betaFeedback

WorksComputer Science

IECLA: Interpretable Emotion Contribution Learning for Affective Image Editing

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.

Introduction

Affective image editing (AIE) seeks to evoke a target emotion while preserving the source image’s identity and structure. This objective differs from an attribute-level instruction: an emotion word specifies a perceptual outcome, but does not indicate which visual evidence should change. The relevant cues may lie in a person’s expression, surrounding objects, or the broader scene context.

Modern image editors combine textual instructions with attention and spatial controls [29, 18, 5, 3, 57 a]. Such controls are effective when the desired attribute or region is explicit, but they leave the source-side decision implicit for affective editing [14, 8]. Existing affective editors expose a related tension: methods that produce stronger emotion shifts can introduce artifacts or structural drift, whereas more constrained methods preserve the source but make weak or misplaced changes when the relevant cues are unspecified [45, 49, 50]. The ambiguity is amplified in complex scenes, where affect is distributed across multiple objects and contextual relations rather than concentrated in a face or dominant subject.

Such distributed affect makes source-emotion evidence difficult to identify and use for editing. Visual emotion datasets and explanatory annotations characterize content associated with perceived affect [31, 22, 1, 44]. Complementary attribution methods expose model evidence through gradients, local surrogates, or perturbations [35, 37, 34, 32, 12, 16]. These resources support recognition and diagnosis, but they do not directly yield reusable, signed object-level contributions that distinguish evidence supporting the source emotion from evidence opposing it. Moreover, localization alone does not specify how to introduce the target emotion while preserving unrelated content [46, 48, 19]. However, effective AIE consistently requires two coupled decisions: identify and characterize the source evidence, then translate that evidence and the target emotion into an appropriate editing operation. Neither decision follows from the target label alone.

To address this need, we introduce IECLA, which learns signed, object-level source-emotion contributions from counterfactual interventions and exposes them as an editing interface. At inference, a frozen evaluator predicts the source emotion from the input, and the contribution predictor assigns regional scores without additional annotations. A target-aware agent combines these signals with the desired emotion to construct constrained instructions for existing editors.

In summary, our main contributions are summarized as follows:

  • We formulate signed object-level counterfactual contributions as a supervision signal for source-emotion attribution. The learned contributions distinguish editable source-emotion evidence from regions to preserve, without requiring a source-emotion label at inference.

  • We build a unified attribution-guided framework that connects the learned evidence to target-aware image editing. It produces interpretable emotion heatmaps and converts attributed evidence into guidance for image editors without retraining their parameters.

  • We validate IECLA through multiple benchmarks, experiments, and backbones. The results show more faithful attribution and stronger target-emotion expression while preserving source structure and naturalness.

Related Work

Visual emotion understanding and attribution. Visual affective computing models depict and elicit viewer’s emotions with categorical labels, response distributions, scene context, and textual explanations [31, 22, 1, 44]. Vision and language approaches further organize affective knowledge into interpretable rationales [26]. Complementary attribution methods expose model evidence through gradients [35, 20, 37], local surrogates [34], or perturbations [32, 12]. EmoCAM combines class-activation maps with detected objects for corpus-scale emotion analysis [10], and counterfactual explanations identify evidence whose alteration changes a prediction [16]. These resources support recognition and diagnosis, but their supervision or outputs are not defined as reusable, signed object-level controls for affective editing. Perturbation-based estimates require multiple evaluator queries, whereas visually plausible maps do not need to be strictly faithful [2, 54, 58]. IECLA builds on this line by learning intervention-grounded contributions with an explicit sign and an editor-facing interface.

Affective image editing and agentic control. General editors use partial noising [29], spatial and attention controls [3, 57, 18], or learned instruction following [5]. Affective editing additionally requires translating emotion into image-dependent changes [52, 40]. EmoEdit learns an emotion adapter from content-aware edit pairs [45], Moodifier derives contextual attributes through affect-specialized representations and multimodal reasoning [49], and MooD provides continuous valence–arousal control [50]. EmoAgent decomposes editing into planning, execution, and critique [27], while EmoKGEdit combines affective localization, knowledge-guided cue transformation, and structure-aware rendering [39] (Zhang & Fan 2026). These approaches advance target-side control. IECLA instead learns signed source-emotion contributions as an explicit interface to target-aware editing.

Method

Task and Framework Overview

Given a source image II and a target emotion tt, affective image editing seeks to evoke tt while preserving unrelated content. IECLA infers the source emotion yy from II, estimates how candidate regions support or oppose yy, and uses this signed evidence to guide an existing editor.

Figure 2 shows the complete pipeline. During training, counterfactual interventions construct region-level contributions, and a context-conditioned predictor learns to estimate them from the original image and masks. At inference, the predicted contributions form an emotion heatmap and separate editable from preservation evidence. A target-aware agent combines this evidence with tt, verifies feasible changes, and selects a budgeted editing plan. After structural review and instruction auditing, the compiled instruction is passed to the existing editor.

IECLA framework

Figure 2. IECLA framework. (a) Counterfactual interventions construct signed supervision. (b) Context-conditioned learning predicts regional contributions. (c) Signed predictions form emotion heatmaps and ground target-aware editing operations.

Counterfactual Contribution Label Construction

To obtain signed region-level supervision, we first infer the source emotion, then measure how individual regions affect its score. Image-level supervision identifies the perceived affect but does not reveal which regions produce it or how they contribute. We use evaluator EE, trained with EmoVerse [17] image-level supervision, to infer the source emotion from the full image:

y=arg⁡max⁡e∈C Ee(I),(1)y = \underset{e \in\mathcal{C}}{\arg\max}\ E_e(I), \tag*{(1)}

where E\mathcal{E} is the emotion vocabulary and Ee(I)E_e(I) is the probability assigned to emotion ee. At inference, the same frozen evaluator is evaluated only on the unmodified image to obtain yy; the user specifies the target emotion tt, and no source-emotion annotation is required.

To assess each candidate region relative to a comparable blur baseline, we compare its blur intervention against a matched reference-mask intervention. Grounding DINO Liu et al. [25] and SAM Kirillov et al. [21] provide candidate regions indexed by ii. Let Ii,blurI_{i,\mathrm{blur}} denote the intervention that blurs candidate region ii, and let Ii,maskI_{i,\mathrm{mask}} denote its matched reference-mask intervention. Their effects on the source-emotion response are defined as:

Δm=Ey(I)−Ey(Ii,blur),Δr=Ey(I)−Ey(Ii,mask),ci=Δm−E[Δr].(2)\Delta_m = E_y(I) - E_y(I_{i,\mathrm{blur}}), \quad\Delta_r = E_y(I) - E_y(I_{i,\mathrm{mask}}), \quad c_i = \Delta_m - \mathbb{E}[\Delta_r]. \tag*{(2)}

The correction estimates the candidate’s effect relative to the average response to comparable interventions. Here, Ey(⋅)E_y(\cdot) denotes the frozen evaluator response for the inferred source emotion yy, whereas E[Δr]\mathbb{E}[\Delta_r] averages the effects of area- and position-matched reference interventions. We retain Δm\Delta_m because averaging it would remove the region-specific signal that cic_i is intended to preserve. The resulting corrected contribution cic_i is the training target: positive and negative values denote effects above and below the reference baseline, respectively.

The resulting cic_i provides the direct supervision interface for the predictor in § 3.3. Repeating the intervention procedure over candidate regions yields one signed training instance per region.

Context-Conditioned Signed Contribution Prediction

To estimate cic_i without interventions, we train a context-conditioned predictor that operates on a single unmodified image. Built on the SigLIP2 vision backbone Tschannen et al. [38], the predictor integrates region-conditioned visual features, the inferred source emotion, and geometric metadata to recover the signed contribution target constructed in § 3.2.

We model the contextual dependence of regional affect by constructing three visual views at both global and local scales: the candidate region, its complementary context, and the complete image. At each scale s∈{g,ℓ}s \in\{g, \ell\}, the backbone encodes these views as risr_i^s, kisk_i^s, and asa^s, respectively. We fuse the three representations with candidate–context difference and product features:

vis=[ris;kis;as;ris−kis;ris⊙kis],hi=[vig;viℓ;ei;Emb⁡(y);ui;gi].(3)\begin{aligned} v_i^s &= [r_i^s; k_i^s; a^s; r_i^s - k_i^s; r_i^s \odot k_i^s],\\ h_i &= [v_i^g; v_i^\ell; e_i; \operatorname{Emb}(y); u_i; g_i]. \tag*{(3)} \end{aligned}

Here, eie_i encodes region semantics, uiu_i summarizes the evaluator state before intervention, and gig_i contains geometric and quality metadata. From the resulting representation hih_i, the prediction head estimates a nonnegative contribution magnitude and probabilities over negative, neutral, and positive signs, which are combined into a signed prediction:

c^i=mi(pi+−pi−),(4)\hat{c}_i = m_i\left(p_i^+ - p_i^-\right), \tag*{(4)}

where mi=softplus⁡(zi)m_i = \operatorname{softplus}(z_i), and pi+p_i^+ and pi−p_i^- are the positive and negative probabilities; the neutral class contributes zero. The predictor must recover not only the value of cic_i, but also its direction and the relative importance of regions within the same image. We therefore use:

L=LHuberw+0.25Labs+0.10Lsigned+0.20Lfocal+0.10Laux.(5)\mathcal{L} = \mathcal{L}_{\mathrm{Huber}}^w + 0.25\mathcal{L}_{\mathrm{abs}} + 0.10\mathcal{L}_{\mathrm{signed}} + 0.20\mathcal{L}_{\mathrm{focal}} + 0.10\mathcal{L}_{\mathrm{aux}}. \tag*{(5)}

The training objective jointly constrains contribution magnitude, within-image ranking, sign, and affective auxiliaries. The weighted Huber term learns contribution magnitude, the absolute-ranking term preserves regional importance, and the signed-ranking and focal terms supervise direction. The auxiliary term retains emotion-intensity and VAD information. Appendix J provides loss definitions and weighting details. Once trained, the predictor directly yields signed regional evidence without counterfactual interventions.

Attribution-Grounded Target Editing

Upon completing training, we apply the predictor to candidate regions identified by Region Discovery in the source image II. The resulting signed scores c^i\hat{c}_i are mapped onto their regions to form an interpretable attribution heatmap, showing which content supports or opposes the inferred source emotion yy. These predictions provide the attribution grounding for subsequent target editing.

The target-aware agent translates the signed regional evidence and target emotion tt into an executable, image-specific editing plan. It verifies the attribution against the visible content, identifies the primary emotional carrier, and selects the least disruptive operations. The primary local action is anchored to attributed source-emotion evidence, while optional atmosphere or minor-object actions are included only when the scene supports; negative evidence imposes preservation constraints.

Finally, a visual and structural critic checks the plan for executability and content preservation. A deterministic compiler assembles the approved operations into an editing prompt, and an integrated-instruction auditor checks it for conflicts before it is passed to the editor. Detailed routing and validation procedures are provided in Appendices § 17 and § 18.

Experiments

Experimental Setup

The locked attribution test contains 7,278 images and 53,180 SAM masks. The training, validation, and test subsets are strictly disjoint. Training uses three NVIDIA A800 GPUs. Cross-dataset evaluation excludes the external datasets’ original labels from training and teacher-emotion selection. Training details and the eight editing emotions are provided in Appendix § 18.

Main editing comparisons and ablations cover 680 source images paired with all eight target emotions, yielding 5,440 tasks across Qwen-Edit (Wu et al., [42]), InstructPix2Pix (Brooks et al., [5]), FLUX.2 [4], LongCat (Meituan LongCat Team et al., [28]), OmniGen2 (Wu et al., [43]), and HiDream (Cai et al., [6]); exact editor variants are listed in Appendix § 18. We compare each editor with its native pipeline using matched source images, targets, seeds, samplers, and resolutions. The teacher EE provides source-emotion inference and training supervision, while the contribution predictor and agent perform candidate scoring and editing planning. Target Success and Δp\Delta p measure teacher-aligned responses under EE, whereas Emo-A and Emo-S provide a second classifier view; none of these post-hoc scores is used for planning. Attribution ranking and localization metrics are computed per image and averaged; Sign Macro-F1 and continuous contribution errors are computed over all valid candidate regions. Confidence intervals use image-level cluster bootstrap.

