acceptodds
Withdrawn from ICLR 2027

EmoWeave: Unified Affective Image Manipulation via Emotion-Aware Understanding and Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Affective Image Manipulation (AIM) aims to modify images according to emotional intents, enabling them to evoke desired affective responses. Existing methods typically focus on either appearance or semantic manipulation, require predefined target emotions, and lack explicit mechanisms for coordinating content and appearance edits toward a shared emotional goal. We introduce Unified Affective Image Manipulation (U-AIM), a task formulation that connects emotion intent inference, editing planning, and image editing. Given a source image with an optional target emotion, U-AIM supports both emotion enhancement and emotion migration by generating and executing coordinated content and appearance editing instructions. To provide supervision for this process, we propose Emotion-Grounded Retrieval Augmentation (EGRA), which retrieves emotion-relevant evidence conditioned on the source-image state and target emotion to construct grounded editing instructions. Combining affective image data with retouching data, we build UAIM-14K, linking source images, emotional intents, coordinated instructions, and edited results. We further introduce UAIMBench to evaluate both planning and end-to-end editing under user-specified and model-inferred emotional intents. Based on this supervision, we develop EmoWeave, a unified multimodal model that integrates emotional reasoning, instruction generation, and instruction-conditioned image editing within a single framework. Experiments on UAIMBench demonstrate that EmoWeave outperforms evaluated unified multimodal baselines and applicable AIM methods in both planning quality and end-to-end editing performance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Unified Affective Image Manipulation, Emotional Edit, Unified Multimodal Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rVKcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces U-AIM, a unified affective image manipulation task that jointly considers semantic and appearance editing under either user-specified or model-inferred emotional intents. The authors propose EGRA to integrate professional retouching data with affective semantic knowledge, constructing the UAIM-14K dataset and UAIMBench benchmark. Based on this supervision, they develop EmoWeave, a unified multimodal model that combines emotional intent inference, coordinated editing planning, and image editing. Experiments demonstrate strong performance in both planning quality and end-to-end affective editing compared with existing unified multimodal and emotion-specific editing baselines.

Critical strengths

1.The problem formulation is comprehensive and practically meaningful. U-AIM jointly models semantic- and appearance-level editing and provides a complete pipeline spanning training-data construction, benchmark development, and unified model training, offering strong practical value for affective image manipulation.

2.The experimental evaluation is comprehensive, with extensive baseline comparisons, component-wise ablations, and rich qualitative visualizations. These analyses provide a relatively thorough assessment of both the planning and editing capabilities of EmoWeave and make the effectiveness of the proposed framework intuitive and easy to understand.

Critical weaknesses

1.The single-reference evaluation in Scenario B may not fully match its autonomous intent-inference setting. Since a source image can plausibly support multiple emotional editing directions, evaluating against only one fixed target emotion with exact-match accuracy may underestimate valid predictions.

2.The end-to-end evaluation is limited to only 100 samples out of the 600-sample benchmark. Given the multiple scenarios, editing types, and emotion categories involved, the resulting per-setting sample sizes are relatively small, which may reduce the statistical power of the reported comparisons. Several EmoACC improvements are also statistically non-significant against strong general-purpose baselines.

Questions for authors

1.Can the authors provide direct human evaluation of the emotional responses elicited by the edited images? Since the current affective evaluation largely relies on EEmo-Logic and VLM judges, a human perceptual study would provide stronger evidence that the edits genuinely evoke the intended emotions.

2.To what extent do EmoWeave's gains come from EGRA/U-AIM rather than from the strong models used to construct the supervision, such as DeepSeek-V4-Pro and FireRed-Edit? An analysis comparing EmoWeave with its supervision generators would help disentangle the contribution of the proposed framework from teacher-model knowledge and biases.

3.Can the authors evaluate whether CAST and EGRA generalize to other UMM backbones or model scales? Since the current experiments use only SenseNova-U1.5-8B-MoT, additional backbone-level ablations would help establish whether the gains primarily come from the proposed method rather than backbone-specific properties.

