acceptodds
Withdrawn from ICLR 2027

IECLA: INTERPRETABLE EMOTION CONTRIBUTION LEARNING FOR AFFECTIVE IMAGE EDITING

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Affective image editing (AIE) aims to evoke a target emotion while preserving the content and structure of a source image. Existing methods improve emotional expression or visual fidelity, but the target emotion alone does not reveal which source regions should be changed and which should remain intact, especially in complex scenes. We introduce Interpretable Emotion Contribution Learning for Affective Image Editing (IECLA), a framework that separates source-emotion attribution from target-aware editing. IECLA constructs signed object-level supervision through counterfactual interventions and trains a context-conditioned contribution predictor to localize evidence that supports or opposes the source emotion. At inference, an image-level evaluator estimates the source emotion directly from the input, while the predictor identifies editable and preservation evidence without repeated interventions. A target-aware editing agent then translates this evidence and the desired emotion into localized, content-aware operations for existing editors. Extensive evaluations show that IECLA predicts counterfactual contributions more faithfully on both in-domain and cross-dataset tests, while external VLM and independent human references support the quality of its localized evidence. Across six editing backbones, IECLA consistently strengthens target-emotion expression while preserving image content and structure, and user studies further show improved perceived target expression and naturalness.

Primary area:
generative models
Keywords:
affective image editing, visual emotion attribution, counterfactual learning, interpretable image editing, multimodal agents

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_xiM7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces IECLA, a method that aims to evoke a target emotion while preserving the content and the structure of the original image. IECLA separates source-emotion attribution from target-aware editing. At first, they extract the source emotion of the image for image-level supervision. Then, their Context-Conditioned Object Contribution Learning, built on the SigLIP2 vision backbone, integrates region-conditioned visual features, the inferred source emotion, and geometric metadata to recover the signed contribution target. Finally, the target-aware agent combines the signed regional evidence and the target emotion to edit the image. They used six editing backbones combined with their method. They claim structural and emotion expression improvements.

Critical strengths

Well-written, well-motivated, and sufficient novelty regarding the architecture.

The proposed method improves emotion and structural preservation in several metrics.

Extensive ablations makes the method's contribution easy to understand.

Sufficient comparisons with six backbones.

Qualitative as well as quantitative results.

The isolation of the contribution of each term of the total loss is essential for understanding the impact of each term.

Nice tables and visualizations make the results interpretable.

Critical weaknesses

Lack of a metric measuring the realism of the edited image.

Some images contain artifacts like fires in some places.

No clear bounds between changing the emotion of the image and changing central image characteristics. For example, objects have entirely changed for example from bat to lollipop. Colors have been changed significantly as well.

A metric that measures the realism of the image would be necessary.

Insufficient justification about the coefficients of the loss function. An ablation of different coefficients is missing.

Questions for authors

Where is the bound between changing significantly the image's colors or objects and altering the emotion of the image?

Could the authors provide realism metrics like MUSIQ or CLIPIQA, or compute distribution realism via FID / KID?

How sensitive is the loss to the coefficients of each term. Why do the authors gave these weights for each term Could they provide an ablation of different loss factors?

Reviewer_sCnqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses the task of affective image editing, where an image changed such that it evokes a target emotion while it has to preseve content and structure of the original image. The authors claim that current methods do not sufficiently addresswhich regions in an image should be changed and which should be left intact. In their method, IECLA, the problem is addressed using three main components: In a first step, the method identifies which parts of the image are responsible for the emotion it envokes. In a second step, a predictor is trained to predict these contributions. In a last step, the predictor output is fed in to a target-aware agent which proposes an executable, image-specific editing plan, than then is executed by a VLM to create the manipulated source image. The method is compared against multiple methods, backbones, and baselines.

My recommendation is to reject the paper. My main concerns are centered around the provided results and evaluation: While the provided quantitive results indicate a clear improvement towards competitors and baselines, the actual visual results reveal multiple bias issues that question the evaluation itself. Additionally, the claimed interpretability is missing discussion.