Attribution Benchmark Results

To evaluate the accuracy of IECLA’s source-emotion attribution, we compare it with two baseline families: conventional heatmap attribution methods and signed attribution models built on

CLIP (Radford et al., [33]), DINOV3 (Siméoni et al., [36]), and SigLIP2 (Tschannen et al., [38]). The signed models are trained on the same counterfactual contribution targets, whereas the nonnegative heatmap methods are evaluated only on unsigned ranking metrics. Table 1 reports region ranking, contribution direction, and magnitude against the frozen evaluator’s counterfactual contributions; this benchmark measures teacher-defined fidelity rather than human causal attribution.

MethodImp. ρ\rho ↑Signed ρ\rho ↑Top-1↑NDCG@5↑Macro-F1↑MAE↓RMSE↓R2R^2 ↑
Heatmap attribution methods
Ablation-CAM (2020)-0.0452–0.22790.5877––––
Grad-CAM (2017)0.0328–0.25250.6313––––
Direct attention0.0432–0.23760.6321––––
Grad-CAM++ (2018)0.0690–0.25670.6443––––
RISE-256 (2018)0.0861–0.25210.6428––––
ADAPT-TF (2026)0.1214–0.27000.6592––––
EigenCAM (2020)0.1249–0.27510.6705––––
HiResCAM (2020)0.1986–0.29870.7051––––
LayerCAM (2021)0.2176–0.29910.7078––––
Vision-encoder contribution methods
CLIP (2021)0.35340.17150.43250.79180.32730.11450.20090.2036
DINOV3 (2025)0.35760.16920.43760.79400.33870.11160.20810.1458
SigLIP2 (2025)0.38190.23140.44450.80500.34540.10200.18330.3370
IECLA (ours)0.58850.50880.58080.89150.37030.09390.17430.4005

Table 1. Comparison with attribution baselines on the locked test set. IECLA achieves the strongest performance across ranking, sign agreement, and contribution regression.

Two patterns distinguish contribution learning from generic saliency. The heatmap methods recover some relevant regions, but their nonnegative outputs cannot represent evidence that opposes the source emotion. The learned vision-encoder models improve regional ranking, while IECLA further leads them in ranking, signed agreement, and regression. This joint improvement is important because the teacher scores concentrate near zero: a near-neutral predictor can attain a low MAE without recovering candidate order or direction. IECLA instead captures which regions matter, whether they support or oppose the source emotion, and their relative contribution strength, Figure 3 illustrates the corresponding localization in complex scenes.

Representative emotion attribution maps on diverse scenes

Figure 3. Representative emotion attribution maps on diverse scenes. Existing approaches mislocalize affective evidence or omit opposing regions. IECLA produces more accurate heatmaps.

Comparison with Native Editors

To assess transfer across editing backbones, Table 2 compares IECLA-guided editing with native emotion-request pipelines on the 5,440 matched tasks defined in § 4.1.

Affective outcomesTarget Success↑\uparrowEmo-A↑\uparrowEmo-S↑\uparrow
EditorNativeIECLA-guidedRel. Δ\DeltaNativeIECLA-guidedRel. Δ\DeltaNativeIECLA-guidedRel. Δ\Delta
Qwen-Edit (2025a)0.29960.3579+19.5%0.25880.3081+19.0%0.18430.2173+17.9%
InstructPix2Pix (2023)0.19060.2127+11.6%0.23290.2704+16.1%0.11030.1508+36.7%
FLUX.2 (2026)0.22830.2643+15.8%0.14650.1522+3.9%0.08670.1508+73.9%
LongCat (2025)0.27740.3099+11.7%0.18010.1844+2.4%0.09760.1053+7.9%
OmniGen2 (2025b)0.27520.3182+15.6%0.22850.2877+25.9%0.12290.1811+47.4%
HiDream (2025)0.26340.2947+11.8%0.22590.2461+8.9%0.07350.1130+53.7%
Structural preservationPSNR↑\uparrowSSIM↑\uparrowLPIPS↓\downarrow
EditorNativeIECLA-guidedRel. Δ\DeltaNativeIECLA-guidedRel. Δ\DeltaNativeIECLA-guidedRel. Δ\Delta
Qwen-Edit (2025a)15.498520.3046+31.0%0.57900.6137+6.0%0.36980.3440-7.0%
InstructPix2Pix (2023)23.464127.7156+18.1%0.60170.6682+11.1%0.38900.1840-52.7%
FLUX.2 (2026)30.847131.9981+3.7%0.41660.4346+4.3%0.41900.3249-22.5%
LongCat (2025)16.043616.6780+4.0%0.45550.4876+7.0%0.32240.2923-9.3%
OmniGen2 (2025b)14.319717.4542+21.9%0.48540.5966+22.9%0.28810.1820-36.8%
HiDream (2025)14.794019.9107+34.6%0.44350.5366+21.0%0.28930.1758-39.2%

Table 2. Comparison between native and IECLA-guided editing across six backbones. IECLA consistently improves target-emotion expression and structural preservation.

Native editors exhibit two complementary failure modes: conservative edits often preserve the source but under-express the requested emotion, whereas stronger changes can alter subject identity or surrounding structure. Table 2 shows that IECLA shifts this trade-off rather than merely increasing edit strength. Across all six editors, affective outcomes improve together with higher PSNR and SSIM and lower LPIPS. The common direction across editors with different native behavior supports attribution and target-aware planning as a transferable control interface, while the variation in gain magnitude reflects the execution capacity of each backbone. Appendix § 15 examines whether the gains persist across seeds and resolutions.

Figure 1, Figure 4, and Figure 5 illustrate the same trend in multi-object and context-dependent scenes: IECLA directs edits toward attributed cues while constraining unrelated content, avoiding the subject drift or scene-wide changes visible in several native outputs. Appendix § 9 provides further Qwen-Edit examples and analyzes the corresponding failure modes in Figure 8 and Figure 9.

Affective image editing with IECLA across diverse scenes and target emotions

Figure 1. Affective image editing with IECLA across diverse scenes and target emotions. Given only a target-emotion word, IECLA produces edits that effectively evoke the requested emotion while preserving the source image’s content and structure.

Eight-emotion editing with Qwen-Edit, with and without IECLA

Figure 4. Eight-emotion editing with Qwen-Edit, with and without IECLA. IECLA consistently elicits the requested emotion while preserving scene structure across all eight target categories.

Qualitative comparison across six editing backbones, with and without IECLA

Figure 5. Qualitative comparison across six editing backbones, with and without IECLA. IECLA improves target-emotion expression across diverse editors while preserving source structure.

Additional qualitative results with IECLA-guided Qwen-Edit

Figure 8. Additional qualitative results with IECLA-guided Qwen-Edit. Each row presents a source image followed by edits toward different target emotions, showing diverse affective changes with consistent source content and composition.

Additional comparisons between native Qwen-Edit and IECLA-guided editing, showing source, Qwen-Edit, and Qwen-Edit + IECLA outputs across emotions

Figure 9. Additional comparisons between native Qwen-Edit and IECLA-guided editing. IECLA strengthens target-emotion expression while preserving scene structure across diverse images.

Cross-Dataset Evaluation

To test transfer beyond the training distribution, we evaluate IECLA on 3,370 images from five external datasets: COCO [24], EmoSet [44], FI [51], Emotion6 [31], and EmoArt [53]. We then test a downstream editing task on Qwen-Edit, comparing its native pipeline with IECLA-guided editing.

Across the five external datasets, IECLA remains the leading attribution predictor (Appendix § 11) and improves downstream editing across natural scenes, affective photographs, and artwork (Table 3). All three preservation metrics improve on every dataset. EmoArt shows smaller structural gains, consistent with artistic and abstract images distributing affective cues across palette, texture, and composition rather than a localized region. Even in this domain, the positive gains show that stronger target-emotion expression does not sacrifice structural preservation.

Affective outcomesTarget Success ↑Emo-A ↑Emo-S ↑
DatasetNativeIECLA-guidedRel. ΔNativeIECLA-guidedRel. ΔNativeIECLA-guidedRel. Δ
COCO0.21880.2939+34.3%0.16870.2375+40.8%0.03940.1015+157.6%
EmoArt0.27570.3664+32.9%0.16270.2397+47.3%0.04250.1120+163.5%
EmoSet0.26420.3396+28.5%0.19490.2642+35.6%0.06850.1114+62.6%
Emotion60.15270.2611+71.0%0.15870.2168+36.6%0.03130.1065+240.3%
FI0.27040.3082+14.0%0.17610.2390+35.7%0.05930.0937+58.0%
Structural preservationPSNR ↑SSIM ↑LPIPS ↓
DatasetNativeIECLA-guidedRel. ΔNativeIECLA-guidedRel. ΔNativeIECLA-guidedRel. Δ
COCO12.852216.3243+27.0%0.27610.3918+41.9%0.49320.2427-50.8%
EmoArt14.489015.7233+8.5%0.34290.3659+6.7%0.23390.2219-5.1%
EmoSet12.742817.0118+33.5%0.32370.4595+42.0%0.42010.2181-48.1%
Emotion613.521817.4566+29.1%0.39440.5366+36.1%0.35340.1719-51.4%
FI14.261017.6349+23.7%0.40660.5093+25.3%0.33890.1675-50.6%

Table 3. Cross-dataset editing with Qwen-Edit. IECLA improves affective expression and structural preservation across all five datasets.

Ablation Study

Table 4 reports systematic ablations of each core component of the IECLA framework to assess its contribution, with qualitative examples in Figure 6. Appendix F provides additional ablation of editing controls and loss ablations, with the latter summarized in Figure 7(b).

ConditionSuccess ↑Δp ↑Emo-A ↑Emo-S ↑CLIP-I ↑G. LPIPS ↓G. SSIM ↑O. LPIPS ↓O. SSIM ↑O. MAE ↓
Qwen-Edit0.29960.27450.25880.18430.84920.36980.57900.41060.30490.1919
Qwen-Edit + IECLA (ours)0.35790.33440.30810.21730.85270.34400.61370.37670.32300.1733
-w/o attribution0.29150.27350.24540.17370.83940.35130.60350.38510.31840.1783
-w/o learning0.30460.28890.27240.20170.84120.34540.58770.41070.31470.1918
-w/o agent0.35400.32980.30240.21600.84000.35430.59900.38740.31710.1800

Table 4. Ablation of IECLA’s attribution, learning, and agent modules.

Qualitative ablation on Qwen-Image-Edit

Figure 6. Qualitative ablation on Qwen-Image-Edit. IECLA better expresses target emotions while preserving structure than the baseline and component-removed variants.

Figure 7. Attribution validation and loss-function ablation. (a) Metric values and normalized rank scores for representative attribution methods. (b) Ablation of the training objective, where IECLA provides the strongest overall affective and structural profile.

Qwen-Edit serves as the native baseline. The ablations follow the three modules in Figure 2: attribution in (a), contribution learning in (b), and target-aware execution in (c). IECLA achieves the strongest overall profile, leading affective metrics and structural measures. Removing attribution produces the largest affective degradation, confirming that source-emotion evidence is the primary signal for target-oriented editing. When the learning module is removed, the attribution output is passed directly from (a) to (c); the resulting decrease across all affective metrics shows that the learned contribution representation provides more useful editing evidence than direct transfer. Removing the agent while retaining attribution yields the strongest affective scores among the ablations, but still lowers every affective metric, indicating that the evidence must be converted into a target-aware plan to realize the full benefit of the framework.

The loss-level ablation further clarifies how the training objective maintains this balance. Focal-sign supervision has the largest affective impact, highlighting the importance of distinguishing supportive from opposing evidence. The ranking terms preserve regional importance and direction, while auxiliary supervision reinforces affective semantics. Although all ablations achieve slightly lower LPIPS, they degrade the other preservation measures and affective outcomes. Overall, the complete objective provides the best balance between target-emotion control and structural fidelity.

Attribution Validation

Attribution validation assesses whether our predicted regional contributions align with independent external annotations. We evaluate three external VLM references on the same 289 images and 1,032 aggregated SAM regions, with method outputs and aggregation fixed while varying only the reference labels. An independent Human Review provides a complementary perceptual evaluation that is excluded from training and model selection. Figure 7(a) visualizes the Human Review comparison, and Appendix § 12 reports the complete protocol and results.

The same profile appears under all three VLM references: IECLA leads in regional ordering, signed agreement, top-region selection, and localization despite varying competitors. This stability indicates that the result is not an isolated match to one external evaluator. Human Review provides a complementary perceptual test, where IECLA also aligns more closely with both the location and direction of affective evidence. In particular, agreement on direction supports negative contributions as meaningful opposing evidence rather than a by-product of the shared counterfactual teacher.