4 comments
AuthorsOpenReview ↗

**[W] Response to rVKc Weaknesses**

W1. Does the single reference in Scenario B underestimate reasonable predictions?

Thank you for this insightful point. We recognize that emotion recommendation can have more than one reasonable answer, so we added human annotators for evaluation, and EmoWeave leads under this evaluation as well. Tab. R1 reports Recommend Acc. (B), which requires an exact match with the preset labels, and a version that counts any recommendation annotators find reasonable. The full human evaluation is in our reply to Q1 in [Q] and will be added to the main text.

Tab. R1. Human evaluation on emotion recommendation in Scenario B (human acceptance rates; 44 samples, 3 annotators each).

Recommend Acc.EmoWeaveQwen3-VL-32B
Preset labels75.00%43.18%
Reasonable emotion84.09%52.27%

W2. How reliable are the end-to-end comparisons given the subset size?

Thank you for this constructive comment. We expanded the end-to-end evaluation from 100 to 200 samples, and EmoWeave keeps its lead over all baselines on all four metrics (Tab. R2). The 200 samples (Scenario A 111, B 89) are stratified-sampled from the 600 benchmark images by category and emotion, so they cover diverse image content (portraits, multi-person images, natural scenery, cities, animals, still life, macro scenes), enhancement and migration across different emotion directions, and all six emotions with a balanced distribution (Fig. 3, L282). Planning quality is evaluated on all 600 samples (Tab. 3).

Tab. R2. End-to-end results. (a) Scenario A + B, 200 samples. (b) Scenario A only, 111 samples; EmoEdit and EmoStyle require a user-specified target, so they are evaluated here.

(a) MethodEmoACCEmo-SSemanticAppearance
EmoWeave0.65000.13546.55006.7600
Qwen-Image-Edit-25110.55000.04885.90506.2600
FireRed-Edit0.54000.05985.99306.1038
BAGEL-Edit0.53000.05175.84504.8100
VeraRetouch0.55000.02374.97805.5930
(b) MethodEmoACCEmo-SSemanticAppearance
EmoWeave0.63960.21106.73876.9324
EmoEdit0.47750.07224.79705.4140
EmoStyle0.3694−0.03074.21203.8420

On the expanded set, EmoWeave's EmoACC is at least 0.10 higher than that of every baseline (0.6500 against at most 0.5500). EmoACC is a top-1 hit rate, so it is the coarser of our two emotion metrics; Emo-S, the shift in the VER probability of the target emotion, separates the methods more clearly, at 0.1354 against 0.0598 for the best baseline.

The Questions are answered in a separate comment, [Q] Response to rVKc Questions. We sincerely thank you for supporting our work, and we hope our replies address your questions.

AuthorsOpenReview ↗

[Q] Response to rVKc Questions

Q1. Can you provide direct human evaluation of the emotional response to the edited images?

Thank you for this important question. We ran a human evaluation of planning quality and U-AIM Performance (Tab. R3). The baselines are representative general-purpose open-source models: Qwen3-VL-32B for planning and FireRed-Edit for editing. Planning quality is evaluated on 100 benchmark samples with 3 annotators each, with agreement 81.34% (Fleiss' κ ≈ 0.558). U-AIM Performance is evaluated in two steps. (1) Without a target hint, annotators see only the edited image and choose the dominant emotion from Ekman-6 plus Neutral (Human EmoACC). (2) In a blind pairwise comparison, they see the source image, the target emotion, and two randomly ordered results, and judge which conveys the target emotion better and which edit is more suitable overall.

Tab. R3. (a) Planning quality vs. Qwen3-VL-32B. (b) U-AIM Performance vs. FireRed-Edit (Tie: no preference).