Critical strengths
  1. The provided quantitative results using the provided evaluation schemes, shows improvements over within and across baseline evaluations.
Critical weaknesses
  1. Interpretability: A major claim of the paper is to provide an interpretable emotional contribution. However, almost no effort has been done to provide any new insight in this regard.
    a) E.g. Fig. 3 provides emotion attribution maps, but completely misses which emotion this refers to, what scales mean etc. An additional analysis on what the model actually learns to attribute for different emotions is completely missing and hence remains non-interpretable. \

  2. Results: While all tables show impressive improvements over the provided baselines, I have concerns about the results and the provided evaluation.
    a) All IECLA visual results for the emotion "anger" (see Fig. 1, 3, 5, Suppl. Fig. 8, 9) show a clear bias to the color red, some items also contain content such as flames. Other emotions also seem to have a clear color bias. Further, the emotion "awe" seems to always have a sun like feature in each created image. This also shows in examples where there is no sky at all (see e.g. Fig. 8, line 7, awe example). Generally, it seems that the model learns to implement a few of these very simple dominant features for each of the emotions. I would challenge that such universal feature should be part of a well working method.
    b) Results are provided on a purely aggregated level. More detailed insights such as per emotion category evaluation is completely missing.

  3. Statements justified by non-related works: At several occasions, provided statements are supported by non-related works. Two examples:
    a) Line 76, 77: "Moreover, localization alone does not specify how to introduce the target emotion while preserving unrelated content" is supported by the paper of Yang et al. 2024: Cross-modal contextualized diffusion models for text-guided visual generation and editing. However, this paper does not make any statement on neither introducing target emotions nor does it have any statement about localization. So it is likely the wrong work cited.
    b) Line 133: "Affective editing additionally requires translating emotion into image-dependent changes". The second citations is the Wei et al., 2025 (Omniedit: Building image editing generalist models through specialist supervision), which is non-related to affective editing or emotions at all

Questions for authors
  • Q1: Can the authors provide a detailed interpretability analysis (see W1).
  • Q2: Can the authors provide a detailed analysis on the described biases in the results?
Reviewer_QVVyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes IECLA, an attribution-guided framework for affective image editing. It constructs signed object-level supervision through counterfactual interventions, trains a context-conditioned contribution predictor, and uses a target-aware agent to translate source-emotion evidence into localized editing instructions. The framework operates with existing image editors without retraining their parameters and is evaluated through attribution benchmarks, cross-dataset experiments, and user studies.

Critical strengths
  1. The paper is clearly organized, and the overall framework is easy to follow.

  2. Experiments across six editing backbones and five external datasets demonstrate improvements in target-emotion expression and structural preservation over the corresponding native editing pipelines.

  3. The component and loss-function ablations, external-reference evaluations, and teacher-replay comparison provide useful analyses of the proposed framework.

Critical weaknesses
  1. Equation (1) assigns every input image to one of eight emotion categories without adequately accounting for emotionally neutral or ambiguous inputs, yet subsequent attribution and editing both depend on this label, making it necessary to evaluate the method’s robustness to source-emotion uncertainty.

  2. Equation (2) defines negative contributions relative to a calibration baseline rather than directly establishing opposition to the source emotion, and although the paper provides human-reference validation, the semantic reliability of negative contributions remains insufficiently characterized.

  3. Although the agent receives the target emotion, the edit/preserve partition relies on source-emotion contributions, yet regions opposing the source emotion may still require modification to achieve the target, and the negative-protection ablation supports this strategy only on average without analysis across source–target emotion pairs.

  4. Each item in the Human Review reference is labeled by only one annotator, and negative-class precision and recall are not reported, leaving inter-annotator agreement and the reliability of negative-contribution predictions insufficiently assessed.

Questions for authors

Please refer to the Critical Weaknesses section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.