User Study

A blinded User Study examines whether attribution-guided planning produces perceptually stronger edits without sacrificing visual naturalness. Forty independent reviewers score 120 paired editing tasks across six editors for target expression and naturalness, while 20 additional heatmap tasks assess the perceptual quality of the underlying attribution maps.

Across all six editors, IECLA-guided outputs receive higher mean ratings for both target expression and naturalness. This consistent joint preference indicates that stronger affect does not come at the expense of perceptual naturalness. Together with the comparison in Table 2, the common direction across editors supports attribution-guided planning as a transferable editing signal, while differences in gain magnitude reflect how effectively each backbone realizes the localized guidance. The accompanying heatmap evaluation provides complementary perceptual evidence for the readability and focus of the underlying attribution. Overall, the User Study primarily supports the downstream editing utility of IECLA; the complete protocol and detailed results are reported in Appendix § 13.

Conclusion

We presented IECLA, a framework for affective image editing that separates source-emotion attribution from target-specific execution. Its counterfactual contribution predictor localizes signed object-level evidence: positive regions provide an editable focus, while negative regions provide preservation evidence. A target-aware agent then translates this evidence and the desired emotion into localized, content-aware editing operations with explicit preservation constraints. Experiments on the locked attribution benchmark and five external datasets, together with reference validation, show consistent attribution quality across evaluation sources and domains. Across six editing backbones, matched comparisons, ablations, qualitative results, and the User Study further demonstrate improved target-emotion expression with stronger structural preservation. These results establish affective attribution as both an explanation of perceived emotion and an actionable interface for controlling general image editors.

AI Use Statement

Generative AI tools supported the development of the conceptual framework, refinement of the experimental methodology, and implementation of experimental code. They were also used to synthesize and analyze related literature, organize references, and improve the clarity and readability of the manuscript. The authors reviewed the resulting text, code, and references and retain full responsibility for all claims, experimental results, and AI-assisted material in the final manuscript.

Ethics Statement

This work uses existing image datasets and pretrained models for affective image editing. Human annotation and the user study were conducted with participants’ informed consent; their protocols are described in Appendices § 12.§ 2 and § 13. Emotion labels describe perceived affect rather than a depicted individual’s internal state. The framework supports creative expression but shares the potential risks of generative image editing, including misleading portrayals and emotional manipulation. Responsible use should respect the privacy of depicted individuals and applicable dataset and model licenses, and disclose image manipulation where authenticity matters.

Reproducibility Statement

§ 3 describes the counterfactual supervision, contribution predictor, and target-aware editing framework, while § 4.1 summarizes the experimental setup. Appendix § 18 documents data provenance and disjoint splits, evaluation metrics, reference calibration, training objectives, baseline implementations, and agent and editor configurations. Appendix § 17 provides the structured editing-agent template. Appendices § 12 and § 13 describe the external-reference and human-evaluation protocols, while Appendices § 14 and § 15 provide additional ablations and stability analyses. Code will be released with the final version of the paper. Exact numerical results may vary with stochastic generation and differences in the software and model-serving environment.

Qualitative Results

Figure 8 complements the examples in Figure 1 with additional IECLA-guided Qwen-Edit outputs. The examples span people, animals, objects, and scene-centric images. Across these settings, the attributed guidance produces target-specific changes through localized semantic cues and scene appearance while retaining recognizable source content and composition.

Figure 9 further compares native Qwen-Edit with the same editor guided by IECLA. Across varied source content and target emotions, attribution-guided editing introduces clearer emotion-specific cues while better preserving the original subject and scene organization.

Taken together, the examples expose two recurring failure modes of native editing. When a source image already carries cues close to contentment, the native editor may introduce text or decorative objects even when a small chromatic or illumination adjustment would suffice. IECLA instead uses positive signed counterfactual contributions as editable evidence and negative contributions as preservation evidence, allowing the agent to make a restrained change while retaining the original composition. Several contentment examples follow this pattern: the guided outputs adjust color and light without adding unrelated content.

For targets such as fear, the native editor often darkens the background or adds red illumination as isolated operations. These cues do not necessarily form a coherent relation with the foreground and can leave the perceived emotion weak or ambiguous. IECLA can route evidence across the subject and its surrounding context, so expression, lighting, and background cues are changed jointly when supported by the image. The same pattern appears in face-dominant scenes, where editing only the expression misses contextual evidence, and in artistic images, where IECLA changes palette or atmosphere while preserving composition. These cases illustrate how signed attribution supplies relational editing evidence rather than a generic instruction to modify the most salient region.

Editing Baselines

This appendix compares IECLA with representative affective and general-purpose image editing methods on the same test set used for the main editing tables. The comparison includes global diffusion editing (SDEdit Meng et al., [29]), instruction-based editing (InsDiff, i.e., InstructDiffusion Geng et al., [15]), spatial control (ControlNet Zhang et al., [57]), subject-preserving editing (BLIP-Diffusion Li et al., [23]), and affective style transfer (CLVA Fu et al., [13] and AIF Weng et al., [41]), together with EmoEdit Yang et al., [45]. All methods are evaluated on the 680 source images and 5,440 source–target tasks used in the main editing comparison; the Qwen-Edit + IECLA row uses the same IECLA-guided Qwen-Edit configuration as the main tables. Since absolute scores depend on the editing backbone, this single-backbone comparison is reported here, while Table 2 presents matched native-versus-guided results across six editors.

Table 5 shows the tension between emotion control and structural preservation across prior approaches. Qwen-Edit + IECLA leads in five of six evaluations, indicating stronger target-emotion expression alongside close agreement with the source image. EmoEdit ranks second on LPIPS, while AIF leads in CLIP-I; the latter metric is also high for CLVA despite its low emotion scores. Thus, high image similarity alone does not imply effective emotion control, whereas IECLA performs well on both affective and preservation measures.

MethodPSNR↑\uparrowSSIM↑\uparrowLPIPS↓\downarrowCLIP-I↑\uparrowEmo-A↑\uparrowEmo-S↑\uparrow
SDEdit [29]14.15470.38200.38410.74810.20500.0929
InsDiff [15]9.86070.38700.39990.73880.05510.0537
ControlNet [57]12.26460.23700.48930.67680.18600.1659
BLIP-Diffusion [23]8.72420.32600.55140.81690.06760.0190
CLVA [13]12.19750.49600.39630.85720.03790.0138
AIF [41]16.13200.60800.41470.99320.08120.0345
EmoEdit [45]18.15640.60900.34590.84130.26470.2149
Qwen-Edit + IECLA (ours)20.30460.61370.34400.85270.30810.2173

Table 5. Comparison with affective and general-purpose editing baselines on the main editing test set. IECLA provides the strongest joint emotion improvement and structure retention.

Cross-Dataset Results

The five external datasets are examined separately under the attribution protocol summarized in Table 11. All tables use the same frozen counterfactual teacher and Grounding DINO and SAM candidate regions. Ranking and localization columns use image-macro aggregation, whereas Sign Macro-F1 and regression columns pool all valid candidate regions; images without valid regions are removed independently within each dataset. Bold and underlined entries mark the best and second-best results, while “–” denotes metrics undefined for nonnegative saliency outputs.

MethodImp. ρ\rho ↑Signed ρ\rho ↑Top-1↑NDCG@3↑Macro-F1↑MAE↓RMSE↓R2R^2 ↑
Heatmap attribution methods
Ablation-CAM (2020)-0.0137–0.28690.5909––––
Grad-CAM (2017)0.0212–0.29670.6214––––
Direct attention0.0304–0.27770.6198––––
Grad-CAM++ (2018)0.0655–0.30530.6326––––
RISE-256 (2018)0.1229–0.31360.6548––––
ADAPT-TF (2026)0.0279–0.28430.6050––––
EigenCAM (2020)0.1056–0.33150.6656––––
HiResCAM (2020)0.1654–0.33320.6869––––
LayerCAM (2021)0.2033–0.34450.6995––––
Vision-encoder contribution methods
CLIP (2021)0.30830.16590.47270.78730.47280.16370.25340.0987
DINOV3 (2025)0.31010.17730.45670.78580.50310.15360.25010.1219
SigLIP2 (2025)0.33510.24230.49580.80100.53310.13960.22110.3137
IECLA (ours)0.50060.48300.58430.87480.76000.08580.18060.5422

Table 11. Pooled cross-dataset attribution results under the shared counterfactual teacher. Ranking metrics are image-weighted, whereas Sign Macro-F1 and regression metrics are pooled over valid regions. IECLA leads all reported metrics.

MethodImagesMasksImp. ρ\rho ↑Signed ρ\rho ↑Top-1↑NDCG@3↑Macro-F1↑MAE↓RMSE↓R2R^2 ↑
Heatmap attribution methods
Ablation-CAM6125008-0.0632–0.17650.5035––––
Grad-CAM61250080.0740–0.23200.5814––––
Direct attention61250080.0507–0.21410.5541––––
Grad-CAM++61250080.0697–0.21900.5737––––
RISE-25661250080.1688–0.24670.5986––––
ADAPT-TF6125008-0.0061–0.19120.5125––––
EigenCAM61250080.1102–0.25330.5950––––
HiResCAM61250080.1840–0.25650.6413––––
LayerCAM61250080.2162–0.28430.6554––––
Vision-encoder contribution methods
CLIP61250080.35490.19950.44440.75300.46690.16170.24580.0102
DINOV361250080.38530.19770.42480.75880.49500.14590.23940.0610
SigLIP261250080.40380.30080.46410.78000.54420.13060.20780.2924
IECLA (ours)61250080.58330.55580.57350.86560.76740.07570.15450.6087

Table 6. Cross-dataset attribution results on COCO (612 images, 5,008 masks).

MethodImagesMasksImp. ρ\rho ↑Signed ρ\rho ↑Top-1↑NDCG@3↑Macro-F1↑MAE↓RMSE↓R2R^2 ↑
Heatmap attribution methods
Ablation-CAM9156304-0.0195–0.26340.5877––––
Grad-CAM91563040.0089–0.27100.6188––––
Direct attention91563040.0395–0.26670.6307––––
Grad-CAM++91563040.0572–0.28310.6345––––
RISE-25691563040.1582–0.29620.6701––––
ADAPT-TF91563040.0659–0.28420.6260––––
EigenCAM91563040.1197–0.32130.6761––––
HiResCAM91563040.1766–0.31260.6955––––
LayerCAM91563040.2472–0.33440.7296––––
Vision-encoder contribution methods
DINOV391563040.30880.21900.46670.79310.51310.15360.24400.1647
CLIP91563040.32140.21590.47100.80220.48560.16250.24820.1358
SigLIP291563040.34060.29200.51040.81550.55120.13940.21790.3340
IECLA (ours)91563040.48040.45190.55300.86360.71970.10340.20250.4247

Table 7. Cross-dataset attribution results on EmoSet (915 images, 6,304 masks).

MethodImagesMasksImp. ρ\rho ↑Signed ρ\rho ↑Top-1↑NDCG@3↑Macro-F1↑MAE↓RMSE↓R2R^2 ↑
Heatmap attribution methods
Ablation-CAM154499660.0103–0.33940.6174––––
Grad-CAM154499660.0023–0.33420.6281––––
Direct attention154499660.0418–0.31540.6378––––
Grad-CAM++154499660.0599–0.35490.6463––––
RISE-256154499660.0925–0.34720.6604––––
ADAPT-TF154499660.0182–0.31870.6253––––
EigenCAM154499660.0972–0.36590.6806––––
HiResCAM154499660.1519–0.37630.6963––––
LayerCAM154499660.1874–0.37890.6995––––
Vision-encoder contribution methods
DINOV3154499660.27330.13450.46440.78260.50170.15050.24860.1187
CLIP154499660.28750.11570.49030.78870.46620.15880.25140.0983
SigLIP2154499660.30020.18690.49610.79560.51930.13730.21910.3155
IECLA (ours)154499660.47700.47150.61010.88250.78840.07460.16750.5999

Table 8. Cross-dataset attribution results on FI (1,544 images, 9,966 masks).