(a) PlannerRecommend Acc. (B)State GroundingExecutabilityInstruction Coherence
Qwen3-VL-32B43.18%70.33%41.00%72.00%
EmoWeave75.00%78.67%81.33%74.67%
(b) MethodHuman EmoACCTarget-emotion preferenceOverall editing preference
EmoWeave0.690062%58%
FireRed-Edit0.580028%30%
Tie—10%12%

Q2. How much of EmoWeave's performance comes from EGRA and U-AIM rather than from the strong models used to build the supervision?

Thank you for this constructive question. EGRA and U-AIM improve results beyond what DeepSeek-V4-Pro and FireRed-Edit provide. We examine the planning and execution sides in turn.

Planning side (DeepSeek-V4-Pro). DeepSeek-V4-Pro is text-only and writes instructions from textual state descriptions during data construction, so it cannot serve as an image-based planner baseline. 1) With the same DeepSeek generator, adding EGRA raises the top-ranked rate of the supervision from 22.9% to 39.7% under Gemini-3.1-Pro and from 14.8% to 46.6% under GPT-5.5 (Fig. 6, L428–431). 2) Planners trained on data built with EGRA recommend the target emotion more accurately and score higher in planning quality than planners trained on data built without EGRA (Tab. R4). 3) The planner is also substantially higher in planning quality than zero-shot planning by other multimodal models, including our base model (Tab. 3, L399–415). Together, these show the improvement comes from EGRA, not DeepSeek.

Tab. R4. Ablation study of EGRA.

Training dataRecommend Acc. (B)Overall Quality (B)Overall Quality (A)
Without EGRA0.67420.76930.7621
With EGRA0.72360.81210.7952

Execution side (FireRed-Edit). FireRed-Edit only generates the target images of the semantic training pairs, while the appearance pairs come from real photographers' retouching. EmoWeave is a unified model in which planning is as important as execution, so the fair comparison is on the whole task: with FireRed-Edit performing the end-to-end task, it trails EmoWeave on all four metrics (Tab. R5).

Tab. R5. End-to-end results on 200 samples (the same samples as Tab. R2(a) in our [W] reply).

MethodEmoACCEmo-SSemanticAppearance
FireRed-Edit0.54000.05985.99306.1038
EmoWeave0.65000.13546.55006.7600

Q3. Do Capability-Aware Asymmetric Training (CAST) and EGRA apply to other unified multimodal model (UMM) architectures and model scales?

Thank you for this valuable question. EGRA and CAST are not tied to our backbone. EGRA builds its supervision independently of the model, and CAST only needs a unified model whose understanding and generation parameters can be updated separately. We further apply these two strategies to another UMM, BAGEL, and the results validate their effectiveness as they improve every metric over the BAGEL base model (Tab. R6). Planning quality is measured on the full benchmark and U-AIM performance on the 200 samples of Tab. R2(a) in our [W] reply.

Tab. R6. Experiments on BAGEL. (a) Ablation study of EGRA on planning quality. (b) Ablation study on U-AIM performance.

(a) SettingRecommend Acc. (B)Overall Quality (B)Overall Quality (A)
BAGEL (base model)0.53110.54130.5503
+ EGRA0.65940.69960.6812
+ EGRA & CAST0.68170.74210.7264
(b) SettingEmoACCEmo-SSemanticAppearance
BAGEL (base model)0.53000.05175.84504.8100
+ EGRA0.56500.07345.97254.9875
+ CAST0.58500.08976.07755.6625
+ EGRA & CAST0.60500.10486.21255.8725

I appreciate the authors’ detailed response and additional experiments, which have alleviated some of my concerns. I am inclined to maintain my positive assessment.

We’re glad that we could address your concerns. Your feedback has clearly helped strengthen the paper. We truly appreciate it, thank you!

Reviewer_E9j9OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript presents a unified framework for affective image manipulation that integrates emotional intent inference, editing planning, and image generation. It introduces Unified Affective Image Manipulation (U-AIM) to support both emotion enhancement and migration. Also, it proposes Emotion-Grounded Retrieval Augmentation (EGRA) to construct coordinated semantic and appearance editing instructions. The authors demonstrate improvements in planning quality and affective editing performance over existing methods. The methodological novelty is limited, as the framework primarily integrates established retrieval-augmented instruction generation, multimodal fine-tuning, and instruction-conditioned image editing techniques. The distinct technical contribution beyond combining these existing components is not sufficiently established.

