao, Haobo Hu, Yujie He, Difei Gao, Haokun Chen, and Libiao Jin. Emoagent: A multi-agent framework for diverse affective image manipulation.IEEE Transactions on Affec- tive Computing, pages 1–18, 2026

    2026

  17. [17]

    Sdedit: Guided image synthesis and editing with stochastic differential equa- tions.arXiv preprint arXiv:2108.01073, 2021

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia- jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equa- tions.arXiv preprint arXiv:2108.01073, 2021

    Pith/arXiv arXiv 2021

  18. [18]

    Emotional category data on images from the interna- tional affective picture system.Behavior research methods, 37(4):626–630, 2005

    Joseph A Mikels, Barbara L Fredrickson, Gregory R Larkin, Casey M Lindberg, Sam J Maglio, and Patricia A Reuter- Lorenz. Emotional category data on images from the interna- tional affective picture system.Behavior research methods, 37(4):626–630, 2005

    2005

  19. [19]

    Coeditor++: Instruction-based vi- sual editing via cognitive reasoning.arXiv preprint arXiv:2603.05518, 2026

    Minheng Ni, Yutao Fan, Zhengyuan Yang, Yeli Shen, Yuxiang Wei, Yaowen Zhang, Lijuan Wang, Lei Zhang, and Wangmeng Zuo. Coeditor++: Instruction-based vi- sual editing via cognitive reasoning.arXiv preprint arXiv:2603.05518, 2026

    arXiv 2026

  20. [20]

    A unified and inter- pretable emotion representation and expression generation

    Reni Paskaleva, Mykyta Holubakha, Andela Ilic, Saman Mo- tamed, Luc Van Gool, and Danda Paudel. A unified and inter- pretable emotion representation and expression generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2447–2456, 2024

    2024

  21. [21]

    Qwen-image technical report.https:// github.com/QwenLM/Qwen-Image-Edit, 2025

    QwenTeam. Qwen-image technical report.https:// github.com/QwenLM/Qwen-Image-Edit, 2025

    2025

  22. [22]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceedings of the 38th International Conference on Machine Learning, pages 8748–8763. PMLR, 2021

    2021

  23. [23]

    Learning multi-level deep representations for image emotion classification.Neu- ral processing letters, 51(3):2043–2061, 2020

    Tianrong Rao, Xiaoxu Li, and Min Xu. Learning multi-level deep representations for image emotion classification.Neu- ral processing letters, 51(3):2043–2061, 2020

    2043

  24. [24]

    A circumplex model of affect.Journal of personality and social psychology, 39(6):1161, 1980

    James A Russell. A circumplex model of affect.Journal of personality and social psychology, 39(6):1161, 1980

    1980

  25. [25]

    IMAGAgent: Orches- trating multi-turn image editing via constraint-aware plan- ning and reflection.arXiv preprint arXiv:2603.29602, 2026

    Fei Shen, Chengyu Xie, Lihong Wang, Zhanyi Zhang, Xin Jiang, Xiaoyu Du, and Jinhui Tang. IMAGAgent: Orches- trating multi-turn image editing via constraint-aware plan- ning and reflection.arXiv preprint arXiv:2603.29602, 2026

    arXiv 2026

  26. [26]

    Empowering visual creativity: A vision- language assistant to image editing recommendations.arXiv preprint arXiv:2406.00121, 2024

    Tiancheng Shen, Jun Hao Liew, Long Mai, Lu Qi, Jiashi Feng, and Jiaya Jia. Empowering visual creativity: A vision- language assistant to image editing recommendations.arXiv preprint arXiv:2406.00121, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing, 13(4): 600–612, 2004

    2004

  28. [28]

    Genartist: Multimodal llm as an agent for unified image gen- eration and editing

    Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal llm as an agent for unified image gen- eration and editing. InAdvances in Neural Information Pro- cessing Systems, pages 128374–128395. Curran Associates, Inc., 2024

    2024

  29. [29]

    Affective image filter: Re- flecting emotions from text to images

    Shuchen Weng, Peixuan Zhang, Zheng Chang, Xinlong Wang, Si Li, and Boxin Shi. Affective image filter: Re- flecting emotions from text to images. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10810–10819, 2023

    2023

  30. [30]

    Autogen: Enabling next-gen LLM applications via multi-agent conversation

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. Autogen: Enabling next-gen LLM applications via multi-agent conversation. InICLR 2024 Workshop on Large Language Model (LLM) Agents, 2024