MethodImagesMasksImp. ρ\rho ↑Signed ρ\rho ↑Top-1↑NDCG@3↑Macro-F1↑MAE↓RMSE↓R2R^2 ↑
Heatmap attribution methods
Ablation-CAM22613350.0104–0.35840.6734––––
Grad-CAM22613350.0531–0.34960.6925––––
Direct attention2261335-0.1231–0.27880.6421––––
Grad-CAM++22613350.1207–0.32740.6914––––
RISE-25622613350.0549–0.36730.6960––––
ADAPT-TF22613350.0520–0.34960.6353––––
EigenCAM22613350.1091–0.38940.7102––––
HiResCAM22613350.1617–0.34960.7149––––
LayerCAM22613350.1476–0.36730.7192––––
Vision-encoder contribution methods
CLIP22613350.25810.16170.47350.80590.45610.20050.30510.0892
DINOV322613350.34580.21800.47790.83660.47310.19500.31220.0463
SigLIP222613350.34880.22830.53980.83870.49820.17750.27640.2523
IECLA (ours)22613350.51880.48220.61060.90020.74000.11070.23240.4716

Table 9. Cross-dataset attribution results on Emotion6 (226 images, 1,335 masks).

MethodImagesMasksImp. ρ\rho ↑Signed ρ\rho ↑Top-1↑NDCG@3↑Macro-F1↑MAE↓RMSE↓R2R^2 ↑
Heatmap attribution methods
Ablation-CAM73533-0.1059–0.17810.5470––––
Grad-CAM735330.0319–0.20550.6264––––
Direct attention73533-0.0193–0.15070.5811––––
Grad-CAM++735330.0793–0.19180.6302––––
RISE-256735330.1492–0.21920.6854––––
ADAPT-TF73533-0.0326–0.13700.5921––––
EigenCAM735330.0586–0.20550.6684––––
HiResCAM735330.1681–0.27400.6730––––
LayerCAM735330.0541–0.17810.6315––––
Vision-encoder contribution methods
CLIP735330.35200.33010.35620.79940.48150.19460.27800.1517
DINOV3735330.36160.26250.36990.82840.52960.17980.27290.1827
SigLIP2735330.38580.34250.43840.79200.52530.17680.25970.2600
IECLA (ours)735330.50280.51030.43840.84880.70690.12080.21760.4801

Table 10. Cross-dataset attribution results on EmoArt (73 images, 533 masks).

The pooled table is not an equal-weight average of the five datasets: its ranking columns weight datasets by the number of valid images, while its sign and regression columns pool all valid regions.

Across all five datasets, IECLA maintains Importance and Signed Spearman correlations above 0.47 and 0.45, respectively, and ranks first on every signed, list-level, and regression metric. The pooled advantage in Table 11 is therefore not driven by a single source. It persists on FI, the largest external set, and on EmoArt, the smallest and most stylistically distinct set, where IECLA ties SigLIP2 on Top-1 but retains the lead in region ordering, signed agreement, and contribution regression.

The method families exhibit a consistent hierarchy across datasets. Vision-encoder contribution predictors generally improve the available ranking metrics over conventional heatmap attribution methods, indicating that semantic representations better match the teacher-defined regional evidence. IECLA further improves upon the strongest vision-encoder baseline in both signed criteria and contribution-error measures. This joint gain shows that its transfer extends beyond selecting the most relevant region to estimating whether each region supports or opposes the source emotion and by how much. Together, the results establish consistent agreement with the frozen counterfactual teacher across natural, affective, and artistic image distributions.

Reference Validation

This appendix collects the complete attribution validation associated with § 4.6. The three VLM references are evaluated on the same 289 images and the same 1,032 aggregated SAM regions. Method prediction maps, heatmaps, and aggregation rules remain fixed across references, and only the external contribution labels vary. No VLM output is used for training. The two Qwen references provide within-family checks, while GPT Terra supplies a reference from a different model family. Table 12–Table 14 report image-macro results for each reference, and Table 15 gives their unweighted mean before the independent Human Review comparison.

MethodImp. ρ\rho ↑\uparrowSigned ρ\rho ↑\uparrowTop-1↑\uparrowNDCG@3↑\uparrowPointing↑\uparrow
Heatmap attribution methods
Ablation-CAM-0.2294–0.40140.70970.2388
Grad-CAM-0.0161–0.47750.77310.2768
Direct attention0.0315–0.48100.77900.2111
Grad-CAM++0.0843–0.53630.79330.2630
EigenCAM0.0538–0.51560.78620.2699
HiResCAM0.1158–0.55710.79290.2837
RISE-2560.0748–0.50170.77380.4394
LayerCAM0.1835–0.56750.81440.2699
ADAPT-TF0.1890–0.56060.80810.2941
Vision-encoder contribution methods
CLIP0.11070.04070.57440.79360.6678
SigLIP20.15270.12800.61590.80830.6851
DINOV30.15070.16880.61250.80830.6782
IECLA (ours)0.61870.48680.87200.93310.8685

Table 12. Attribution agreement with the Qwen Chat reference. IECLA leads all five metrics.

MethodImp. ρ\rho ↑\uparrowSigned ρ\rho ↑\uparrowTop-1↑\uparrowNDCG@3↑\uparrowPointing↑\uparrow
Heatmap attribution methods
Ablation-CAM-0.1633–0.40480.72310.2388
Grad-CAM0.0698–0.51210.79590.2803
Direct attention0.0077–0.42210.76640.2215
Grad-CAM++0.0863–0.51210.79710.2803
EigenCAM0.0460–0.47060.78510.2803
HiResCAM0.1044–0.53630.79600.2941
RISE-256-0.0277–0.42560.75240.4394
LayerCAM0.1517–0.52550.81190.2837
ADAPT-TF0.0418–0.48440.78320.2699
Vision-encoder contribution methods
CLIP0.18500.02310.58820.81210.6609
SigLIP20.21380.07050.60550.82750.6644
DINOV30.22520.06260.61590.82610.6574
IECLA (ours)0.61650.53240.87890.93230.8547

Table 14. GPT Terra reference results. IECLA retains its lead under a different VLM family.

MethodImp. ρ\rho ↑\uparrowSigned ρ\rho ↑\uparrowTop-1↑\uparrowNDCG@3↑\uparrowPointing↑\uparrow
Heatmap attribution methods
Ablation-CAM-0.2191–0.39450.70510.2342
Grad-CAM0.0105–0.49360.77560.2734
Direct attention0.0244–0.46250.77250.2111
Grad-CAM++0.0898–0.53170.78830.2641
EigenCAM0.0397–0.49710.78080.2676
HiResCAM0.1299–0.55940.79440.2826
RISE-2560.0602–0.48330.76650.4325
LayerCAM0.1900–0.55710.81400.2664
ADAPT-TF0.1533–0.54100.79780.2849
Vision-encoder contribution methods
CLIP0.13480.03800.57550.79610.6609
SigLIP20.16020.11180.60550.80770.6690
DINOV30.16410.13520.60550.80580.6643
IECLA (ours)0.61170.51000.87660.93220.8616

Table 15. Mean agreement across VLM references. IECLA leads every metric.

VLM Reference

Qwen Chat and Qwen Reasoner both use Qwen3.8 through the official Qwen API, under direct-response and reasoning configurations, respectively. GPT Terra uses GPT-5.6-Terra. These models serve only as external attribution references and do not participate in contribution-predictor training, checkpoint selection, or editing-agent planning.

MethodImp. ρ\rho ↑\uparrowSigned ρ\rho ↑\uparrowTop-1↑\uparrowNDCG@3↑\uparrowPointing↑\uparrow
Heatmap attribution methods
Ablation-CAM-0.2647–0.37720.68260.2249
Grad-CAM-0.0222–0.49130.75780.2630
Direct attention0.0340–0.48440.77220.2007
Grad-CAM++0.0988–0.54670.77460.2491
EigenCAM0.0193–0.50520.77120.2526
HiResCAM0.1695–0.58480.79420.2699
RISE-2560.1335–0.52250.77340.4187
LayerCAM0.2348–0.58130.81580.2457
ADAPT-TF0.2291–0.57790.80210.2907
Vision-encoder contribution methods
CLIP0.10880.05010.56400.78270.6540
SigLIP20.11420.13700.59520.78730.6574
DINOV30.11630.17410.58820.78310.6574
IECLA (ours)0.59980.51080.87890.93130.8616

Table 13. Qwen Reasoner reference results. IECLA retains its lead across all five metrics.

The reference-specific results expose a stability pattern that is hidden by the average alone. Across Qwen Chat, Qwen Reasoner, and GPT Terra, IECLA’s Top-1, NDCG@3, and Pointing scores vary by at most 0.014, although the strongest competing method changes with the metric and reference. Importance and Signed Spearman show larger reference-dependent shifts, yet IECLA remains ahead of the competing methods. The result is therefore supported by agreement across references rather than an isolated match to one evaluator.

The baseline profiles also separate localization from signed attribution. Heatmap methods can obtain competitive NDCG@3 or Pointing on individual references, but their nonnegative outputs cannot represent opposing evidence. Vision-encoder contribution predictors add signed estimates and improve region localization, while their signed agreement remains substantially below IECLA. The simultaneous gain in ranking, sign, and Pointing indicates that the contribution predictor aligns these complementary properties instead of optimizing a single localization criterion.

Human Reference

Human Review starts from 320 emotion-balanced images; fixed usability filtering and human-region merging define the reported evaluation set. Forty annotators complete disjoint batches, with each item assigned to one annotator. All methods use the same SAM candidate masks and region mapping. We compute the ranking and localization metrics in Table 16 per image and average over images. Signed metrics are reported only for methods that produce signed contributions.

MethodImp. ρ\rho ↑Signed ρ\rho ↑Top-1 ↑NDCG@3 ↑Pointing ↑
Heatmap attribution methods
Ablation-CAM-0.2207–0.55020.78090.3218
Grad-CAM-0.0450–0.59520.81590.3426
Direct attention-0.0294–0.57440.82140.2595
Grad-CAM++0.0115–0.63320.82720.3391
EigenCAM0.0491–0.62980.83210.3495
HiResCAM0.0707–0.66780.84740.3495
RISE-2560.1264–0.65740.83970.4706
LayerCAM0.1526–0.68860.85720.3460
ADAPT-TF0.1942–0.67130.84640.3529
Vision-encoder contribution methods
CLIP0.07260.05470.68510.84920.7543
DINOV30.13730.11640.70930.86620.7716
SigLIP20.14120.09210.70930.86190.7509
IECLA (ours)0.50010.45250.86850.93150.8478

Table 16. Attribution agreement with independent Human Review annotations. IECLA leads in region ranking, signed agreement, and localization.

Comparing Table 16 with the VLM average in Table 15 reveals which properties transfer most closely to human judgments. IECLA’s NDCG@3 is nearly unchanged, while Top-1 and Pointing decrease only modestly; the larger reductions occur in Importance and Signed Spearman correlation. Thus, the leading region and top-weighted ordering are more stable across reference types than the full ordering of regional contributions. IECLA remains first throughout, whereas the strongest alternative changes between ADAPT-TF, DINOV3, and SigLIP2 depending on the criterion.

Signed contribution classification. Signed direction is evaluated independently on the same Human Review annotations. The protocol starts from 320 images and 1,319 annotated regions; filtering leaves 289 images and 1,032 human regions. When a human region overlaps several SAM masks, source-mask scores are aggregated with source-mask-area weights. The Sign Macro-F1 values in Table 17 use image-macro aggregation with 10,000 image-level bootstrap resamples, and signed methods use the fixed threshold τ=0.1\tau= 0.1. Nonnegative attention and saliency methods are omitted because they do not produce signed outputs.

MethodSign Macro-F1↑\uparrow
DINOV30.1530
CLIP0.1711
SigLIP20.1765
IECLA (ours)0.3601

Table 17. Signed contribution classification against Human Review.

IECLA reaches a Sign Macro-F1 of 0.3601, more than twice the strongest competing score of 0.1765. Because this classification uses independent Human Review annotations and a fixed sign threshold, the result complements the continuous signed-correlation analysis. IECLA aligns more closely with both regional ordering and the discrete distinction between evidence that supports or opposes the perceived source emotion.

This comparison tests the usefulness of the calibrated sign for distinguishing supporting and opposing perceptual evidence; it does not establish an absolute causal polarity for every candidate region.

User Study

This appendix provides the complete protocol and detailed results for the blinded User Study in Section 4.7. The study has two complementary components: a heatmap evaluation of whether the attributed regions are accurate and easy to interpret, and a paired editing evaluation of whether the same localized evidence improves target-emotion expression while preserving naturalness. We randomize candidate order independently within each task.