Critical strengths
  • The paper addresses a relevant problem in affective image manipulation by integrating emotional intent inference, semantic editing, and appearance editing within a unified framework. The proposed formulation provides a structured approach to both emotion enhancement and migration.
Critical weaknesses
  • The conceptual novelty of U-AIM is limited. Existing affective image manipulation approaches have already explored emotion-conditioned editing, semantic and appearance modifications, and planning-based editing frameworks. EmoAgent has investigated emotion-aware planning and editing through knowledge-guided strategies and tool execution. Although U-AIM integrates these capabilities into a unified formulation, the manuscript does not sufficiently establish a fundamentally new problem setting or methodological distinction from prior approaches.
  • The methodological contribution of EGRA is relatively incremental. Its core procedure combines existing emotion-annotated datasets, textual visual-state representations, embedding-based retrieval, clustering, and LLM-based instruction generation. Although this provides structured supervision for affective editing, these components are largely established techniques, and their integration does not introduce a substantially new retrieval or instruction-generation mechanism.
  • The framework builds upon the existing SenseNova-U1.5 backbone and applies sequential planner and executor fine-tuning with selective parameter freezing. Moreover, semantic and appearance editing are learned from separate editing pairs, while their joint execution relies on the backbone's existing ability to follow combined instructions. The manuscript does not clearly establish a novel mechanism for coordinating these two editing operations beyond instruction generation and conventional fine-tuning.
Questions for authors
  • The main claimed contribution is the unified coordination of semantic and appearance editing. However, the executor is trained in separate semantic, and appearance editing pairs and jointly executes their instructions at inference. What specific mechanism enables coordination beyond the pretrained backbone's existing instruction-following capabilities, and how does this fundamentally differ from conventional instruction-conditioned image editing?
2 comments
AuthorsOpenReview ↗

[W] Response to E9j9 Weaknesses

W1. How does U-AIM differ from existing affective image manipulation methods, including EmoAgent?

Thank you for this question. U-AIM differs in two ways: it supports two types of manipulation (with or without a specified target emotion) and requires coherent editing.

  1. Two manipulation scenarios. U-AIM supports both Scenario A (User-Specified) and Scenario B (Emotion Inferred), where the system infers a suitable target emotion. EmoAgent takes a user-specified target emotion (Introduction, L45).

  2. Coherent editing. U-AIM requires the specified or recommended emotion to translate into semantic and appearance edits that are emotionally consistent and logically coherent, and we enforce this at the task, data, and method levels. EmoAgent verifies each edit against the target emotion, but states no consistency requirement across its tools' edits.

W2. What does EGRA contribute beyond its individually standard components?

Thank you for this question. We clarify a possible misunderstanding: the contribution of EGRA is not retrieval itself, but converting real retouching edits into evidence with measured emotional effect, and generating semantic and appearance instructions jointly from this evidence and emotion knowledge.

  1. Linking emotion knowledge with real editing experience. EmoVerse links emotions to visual content but has no edits; AetherRetouch has real edits but no emotion labels. VER measures the emotion before and after each retouching edit, turning retouching pairs into editing records with measured emotion change (Appendix, L709). Retrieval by source-image state and target emotion returns emotional evidence (what supports the emotion) and operational evidence (how similar images were edited toward it).

  2. Joint construction under one target emotion. One reasoning model writes both instructions from both kinds of evidence, so the two edits share one emotion direction and fit each other.

With the same DeepSeek generator, adding EGRA raises the share of cases ranking first among four supervision strategies from 22.9% to 39.7% (Gemini-3.1-Pro) and from 14.8% to 46.6% (GPT-5.5) (Fig. 6, L428–431). Two further ablations (Tab. R1–R2) show that (a) generating the two instructions jointly improves all four instruction metrics, most on Instruction Coherence (+0.082); and (b) planners trained on data built with EGRA recommend the target emotion more accurately in Scenario B (+0.049) and improve Overall Quality in both Scenario B (+0.043) and Scenario A (+0.033).