    2024

  31. [31]

    MUSE: Manipulating unified framework for synthesizing emotions in images via test-time optimization

    Yingjie Xia, Xi Wang, Jinglei Shi, Vicky Kalogeiton, and Jian Yang. MUSE: Manipulating unified framework for synthesizing emotions in images via test-time optimization. arXiv preprint arXiv:2511.21051, 2025

    arXiv 2025

  32. [32]

    Beyond pix- els: Visual metaphor transfer via schema-driven agentic rea- soning.arXiv preprint arXiv:2602.01335, 2026

    Yu Xu, Yuxin Zhang, Juan Cao, Lin Gao, Chunyu Wang, Oliver Deussen, Tong-Yee Lee, and Fan Tang. Beyond pix- els: Visual metaphor transfer via schema-driven agentic rea- soning.arXiv preprint arXiv:2602.01335, 2026

    arXiv 2026

  33. [33]

    Fed-bench: A cross-granular bench- mark for disentangled evaluation of facial expression editing

    Fengjian Xue, Xuecheng Wu, Heli Sun, Yunyun Shi, Shi Chen, Liangyu Fu, Jinheng Xie, Dingkang Yang, Hao Wang, Junxiao Xue, et al. Fed-bench: A cross-granular bench- mark for disentangled evaluation of facial expression editing. arXiv preprint arXiv:2603.29697, 2026

    arXiv 2026

  34. [34]

    EmoSet: A large- scale visual emotion dataset with rich attributes

    Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischin- ski, Danny Cohen-Or, and Hui Huang. EmoSet: A large- scale visual emotion dataset with rich attributes. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 20383–20394, 2023

    2023

  35. [35]

    Emogen: Emo- tional image content generation with text-to-image diffusion models

    Jingyuan Yang, Jiawei Feng, and Hui Huang. Emogen: Emo- tional image content generation with text-to-image diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6358–6368, 2024

    2024

  36. [36]

    EmoStyle: Emotion-driven image stylization.arXiv preprint arXiv:2512.05478, 2025

    Jingyuan Yang, Zihuan Bai, and Hui Huang. EmoStyle: Emotion-driven image stylization.arXiv preprint arXiv:2512.05478, 2025

    arXiv 2025

  37. [37]

    Emoedit: Evoking emotions through image manipulation

    Jingyuan Yang, Jiawei Feng, Weibin Luo, Dani Lischin- ski, Daniel Cohen-Or, and Hui Huang. Emoedit: Evoking emotions through image manipulation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24690–24699, 2025

    2025

  38. [38]

    ReAct: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  39. [39]

    Dream- steerer: Enhancing source image conditioned editability us- ing personalized diffusion models

    Zhengyang Yu, Zhaoyuan Yang, and Jing Zhang. Dream- steerer: Enhancing source image conditioned editability us- ing personalized diffusion models. InAdvances in Neural In- formation Processing Systems, pages 120699–120734. Cur- ran Associates, Inc., 2024

    2024

  40. [40]

    CoEmoGen: Towards semantically-coherent and scalable emotional image content generation.arXiv preprint arXiv:2508.03535, 2025

    Kaishen Yuan, Yuting Zhang, Shang Gao, Yijie Zhu, Wenshuo Chen, and Yutao Yue. CoEmoGen: Towards semantically-coherent and scalable emotional image content generation.arXiv preprint arXiv:2508.03535, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    Mira: Multimodal iterative reasoning agent for image editing.arXiv preprint arXiv:2511.21087, 2025

    Ziyun Zeng, Hang Hua, and Jiebo Luo. Mira: Multimodal iterative reasoning agent for image editing.arXiv preprint arXiv:2511.21087, 2025

    arXiv 2025

  42. [42]

    EmoKGEdit: Training-free affective injection via visual cue transformation.arXiv preprint arXiv:2601.12326, 2026

    Jing Zhang and Bingjie Fan. EmoKGEdit: Training-free affective injection via visual cue transformation.arXiv preprint arXiv:2601.12326, 2026

    arXiv 2026

  43. [43]

    Affective image edit- ing: Shaping emotional factors via text descriptions.Int

    Peixuan Zhang, Shuchen Weng, Chengxuan Zhu, Binghao Tang, Zijian Jia, Si Li, and Boxin Shi. Affective image edit- ing: Shaping emotional factors via text descriptions.Int. J. Comput. Vis., 134(1):16, 2025

    2025

  44. [44]

    Efros, Eli Shecht- man, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2018

    2018