Protocol and scoring. Forty independent reviewers score the anonymized candidates on a five-level scale from −2 to 2. Heatmap candidates are evaluated by source-evidence accuracy and focus clarity; editing pairs are evaluated by target expression and naturalness. After decoding the blinded assignments, we aggregate reviewer scores. Table 18 summarizes the evaluation units and the corresponding ordinal reliability coefficients.

ComponentTasksEvaluation unitsKrippendorff’s α\alpha
Heatmap205200.8392
Editing120 paired4800.7364
Overall–1,000–

Table 18. Ordinal inter-rater reliability in the blinded User Study. Each evaluation unit is rated by all 40 reviewers.

The two components exhibit different levels of ordinal inter-rater reliability. Heatmap ratings reach a Krippendorff’s α\alpha of 0.8392, while the paired editing ratings reach 0.7364. These coefficients are computed from the unaggregated ratings of all 40 reviewers and measure reviewer consistency rather than method superiority. Reviewer means are used only for the subsequent task-level comparisons, so the reliability results support the interpretation of the study without replacing the task-level tests.

Heatmap evaluation. Each heatmap task fixes the source image and target emotion, then presents anonymized candidate maps generated under the same candidate-region protocol. The two criteria are averaged per candidate and then across the 20 tasks. The complete comparison is reproduced in Table 19. Its Friedman statistic and Kendall’s WW indicate that the observed ordering is descriptive because the repeated-measures test does not reach significance.

MethodMean score↑\uparrowTop score↑\uparrow
LayerCAM0.77510
CLIP0.92512
HiResCAM0.95013
ADAPT-TF0.95013
DINOV30.97513
Grad-CAM++0.97514
EigenCAM0.97513
RISE-2561.02515
SigLIP21.02514
Direct attention1.02514
Grad-CAM1.02515
Ablation-CAM1.02516
IECLA (ours)1.07518

Table 19. Detailed heatmap User Study results. Mean score averages source-evidence accuracy and focus clarity; top-score counts include task-level ties among 20 tasks.

The heatmap scores are tightly clustered: most methods lie between 0.925 and 1.025, and task-level ties are frequent. Within this compressed range, IECLA attains the highest mean and appears among the top-scoring candidates on 18 of 20 tasks.

Editing evaluation. For each editor, the same 20 source–target jobs are rendered with the native configuration and with IECLA-guided localized evidence. Each pair fixes the source image, target emotion, editor, and corresponding generation configuration, while the A/B order is randomized independently within each task. The table reports reviewer means and within-editor differences. Mean ratings for both target expression and naturalness increase across all six editors, consistent with the main-paper comparison in Section 4.7. Reviewer means are computed within each evaluation unit before the paired task-level comparison, while ordinal reliability is reported separately for the heatmap and editing components in Table 18.

The sign of the effect is stable across editors, whereas its magnitude is editor dependent. Qwen-Edit and HiDream show the largest target-expression gains, while Qwen-Edit and FLUX.2 show the largest naturalness gains. OmniGen2 already has the highest native target-expression score yet still improves by 0.800, so the benefit is not limited to correcting editors with weak native emotion control. InstructPix2Pix and LongCat exhibit smaller but positive changes on both criteria, indicating that the response to localized evidence varies across editing architectures.

The paired pattern also clarifies the relation between the two subjective objectives. Stronger target expression does not coincide with a naturalness decrease for any of the six editors in this evaluation; all naturalness differences remain positive. Together with the variation in gain magnitude, these point estimates suggest that localized IECLA evidence is useful across editing architectures, while the realized benefit remains dependent on the underlying editor.

EditorTarget expression ↑Naturalness ↑
NativeIECLA-guidedΔNativeIECLA-guidedΔ
Qwen-Edit0.06251.325+1.26250.3751.025+0.650
HiDream0.2001.400+1.2000.4250.675+0.250
InstructPix2Pix0.2250.400+0.1750.5000.575+0.075
FLUX.2-0.3500.150+0.5001.0501.425+0.375
OmniGen20.6501.450+0.8000.3500.600+0.250
LongCat-0.0750.075+0.1500.1250.200+0.075

Table 20. Paired editing ratings across six backbones. IECLA guidance improves both target expression and naturalness for every editor.

Additional Ablations

Table 21 isolates training and editing decisions that are not separated in the main end-to-end ablation. All variants use the same 680-image, 5,440-task evaluation pool, predictor architecture, editor, and metric protocol; only the named component or supervision term is changed.

VariantTarget Success ↑Δp ↑Emo-A ↑Emo-S ↑CLIP-I ↑G. LPIPS ↓G. SSIM ↑O. LPIPS ↓O. SSIM ↑O. MAE ↓
IECLA0.35790.33440.30810.21730.85270.34400.61370.37670.32300.1733
-w/o Labs\mathcal{L}_{abs}0.34120.32100.29010.21500.84210.35070.60000.37210.31850.1831
-w/o Lsigned\mathcal{L}_{signed}0.34710.32260.29210.21500.84150.35330.59750.37510.31730.1842
-w/o Lfocal\mathcal{L}_{focal}0.33900.31640.28990.21100.84380.35150.59960.37290.31850.1834
-w/o Laux\mathcal{L}_{aux}0.34650.32210.29140.21430.84180.35340.59690.37520.31690.1842
-w/o fusion0.35420.33180.29210.21390.84510.34760.60730.38510.32140.1746
-w/o template0.35180.32800.29690.21310.83970.35240.60350.38620.31800.1793
-w/o negative protection0.35040.33350.28200.19800.84540.34400.60280.38390.31670.1774
Random matched0.27390.25070.27060.16580.84810.34420.60510.37720.31430.1716

Table 21. Additional editing and loss-function ablations. IECLA provides the strongest overall balance between target-emotion control and structural preservation.

The complete IECLA configuration is the reference condition. The “w/o fusion” variant removes global–local context fusion from the contribution predictor, testing whether regional evidence benefits from joint local and scene-level representations. The “w/o template” variant retains attribution and agent planning but omits the structured editing template, testing whether the template converts evidence into consistent editing instructions. The “w/o learning” variant bypasses the contribution predictor and passes the corrected counterfactual contributions directly from attribution to target-aware editing, testing whether learned contribution representations improve over direct evidence transfer. These settings complement the main-paper “w/o attribution”, “w/o learning”, and “w/o agent” conditions rather than duplicating them.

Two controls target the use of signed evidence during editing. In w/o negative protection, positive contribution regions and the affective instruction are unchanged, but the prompt omits preservation constraints derived from negatively contributing regions. This tests whether negative contributions act as preservation evidence rather than merely being unused scores. In Random matched, the attributed positive region is replaced by a deterministically sampled candidate satisfying the same matching criteria, while prompt structure and editing strength are retained. This control tests whether the gain comes from selecting emotionally relevant evidence instead of from structured prompting alone. The fixed seed and image identifier make the sampled candidate reproducible.

The remaining rows remove one loss from Equation (5) at a time while keeping the predictor, data, and downstream editing protocol fixed. “w/o Labs\mathcal{L}_{abs}” removes absolute ranking supervision and tests contribution-magnitude ordering; “w/o Lsigned\mathcal{L}_{signed}” removes signed ranking supervision and tests direction-aware ordering; “w/o Lfocal\mathcal{L}_{\mathrm{focal}}” removes focal-sign supervision and tests emphasis on difficult or imbalanced sign decisions; and “w/o Laux\mathcal{L}_{\mathrm{aux}}” removes auxiliary supervision and tests the contribution of the auxiliary affective signals. Bold and underlined entries denote the best and second-best point estimates in each column.

The editing controls separate evidence selection from instruction structure. Removing negative protection lowers all four affective measures and reduces outside-region SSIM, while G. LPIPS remains tied with IECLA at 0.3440. Negative contributions therefore constrain unnecessary changes without simply suppressing the edit. Random matched causes the largest affective drop, reaching 0.2739 Target Success and 0.1658 Emo-S, even though it obtains the lowest O. MAE (0.1716). The combination of weak emotion control and conservative pixel change shows that the selected region, rather than structured prompting alone, determines whether the edit is useful. Removing fusion or the template produces smaller but consistent losses in affective and structural measures, with the template ablation particularly lowering CLIP-I to 0.8397 and increasing O. LPIPS to 0.3862.

The main-paper component ablations expose three distinct roles in the editing pipeline. Removing attribution leaves the agent without region-specific affective evidence, so its planning cannot reliably identify where the target emotion should be expressed or which regions should be preserved. Removing learning retains the attribution signal but bypasses its learned calibration, context integration, and stable magnitude and sign ordering; direct evidence transfer therefore provides a noisier control signal and permits more unintended changes outside the selected region. Removing the agent retains the learned contribution representation, allowing the editor to use localized evidence directly, but loses the structured decomposition of target-emotion actions and preservation constraints that makes the evidence consistently executable. Together, the three conditions separate the roles of evidence construction, evidence calibration, and evidence-to-instruction planning.

The loss ablations reveal a complementary training effect. Removing focal-sign supervision gives the largest affective degradation among the four loss variants, whereas removing either ranking term or auxiliary supervision produces smaller declines across the affective metrics. Individual variants can improve one preservation distance, such as O. LPIPS after removing absolute ranking, but none matches IECLA’s joint profile: the complete model is best on all four affective measures, CLIP-I, G. SSIM, and O. SSIM, while tying the best G. LPIPS value and remaining second-best on O. MAE. The results link the full objective to three functions of the framework: ordering relevant evidence, preserving the direction of signed contributions, and maintaining structural fidelity during editing.

Stability Analysis

Robustness to sampling variation is assessed with Qwen-Edit, using the configuration in Appendix J, on 16 routable images stratified by the eight source emotions, with two images per class. Each image is edited toward the other seven emotions, yielding 112 matched tasks. The source image, target emotion, agent configuration, routing protocol, and SAM regions are fixed. Native receives only the global target-emotion instruction; IECLA-guided editing augments it with at most three human-related SAM regions whose predicted positive contribution exceeds 0.05. Each pair shares its source image, target emotion, seed, and resolution.

Seeds 48141, 48142, and 48143 are evaluated at the main resolution 832×480832 \times480 and the adjacent resolution 1024×5761024 \times576. Each resolution contains 112×2×3=672112 \times2 \times3 = 672 outputs, or 1,344 in total. Target Success is the frozen EmoVerse classifier’s target decision, and Target probability Δ\Delta measures change relative to the source image. Preservation is measured outside the hard union of routed SAM masks. Affective IECLA-guided–Native differences use 10,000 source-level paired bootstrap resamples. Preservation values remain point differences because verified paired intervals were not retained; resolution-level standard deviations are computed across the three seed-level means.

IECLA-guided editing improves Target Success and target probability for every seed and resolution. At 1024×5761024 \times576, all affective intervals exclude zero; the three-seed mean gains are +0.101±0.005+0.101 \pm0.005 for Target Success and +0.095±0.015+0.095 \pm0.015 for target probability. At 832×480832 \times480, the corresponding gains are smaller at +0.054±0.009+0.054 \pm0.009 and +0.053±0.010+0.053 \pm0.010; the seed-48141 intervals cross zero, while the other two exclude zero. Outside-mask LPIPS decreases and SSIM increases in all six settings. The repeated directions support sampling stability within each resolution, while their magnitudes indicate sensitivity to output resolution.

Table 22: Stability across seeds and resolutions. IECLA shows favorable affective and outside-mask preservation differences in all six settings.

SeedΔ\Delta Target Success [95% CI]Δ\Delta Target prob. [95% CI]Outside-mask Δ\Delta LPIPSOutside-mask Δ\Delta SSIM
Resolution: 1024 ×\times 576
48141+0.098 [+0.027, +0.179]+0.098 [+0.034, +0.169]-0.010+0.001
48142+0.107 [+0.018, +0.214]+0.108 [+0.035, +0.199]-0.018+0.002
48143+0.098 [+0.036, +0.170]+0.078 [+0.036, +0.128]-0.005+0.010
Resolution: 832 ×\times 480
48141+0.045 [-0.009, +0.107]+0.044 [-0.013, +0.107]-0.011+0.005
48142+0.054 [+0.009, +0.116]+0.051 [+0.004, +0.108]-0.008+0.001
48143+0.062 [+0.018, +0.107]+0.063 [+0.019, +0.104]-0.005+0.008

Table 22. Stability across seeds and resolutions. IECLA shows favorable affective and outside-mask preservation differences in all six settings.

The stability evidence is limited to 16 source images, three seeds, two resolutions, and one editor. It supports consistency of the observed direction under this protocol, rather than universal robustness across editors, samplers, resolutions, or image populations.

Teacher Replay and Predictor Comparison

This appendix isolates whether the learned contribution predictor improves on direct replay of the intervention-based supervision that defines its target. The comparison is useful because the replay is an exact counterfactual measurement, whereas the predictor can denoise that signal and generalize it to the reference-validation protocol.

Experimental setup. The replay uses the same frozen Phase-1 emotion evaluator EE that constructs the corrected counterfactual contributions in § 3.2. For each of the 1,032 aggregated candidate regions on the 289-image reference-validation set, we infer the source emotion with EE, apply the candidate blur intervention, apply its matched reference intervention, and compute the signed contribution with Equation (2). No predictor is trained or selected for this condition. Candidate masks, reference-region matching, prediction-map rendering, region aggregation, sign threshold, and image-macro metric computation are identical to the reference-validation protocol. The replay is compared with IECLA’s learned predictor against Human Review, Qwen Chat, Qwen Reasoner, GPT Terra, and their unweighted VLM mean.

The replay results in Table 23 show that the learned predictor is not merely a compressed copy of the teacher’s ranking. IECLA is higher on every metric for the three individual VLM references and on the VLM mean, while it also improves Human Review agreement on all metrics except the tied Top-5 score. The largest gaps occur in signed measures: on the VLM mean, Signed Spearman rises from 0.4330 to 0.5100, Sign Accuracy from 0.3356 to 0.5333, and Sign F1 from 0.2652 to 0.4695. This pattern is consistent with the predictor learning a more stable signed representation from the intervention supervision rather than reproducing a single noisy replay.

Table 23: Teacher replay versus the IECLA predictor under the reference-validation protocol.

ReferenceImp. ρ\rho ↑ ReplayImp. ρ\rho ↑ IECLAImp. ρ\rho ↑ Δ\DeltaSigned ρ\rho ↑ ReplaySigned ρ\rho ↑ IECLASigned ρ\rho ↑ Δ\DeltaTop-1 ↑ ReplayTop-1 ↑ IECLATop-1 ↑ Δ\DeltaTop-3 ↑ ReplayTop-3 ↑ IECLATop-3 ↑ Δ\DeltaTop-5 ↑ ReplayTop-5 ↑ IECLATop-5 ↑ Δ\Delta
Human Review0.37300.5001+0.12710.40940.4525+0.04310.80970.8685+0.05880.97580.9896+0.01380.99650.9965+0.0000
Qwen Chat0.43520.6187+0.16650.45220.4868+0.03460.77510.8720+0.09690.97920.9896+0.01040.99311.0000+0.0069
Qwen Reasoner0.43300.5998+0.16680.41950.5108+0.09130.77850.8789+0.10040.97230.9792+0.00690.99311.0000+0.0069
GPT Terra0.44770.6165+0.16880.42750.5324+0.10490.76120.8789+0.11770.96540.9827+0.01730.99311.0000+0.0069
VLM mean0.44430.6117+0.16740.43300.5100+0.07700.77160.8766+0.10500.97230.9838+0.01150.99311.0000+0.0069
ReferenceNDCG@3† ReplayNDCG@3† IECLANDCG@3† Δ\DeltaNDCG@5† ReplayNDCG@5† IECLANDCG@5† Δ\DeltaPointing† ReplayPointing† IECLAPointing† Δ\DeltaSign Accuracy† ReplaySign Accuracy† IECLASign Accuracy† Δ\DeltaSign F1† ReplaySign F1† IECLASign F1† Δ\Delta
Human Review0.90710.9315+0.02440.93610.9537+0.01760.79930.8478+0.04850.20550.4320+0.22650.15220.3601+0.2079
Qwen Chat0.89340.9331+0.03880.92290.9517+0.02880.81310.8685+0.05540.33340.5164+0.18300.26300.4500+0.1870
Qwen Reasoner0.89140.9313+0.03990.91470.9489+0.03420.79930.8616+0.06230.36600.5440+0.17800.28800.4842+0.1962
GPT Terra0.88510.9323+0.04720.91870.9521+0.03340.79240.8547+0.06230.30740.5395+0.23210.24460.4745+0.2299
VLM mean0.89030.9322+0.04190.91880.9509+0.03210.80160.8616+0.06000.33560.5333+0.19770.26520.4695+0.2043

Table 23. Teacher replay versus the IECLA predictor under the reference-validation protocol.

The result also clarifies the role of direct replay. The teacher replay remains the exact intervention-defined source of supervision, but its direct maps are weaker than the learned predictor under the independent VLM and Human Review references. At deployment, replay would require repeated evaluator calls for candidate and matched reference interventions, whereas IECLA reuses one source-emotion inference and applies the learned predictor once per region. Although latency was not measured in this experiment, this replacement reduces counterfactual computation and is expected to lower inference time.

Agent Template

The extended agent retains attribution-grounded local editing while admitting two optional, independently controlled channels for scene atmosphere and minor-object operations. Editable and preservation evidence remain distinct throughout planning. A fourth round audits the complete assembled instruction after the structured plan has passed visual and structural review. New-object proposals are treated as context-conditioned editing hypotheses rather than observed attribution evidence. Figure 10 summarizes the paper-level template.

Extended attribution-grounded editing agent template

Figure 10. Extended attribution-grounded editing agent template. Verified evidence is routed through local, atmosphere, and optional minor-object operations; structural review and whole-instruction auditing constrain the final editor instruction.

Locked request source image II; predicted source emotion yy; target emotion tt; attributed regions ri=(id,label,bbox,c^i)r_i = (\mathrm{id}, \mathrm{label}, \mathrm{bbox}, \hat{c}_i) partitioned into editable and preservation evidence; verified placement spaces; output geometry. Region IDs, boxes, emotion pair, and output size are immutable.

Round 1A: Evidence verification Verify editable and preservation evidence, the primary emotional carrier, visible affective cues, scene lighting and palette, existing minor objects, and feasible placement spaces. Report observed evidence and editability risks only.

Round 1B: Object association Propose up to three context-compatible minor-object hypotheses from verified objects or placement spaces. Prefer modify over remove over add; returning none is valid.

Editing hypothesis, not attribution evidence

Round 2: Budgeted route and action planning

Local cue Existing editable content; reduce the source cue and/or establish a visible target cue.

Scene atmosphere Approved photometric or environmental dimensions; preserve scene identity and geometry.

Minor object Modify, add, or remove one verified secondary object under placement and occlusion constraints.

Select the least disruptive valid route; use at most one operation per channel.

Round 3: Visual and structural critic Check executability, emotional completeness, identity and geometry preservation, object existence or placement, occlusion, perspective, and cross-channel consistency. Approve, minimally revise, drop an optional channel, or mark the task invalid.

Draft compiler Assemble approved actions into an ordered draft.

Round 4 auditor Check emotion logic, referents, conflicts, action load, and route consistency.

Final compiler Validate locked fields and apply one protection profile.

Image editor Apply the audited instruction to the source image.

Control invariants. One primary emotional action; at most one atmosphere action and one minor-object action. Protect regions identified as preservation evidence, people, and primary objects; preserve the selected route’s structural invariants; and keep attribution evidence distinct from proposed additions. Invalid plans are not silently replaced by a native prompt or unchanged input.

The route-specific compiler emits only approved clauses and one compatible protection profile. This avoids contradictory constraints such as simultaneously adding an object and requiring object count to remain unchanged.

Implementation Details

This appendix specifies the training configuration and editing controls used in § 3 and § 4.1.

Data Provenance and Splits

The contribution-training, locked attribution-test, and editing-evaluation pools are constructed from EmoVerse. Images belonging to the external evaluation datasets, including EmoSet and EmoArt, are removed before contribution-model training. The training pool supplies 1,056,949 candidate masks, whereas the image-disjoint locked attribution test contains 7,278 images and 53,180 masks. The editing benchmark uses a separate set of 680 challenging images selected to include multi-subject scenes, scenes without a dominant subject, and scenes containing many objects. Deduplication and overlap checks are applied before the final splits, leaving the training, locked attribution-test, and editing-evaluation pools mutually disjoint.

Editing Tasks and Evaluation Metrics

Task construction. The main editing pool contains 680 source images, each paired with all eight target emotions: amusement, anger, awe, contentment, disgust, excitement, fear, and sadness. This construction yields 680×8=5,440680 \times8 = 5{,}440 source–target task keys, including 680 same-emotion tasks (t=yt = y) and 4,760 cross-emotion tasks (t≠yt \ne y), where the frozen image-level evaluator supplies the source emotion yy. The six editors use the same task keys, and Native and IECLA-guided conditions share the source image, target emotion, seed, sampler, and resolution rule. Table 2, Table 4, and Table 21 use the same evaluation set for their corresponding Qwen-Edit conditions, which accounts for their identical shared entries.

Cross-dataset editing applies every target emotion to every image used in the attribution evaluation. The 3,370 external images therefore produce 26,960 editing tasks, comprising 3,370 same-emotion and 23,590 cross-emotion tasks. Table 3 averages over 4,896 COCO tasks, 7,320 EmoSet tasks, 12,352 FI tasks, 1,808 Emotion6 tasks, and 584 EmoArt tasks. The stability analysis in Appendix § 15 is a separate diagnostic: its 16 images are edited only toward the other seven emotions, giving 112 tasks for each seed and resolution.

EvaluationImagesTargets/imageSame-emotionCross-emotionTasks
Main editing pool68086804,7605,440
COCO61286124,2844,896
EmoSet91589156,4057,320
FI1,54481,54410,80812,352
Emotion622682261,5821,808
EmoArt73873511584
Cross-dataset total3,37083,37023,59026,960
Stability, per seed/resolution1670112112

Table 24. Editing-task accounting. The main and cross-dataset evaluations include all eight targets, whereas the controlled cross-emotion diagnostics exclude t=yt = y.

Affective metrics. Let IsI_s and IeI_e denote the source and edited images, and let tt be the requested emotion. The frozen EmoVerse-supervised evaluator EE defines

Success⁡(Ie,t)=I[arg⁡max⁡e Ee(Ie)=t],\operatorname{Success}(I_e,t)=\mathbb{I}\left[\underset{e}{\arg\max}\ E_e(I_e)=t\right],
Δp(Is,Ie,t)=Et(Ie)−Et(Is).(6)\Delta p(I_s,I_e,t)=E_t(I_e)-E_t(I_s). \tag*{(6)}

An independent CLIP-based emotion classifier CC provides a second evaluation:

Emo-A⁡(Ie,t)=I[arg⁡max⁡e Ce(Ie)=t],\operatorname{Emo\text{-}A}(I_e,t)=\mathbb{I}\left[\underset{e}{\arg\max}\ C_e(I_e)=t\right],
Emo-S⁡(Is,Ie,t)=Ct(Ie)−Ct(Is).(7)\operatorname{Emo\text{-}S}(I_s,I_e,t)=C_t(I_e)-C_t(I_s). \tag*{(7)}

The classifier CC follows the EmoEdit evaluation model: a frozen CLIP ViT-B/32 image encoder feeds a two-layer classifier with dimensions 512 →\to 256 →\to 8 and a softmax output. Its released checkpoint was pretrained on EmoSet. We performed image-level overlap checks between its EmoSet training data and the EmoSet subset used in our cross-dataset evaluation, confirming that the two sets are disjoint. Thus, Target Success and Emo-A are target-class accuracies from different held-out evaluators, while Δp\Delta p and Emo-S measure target-probability changes from the corresponding evaluators. Reported entries are arithmetic means over the task set specified for each experiment; source-image cluster resampling is used when confidence intervals are reported.

Evaluator roles and metric provenance. The frozen evaluator EE provides the source-emotion label and counterfactual supervision, while the contribution predictor scores candidate regions and the target-aware agent plans edits from those scores. The classifier CC is used only for post-hoc Emo-A and Emo-S evaluation; its scores are never fed back to the predictor or agent. VLM references, Human Review, and the User Study provide complementary external and perceptual evaluations.

Global preservation metrics. All source and edited images are converted to RGB and resized to 512 ×\times 512 with Lanczos interpolation. PSNR, SSIM, and LPIPS compare the complete images, while CLIP-I is the cosine similarity between their CLIP image embeddings. In Table 4 and Table 21, the prefix G. denotes the same complete-image evaluation. G. SSIM is computed on uint8 images with data range 255. G. LPIPS uses the AlexNet LPIPS backbone on float32 images scaled to [0, 1]. We call the official LPIPS implementation with normalize=True, which internally maps these inputs to the expected [−1, 1] range:

G. LPIPS=DLPIPS(Is,Ie).(8)\mathrm{G.\ LPIPS}=D_{\mathrm{LPIPS}}(I_s,I_e). \tag*{(8)}

All compared conditions use this identical normalized implementation. The reported values therefore follow the standard LPIPS preprocessing while retaining the common [0, 1] image representation used by the rest of our evaluation code.

Common outside-region protocol. Outside-region metrics measure changes beyond a common allowed editing area. For each task, the selected binary SAM masks from all compared conditions are resized to 512 ×\times 512 by nearest-neighbor interpolation and combined by a hard union,

A=⋃v∈VMv,O=¬A,(9)A=\bigcup_{v\in\mathcal{V}}M_v,\qquad O=\neg A, \tag*{(9)}

where V\mathcal{V} is the set of variants in the comparison. Fixing AA across variants prevents a condition from improving its preservation score by declaring a larger editable region. In the loss-function ablation, V\mathcal{V} contains the complete objective and the four loss-removed variants. Their common allowed region covers 22.7521% of the image on average, leaving 77.2479% for outside-region evaluation. This experiment uses the same 680 source images and all eight target emotions, yielding 680 ×\times 8 = 5,440 tasks.

O. MAE averages the RGB absolute error over OO:

O. MAE=1∣O∣∑p∈O∥Ie(p)−Is(p)∥13.(10)\mathrm{O.\ MAE}=\frac{1}{|O|}\sum_{p\in O}\frac{\lVert I_e(p)-I_s(p)\rVert_1}{3}. \tag*{(10)}

O. SSIM first computes the full-resolution SSIM map and then averages its values over OO. For O. LPIPS, the allowed region is restored to the source image,

Ioutside=A⊙Is+(1−A)⊙Ie,O. LPIPS=DLPIPS(Is,Ioutside).(11)I_{\mathrm{outside}}=A\odot I_s+(1-A)\odot I_e,\qquad\mathrm{O.\ LPIPS}=D_{\mathrm{LPIPS}}(I_s,I_{\mathrm{outside}}). \tag*{(11)}

O. LPIPS is therefore an outside-focused perceptual distance rather than a pixelwise regional average; convolutional receptive fields may still cross the mask boundary.

The affective and preservation metrics are interpreted jointly: minimal image change is not beneficial when the requested emotion is not expressed, while improved affective scores accompanied by poor outside-region preservation indicate unnecessary rewriting.

Counterfactual Reference Calibration

The reference correction in Equation (2) uses dataset-level calibration rather than a fixed number kk of reference masks for each candidate. We sample Nref=10,000N_{\mathrm{ref}} = 10{,}000 training records and generate one reference intervention per record. Reference regions are matched to candidate regions in nominal area, spatial position, and shape. The resulting reference effects are averaged within source-emotion and area bins, using five area ranges: [0,0.005)[0, 0.005), [0.005,0.02)[0.005, 0.02), [0.02,0.10)[0.02, 0.10), [0.10,0.30)[0.10, 0.30), and [0.30,1][0.30, 1]. Each emotion–area cell contains at least 30 references; otherwise, the area-level and then global mean provide the fallback. Thus, E[Δr]\mathbb{E}[\Delta_r] denotes the corresponding calibration mean, not a per-region Monte Carlo average.

Both candidate and reference interventions replace the matched region with a radius-24 Gaussian-blurred version. Matching area, position, and shape ensures comparable intervention geometry, while averaging over the calibration set estimates the blur-induced nuisance effect. Subtracting this expectation leaves the candidate-specific residual as the corrected contribution.

We interpret cic_i on this calibrated relative scale: ci<0c_i < 0 indicates an effect below the matched-reference mean, while its role as an opposing editing cue is validated against Human Review.

Predictor Training and Objective

Predictor training. The contribution predictor is trained on 1,056,949 candidate masks and validated on 58,591 disjoint masks. Every loss term uses the corrected counterfactual contribution cic_i from Equation (2) as its supervision target, and checkpoint selection minimizes validation MAE on these corrected contributions. Training uses three NVIDIA A800 GPUs.

Supervision and weighted regression. Let ci^\widehat{c_i} denote the predicted contribution and τ=0.1\tau= 0.1 the fixed sign threshold. The target sign is negative for ci<−τc_i < -\tau, neutral for ∣ci∣≤τ\lvert c_i\rvert\le\tau, and positive for ci>τc_i > \tau. We define the unit-transition Huber penalty as

ρ(e)={12e2,∣e∣≤1,∣e∣−12,∣e∣>1.(12)\rho(e) = \begin{cases} \frac{1}{2}e^2, & \lvert e\rvert\le1, \\ \lvert e\rvert- \frac{1}{2}, & \lvert e\rvert> 1. \end{cases} \tag*{(12)}

Let bib_i be the evaluator’s pre-intervention class and y~i\widetilde{y}_i the image-level training label. The sample weight

wi=31[ci<−τ]1.51[∣ci∣>τ]0.51[bi≠y~i](13)w_i = 3^{\mathbb{1}[c_i<-\tau]}1.5^{\mathbb{1}[\lvert c_i\rvert>\tau]}0.5^{\mathbb{1}[b_i\ne\widetilde{y}_i]} \tag*{(13)}

combines the three multipliers rather than selecting only one of them. The normalized regression term is

LHuberw=∑iwiρ(ci^−ci)∑iwi.(14)\mathcal{L}_{\mathrm{Huber}}^{w} = \frac{\sum_i w_i\rho(\widehat{c_i}-c_i)}{\sum_i w_i}. \tag*{(14)}

Within-image ranking. Let g(i)g(i) identify the source image of region ii. Absolute and signed ranking use, respectively, the valid pair sets

Pabs={(i,j):g(i)=g(j), i<j, ∣ci∣≠∣cj∣},Psigned={(i,j):g(i)=g(j), i<j, ci≠cj}.(15)\mathcal{P}_{\mathrm{abs}} = \{(i,j): g(i)=g(j),\ i<j,\ \lvert c_i\rvert\ne\lvert c_j\rvert\}, \qquad \mathcal{P}_{\mathrm{signed}} = \{(i,j): g(i)=g(j),\ i<j,\ c_i\ne c_j\}. \tag*{(15)}

With σ(⋅)\sigma(\cdot) denoting the sigmoid function, the two losses are

Labs=−1∣Pabs∣∑(i,j)∈Pabslog⁡σ(sija(∣ci^∣−∣cj^∣)),sija=sign⁡(∣ci∣−∣cj∣),\mathcal{L}_{\mathrm{abs}} = -\frac{1}{\lvert\mathcal{P}_{\mathrm{abs}}\rvert}\sum_{(i,j)\in\mathcal{P}_{\mathrm{abs}}}\log\sigma\left(s_{ij}^{a}(\lvert\widehat{c_i}\rvert-\lvert\widehat{c_j}\rvert)\right),\quad s_{ij}^{a}=\operatorname{sign}(\lvert c_i\rvert-\lvert c_j\rvert),
Lsigned=−1∣Psigned∣∑(i,j)∈Psignedlog⁡σ(sijs(ci^−cj^)),sijs=sign⁡(ci−cj).(16)\mathcal{L}_{\mathrm{signed}} = -\frac{1}{\lvert\mathcal{P}_{\mathrm{signed}}\rvert}\sum_{(i,j)\in\mathcal{P}_{\mathrm{signed}}}\log\sigma\left(s_{ij}^{s}(\widehat{c_i}-\widehat{c_j})\right),\quad s_{ij}^{s}=\operatorname{sign}(c_i-c_j). \tag*{(16)}

The absolute term orders regions by contribution strength, whereas the signed term preserves their ordering on the corrected contribution scale.

Sign and affective auxiliary supervision. For the three sign classes qi∈{−,0,+}q_i \in\{-, 0, +\}, let piqip_i^{q_i} be the predicted probability of the target class. With class weights (α−,α0,α+)=(4.0,0.3,1.1)(\alpha_{-}, \alpha_0, \alpha_{+}) = (4.0, 0.3, 1.1) and focal exponent γ=2\gamma= 2, the sign-classification loss is

Lfocal=−∑iαqi(1−piqi)2log⁡piqi∑iαqi(17)\mathcal{L}_{\mathrm{focal}} = -\frac{\sum_i \alpha_{q_i}(1-p_i^{q_i})^2 \log p_i^{q_i}}{\sum_i \alpha_{q_i}} \tag*{(17)}

The auxiliary heads predict the intensity difference diintd_i^{\mathrm{int}} and the three-dimensional VAD difference diVAD\mathbf{d}_i^{\mathrm{VAD}} associated with the same intervention. For a batch of BB regions, their loss is

Laux=1B∑iρ(d^iint−diint)+0.23B∑i∑k=13ρ(d^ikVAD−dikVAD).(18)\mathcal{L}_{\mathrm{aux}} = \frac{1}{B}\sum_i \rho(\hat{d}_i^{\mathrm{int}}-d_i^{\mathrm{int}}) + \frac{0.2}{3B}\sum_i\sum_{k=1}^{3}\rho(\hat{d}_{ik}^{\mathrm{VAD}}-d_{ik}^{\mathrm{VAD}}). \tag*{(18)}

The coefficient 0.2 applies within Laux\mathcal{L}_{\mathrm{aux}}, Equation (5) further weights the full auxiliary term by 0.10.

Grouped sampling and pair construction. Candidate regions are grouped by source image, image groups are shuffled each epoch, and complete groups are packed without splitting into batches of up to 128 regions per process. With three training processes, this gives the effective batch size of 384 reported above. Both ranking terms enumerate every valid unordered pair within each image group; they use neither random pair subsampling nor a ranking margin. Tied targets are omitted, and image groups with fewer than two regions contribute zero to the ranking terms.

Attribution Baseline Implementations

The vision-encoder contribution baselines are trained comparison versions used to produce signed heatmaps. They share the candidate masks, counterfactual supervision, and mask-to-pixel rendering protocol of IECLA, but use simpler backbone-specific predictors. The CLIP version combines frozen ViT-L/14 region-crop and full-image features with emotion conditioning, object labels, and region metadata. The SigLIP2 version uses a frozen SigLIP2-base encoder with region and full-image context, emotion and object conditions, and the same metadata. The DINOv3 version pools frozen ViT-7B/16 patch tokens within each SAM mask and combines the region representation with full-image context, emotion and object conditions, and metadata. Each predictor outputs a signed regional contribution that is projected onto the fixed masks; these baselines are trained specifically for heatmap comparison and are not native attention maps.

Grad-CAM, Grad-CAM++, HiResCAM, LayerCAM, EigenCAM, and Ablation-CAM use the frozen eight-class emotion evaluator EE as their target network and explain its inferred source-emotion class. RISE-256 uses the same evaluator with 256 random masks sampled on a 7×77 \times7 grid. Each resulting saliency map is resized to the input resolution, and its score for a candidate region is the mean saliency within the corresponding fixed SAM mask. Direct attention instead uses the CLIP visual encoder from the CLIP contribution baseline: multi-head self-attention is augmented with the identity connection and rolled out across transformer layers to obtain the CLS-to-patch map. ADAPT-TF likewise follows its own Qwen image-token attention path rather than the evaluator EE. These heatmap methods produce nonnegative region scores and are therefore evaluated only with unsigned ranking and localization metrics.

Agent constraints and candidate selection. The predictor supplies up to three attributed regions with c^i>0.05\hat{c}_i > 0.05 as editable evidence and retains strongly negative regions as preservation evidence. The agent verifies a primary emotional carrier, and the local action remains anchored to the editable evidence. Scene-atmosphere changes and minor-object operations are optional, independently budgeted channels that require visible scene support; a proposed added object is recorded as an editing hypothesis rather than attributed evidence. At most one primary, one atmosphere, and one minor-object operation are retained. A visual and structural critic checks feasibility and preservation, after which a deterministic draft compiler and integrated-instruction auditor instantiate the template in Appendix I. Reporting evaluators remain held out from planning and candidate selection.

Agent VLM configuration. All agent rounds use the same locally served vision-language model under the API identifier qwen3-vision. The model is loaded from Qwen3-6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q8_K_P.gguf together with mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf. It uses the qwen35moe architecture with 34.66B total parameters and approximately 3B active parameters, and the language checkpoint uses Q8_K_P quantization. Thinking is enabled for every round; decoding uses temperature 0.1, the server default top-p of 0.95, at most 1,024 output tokens per round, and an 8,192-token service context.

The original image is transmitted as base64 data without client-side resizing. The vision encoder has a nominal input size of 768 and a patch size of 16, accepts dynamic inputs from 8,192 to 4,194,304 pixels, and uses a qwen3vl_merger projector with spatial merge size 2.

Editor configurations. The editing experiments use Qwen-Edit, InstructPix2Pix, FLUX.2, LongCat, OmniGen2, and HiDream. Qwen-Edit denotes Qwen-Image-Edit-2511, FLUX.2 denotes FLUX.2 Klein 4B, LongCat denotes LongCat-Turbo, and HiDream denotes HiDream-E1.1; InstructPix2Pix and OmniGen2 use their official released checkpoints. The stability analysis in Appendix § 15 runs Qwen-Edit with Lightning four-step sampling.

Editing targets. The eight target emotions are amusement, anger, awe, contentment, disgust, excitement, fear, and sadness. Native and IECLA-guided outputs share the source image, target emotion, seed, sampler, and resolution rule within each editor. Absolute scores across different backbones are not interpreted as a ranking of editors.

References

  1. [1]Panos Achlioptas, Maks Ovsjanikov, Kilichbek Haydarov, Mohamed Elhoseiny, and Leonidas Guibas. Artemis: Affective language for visual art. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11564–11574, 2021. doi: 10.1109/CVPR46437.2021.01140.DOI
  2. [2]Julius Adebayo, Justin Gilmer, Michael Muellly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. arXiv preprint arXiv:1810.03292, 2018. doi: 10.48550/arXiv.1810.03292.DOI
  3. [3]Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 18187–18197, 2022. doi: 10.1109/CVPR52688.2022.01767.DOI
  4. [4]Black Forest Labs. FLUX.2 [klein]: Fast, efficient image generation. https://bfl.ai/models/flux-2-klein, 2026.
  5. [5]Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 18392–18403, 2023. doi: 10.1109/CVPR52729.2023.01764.DOI
  6. [6]Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang, Fengbin Gao, Peihan Xu, et al. HiDream-I1: A high-efficient image generative foundation model with sparse diffusion transformer. arXiv preprint arXiv:2505.22705, 2025. doi: 10.48550/arXiv.2505.22705.DOI
  7. [8]Weidong Chen, Cheng Ye, Zhendong Mao, Peipei Song, Xinyan Liu, Lei Zhang, Xiaojun Chang, and Yongdong Zhang. Face-net: Factual calibration and emotion augmentation for retrieval-enhanced emotional video captioning. arXiv preprint arXiv:2603.17455, 2026.arxiv.org/abs/2603.17455
  8. [10]Youssef Doulfoukar, Laurent Mertens, and Joost Vennekens. Emocam: Toward understanding what drives CNN-based emotion recognition. arXiv preprint arXiv:2407.14314, 2024. doi: 10.48550/arXiv.2407.14314.DOI
  9. [12]Ruth C. Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE International Conference on Computer Vision, pp. 3449–3457, 2017. doi: 10.1109/ICCV.2017.371.DOI
  10. [13]Tsu-Jui Fu, Xin Eric Wang, and William Yang Wang. Language-driven artistic style transfer. In European Conference on Computer Vision, pp. 717–734, 2022.DOI
  11. [14]Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. In International Conference on Learning Representations, volume 2024, pp. 54820–54833, 2024.arxiv.org/abs/2309.17102
  12. [15]Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Han Hu, Dong Chen, and Baining Guo. Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895, 2023. doi: 10.48550/arXiv.2309.03895.DOI
  13. [16]Yash Goyal, Ziyan Wu, Jan Ernst, Dhruv Batra, Devi Parikh, and Stefan Lee. Counterfactual visual explanations. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 2376–2384. PMLR, 2019.arxiv.org/abs/1904.07451
  14. [17]Yijie Guo, Dexiang Hong, Weidong Chen, Zihan She, Cheng Ye, Xiaojun Chang, and Zhendong Mao. Emoverse: A mlms-driven emotion representation dataset for interpretable visual emotion analysis. arXiv preprint arXiv:2511.12554, 2025.arxiv.org/abs/2511.12554
  15. [18]Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In International Conference on Learning Representations, 2023. URL https://arxiv.org/abs/2208.01626.
  16. [19]Dexiang Hong, Yijie Guo, Weidong Chen, Xinyan Liu, Zixuan Zou, Zhendong Mao, and Yongdong Zhang. Emostyle: Affective conditioning of style-specialist experts for emotional image generation. arXiv preprint arXiv:2607.10165, 2026.arxiv.org/abs/2607.10165
  17. [20]Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875–5888, 2021. doi: 10.1109/TIP.2021.3089943.DOI
  18. [21]Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4015–4026, 2023. doi: 10.1109/ICCV51070.2023.00371.
  19. [22]Ronak Kosti, Jose M. Alvarez, Adria Recasens, and Agata Lapedriza. Context based emotion recognition using EMOTIC dataset. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(11):2755–2766, 2020. doi: 10.1109/TPAMI.2019.2916866.DOI
  20. [23]Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing, 2023. URL https://arxiv.org/abs/2305.14720.
  21. [24]Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision, pp. 740–755, 2014. doi: 10.1007/978-3-319-10602-1_48.DOI
  22. [25]Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pp. 38–55, 2024. doi: 10.1007/978-3-031-72970-6_3.
  23. [26]Fanhang Man, Xiaoyue Chen, Huandong Wang, Baining Zhao, Han Li, and Xinlei Chen. VAEER: Visual attention-inspired emotion elicitation reasoning. arXiv preprint arXiv:2505.24342, 2025. doi: 10.48550/arXiv.2505.24342.DOI
  24. [27]Qi Mao, Haobo Hu, Yujie He, Difei Gao, Haokun Chen, and Libiao Jin. Emoagent: A multi-agent framework for diverse affective image manipulation. arXiv preprint arXiv:2503.11290, 2025. doi: 10.48550/arXiv.2503.11290.
  25. [28]Meituan LongCat Team, Hanghang Ma, Haoxian Tan, Jiale Huang, Junqiang Wu, Jun-Yan He, Lishuai Gao, Songlin Xiao, Xiaoming Wei, Xiaoqi Ma, Xunliang Cai, Yayong Guan, and Jie Hu. LongCat-Image technical report. arXiv preprint arXiv:2512.07584, 2025. doi: 10.48550/arXiv.2512.07584.DOI
  26. [29]Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. URL https://arxiv.org/abs/2108.01073.
  27. [31]Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 860–868, 2015. doi: 10.1109/CVPR.2015.7298687.DOI
  28. [32]Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: Randomized input sampling for explanation of black-box models. In British Machine Vision Conference 2018, pp. 151. BMVA Press, 2018.arxiv.org/abs/1806.07421
  29. [33]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. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pp. 8748–8763. PMLR, 2021.arxiv.org/abs/2103.00020
  30. [34]Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Why should i trust you?: Explaining the predictions of any classifier. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, pp. 97–101, 2016. doi: 10.18653/v1/N16-3020.DOI
  31. [35]Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision, pp. 618–626, 2017. doi: 10.1109/ICCV.2017.74.DOI
  32. [36]Oriane Siméoni, Huy V. Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeon Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julien Mairal, Hervé Jégou, Patrick Labatut, and Piotr Bojanowski. DINOv3. arXiv preprint arXiv:2508.10104, 2025. doi: 10.48550/arXiv.2508.10104.
  33. [37]Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pp. 3319–3328. PMLR, 2017.
  34. [38]Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdul-mohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier Hénaff, Jeremiah Harmsen, Andreas Steiner, and Xiaohua Zhai. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786, 2025. doi: 10.48550/arXiv.2502.14786.DOI
  35. [39]Liping Wang, Cheng Ye, Weidong Chen, Peipei Song, Bo Hu, and Zhendong Mao. A multi-agent framework with structured reasoning and reflective refinement for multimodal empathetic response generation. arXiv preprint arXiv:2604.18988, 2026.arxiv.org/abs/2604.18988
  36. [40]Cong Wei, Zheyang Xiong, Weiming Ren, Xeron Du, Ge Zhang, and Wenhu Chen. OmniEdit: Building image editing generalist models through specialist supervision. In International Conference on Learning Representations, volume 2025, pp. 259–271, 2025.arxiv.org/abs/2411.07199
  37. [41]Shuchen Weng, Peixuan Zhang, Zheng Chang, Xinlong Wang, Si Li, and Boxin Shi. Affective image filter: Reflecting emotions from text to images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10810–10819, 2023.DOI
  38. [42]Chenfei Wu, Jiahаo Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-Image technical report. arXiv preprint arXiv:2508.02324, 2025a. doi: 10.48550/arXiv.2508.02324.DOI
  39. [43]Chenyuаn Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, et al. OmniGen2: Towards instruction-aligned multimodal generation. arXiv preprint arXiv:2506.18871, 2025b. doi: 10.48550/arXiv.2506.18871.DOI
  40. [44]Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Emoset: A large-scale visual emotion dataset with rich attributes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 20326–20337, 2023. doi: 10.1109/ICCV51070.2023.01864.DOI
  41. [45]Jingyuan Yang, Jiawei Feng, Weibin Luo, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. EmoEdit: Evoking emotions through image manipulation. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 24690–24699, IEEE, 2025.arxiv.org/abs/2405.12661
  42. [46]Ling Yang, Zhilong Zhang, Zhaоchen Yu, Jingwei Liu, Minkai Xu, Stefano Ermon, and Bin Cui. Cross-modal contextualized diffusion models for text-guided visual generation and editing. In International Conference on Learning Representations, volume 2024, pp. 13972–13997, 2024.
  43. [48]Cheng Ye, Weidong Chen, Jingyu Li, Lei Zhang, and Zhendong Mao. Dual-path collaborative generation network for emotional video captioning. In Proceedings of the 32nd ACM International Conference on Multimedia, pp. 496–505, 2024.arxiv.org/abs/2408.03006
  44. [49]Jiarong Ye and Sharon X. Huang. Mmodifier: MLLM-enhanced emotion-driven image editing. arXiv preprint arXiv:2507.14024, 2025. doi: 10.48550/arXiv.2507.14024.DOI
  45. [50]Xinyi Yin, Yiduo Wang, Tingqi Hu, Meicong Si, Yunyun Shi, Shi Chen, Hao Wang, Junciao Xue, and Xuecheng Wu. Mood: Perception-enhanced efficient affective image editing via continuous valence–arousal modeling. arXiv preprint arXiv:2605.02521, 2026. doi: 10.48550/arXiv.2605.02521.DOI
  46. [51]Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. Building a large scale dataset for image emotion recognition: The fine print and the benchmark. In Proceedings of the AAAI Conference on Artificial Intelligence, 2016. doi: 10.1609/aaai.v30i1.9987.DOI
  47. [52]Kaishen Yuan, Yuting Zhang, Shang Gao, Yijie Zhu, Wenshuo Chen, and Yutao Yue. Coemogen: Towards semantically coherent and scalable emotional image content generation. In International Conference on Learning Representations, volume 2026, pp. 6341–6361, 2026.arxiv.org/abs/2508.03535
  48. [53]Cheng Zhang, Hongxia Xie, Bin Wen, Songhan Zuo, Ruoxuan Zhang, and Wen-Huang Cheng. Emoart: A multidimensional dataset for emotion-aware artistic generation. In Proceedings of the 33rd ACM International Conference on Multimedia, pp. 12644–12650, 2025a. doi: 10.1145/3746027.3758201.DOI
  49. [54]Hui Zhang, Dexiang Hong, Maoke Yang, Yutao Cheng, Zhao Zhang, Weidong Chen, Jie Shao, Xinglong Wu, Zuxuan Wu, and Yu-Gang Jiang. Creatidesign: A unified multi-conditional diffusion transformer for creative graphic design. In International Conference on Learning Representations, volume 2026, pp. 111201–111214, 2026.arxiv.org/abs/2505.19114
  50. [57]Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 3813–3824, 2023b. doi: 10.1109/ICCV51070.2023.00355.DOI
  51. [58]Zhao Zhang, Yutao Cheng, Dexiang Hong, Maoke Yang, Gonglei Shi, Lei Ma, Hui Zhang, Jie Shao, and Xinglong Wu. Creatiposter: Towards editable and controllable multi-layer graphic design generation. arXiv preprint arXiv:2506.10890, 2025b.arxiv.org/abs/2506.10890

Paper details

vlnc:W_kxd69kxx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Generative AI tools

View statement
Looking for related papers…
Contents