Tab. R1. Joint vs. separate instruction generation.

GenerationState GroundingExecutabilityInstruction CoherenceOverall Quality
Separate0.77150.81180.69820.7630
Joint (EGRA)0.79360.84070.78000.8029

Tab. R2. Ablation study of EGRA.

Training dataRecommend Acc. (B)Overall Quality (B)Overall Quality (A)
Without EGRA0.67420.76930.7621
With EGRA0.72360.81210.7952

W3. How are semantic and appearance edits coordinated, and does it go beyond the pretrained model and conventional fine-tuning?

Thank you for this question. Coordination in EmoWeave is designed into the supervision, not into a dedicated training mechanism. EGRA maps one target emotion to content and appearance edits built to fit each other, so training on it teaches coordinated planning. The unified multimodal model (UMM) backbone can already execute both edit types together to some extent, which is why we chose it; its zero-shot plans are less grounded and less executable (Tab. 3, L399–415). The executor is trained per edit type because the data sources are heterogeneous (real photographers' retouching and semantic editing), with both types mixed in each batch; Capability-Aware Asymmetric Training (CAST) updates non-overlapping parameters for the planner and executor.

Tab. R3 varies executor training and execution on the same plans (200 samples; extends Tab. 4, L449–450).

ExecutorEmoACCEmo-SSemanticAppearance
Pretrained0.62000.09016.40006.6400
Semantic only0.63500.11376.47006.5750
Appearance only0.63000.10276.42506.6575
Both (EmoWeave)0.65000.13546.55006.7600
Both, two passes0.61000.08316.49506.6075

Training on both types is best on all four metrics, while semantic-only training lowers Appearance. Executing the same plans in two passes falls below even the pretrained executor on EmoACC, Emo-S, and Appearance, consistent with the two instructions being planned as one set.

The Question is answered in [Q] Response to E9j9 Questions. We sincerely hope our replies resolve your concerns, and would appreciate your reconsideration of the score.

AuthorsOpenReview ↗

[Q] Response to E9j9 Questions

Q. How is coordination achieved beyond the pretrained model, and how does it differ from conventional instruction-based editing?

Thank you for this question. We would like to clarify that coordination is learned from direct supervision: each UAIM-14K sample pairs a source image and a target emotion with a semantic and an appearance instruction generated together (Sec. 3).

  1. Beyond the pretrained model. Training on this supervision is what lifts planning beyond the pretrained model: our planner outperforms zero-shot planning by other large models, including our base model, in planning quality (Tab. 3, L399–415). The executor is trained on single-type pairs, mixed within each batch, because the data sources are heterogeneous; it then carries out the planned set in one pass, building on the UMM's ability to execute both edit types together. Given the same plans, training on both types improves on the pretrained executor on all four metrics (EmoACC 0.6200 → 0.6500, Emo-S 0.0901 → 0.1354; Tab. R3 in [W] Response to E9j9 Weaknesses). Executing the plans in two passes instead of one lowers all four metrics (EmoACC 0.6500 → 0.6100, Emo-S 0.1354 → 0.0831), consistent with the two instructions being planned as one set and losing effect when split.

Generating the two instructions separately rather than jointly, with the executor unchanged, tests more directly whether joint generation is what coordinates the edits. Joint generation improves State Grounding (0.7715 → 0.7936), Executability (0.8118 → 0.8407), Instruction Coherence (0.6982 → 0.7800), and Overall Quality (0.7630 → 0.8029) (Tab. R1 in [W] Response to E9j9 Weaknesses).

  1. Difference from instruction-based editing. At execution, EmoWeave performs instruction-based editing. The difference lies in the task: conventional instruction-based editing executes an instruction supplied by the user, whereas in U-AIM the model itself infers or follows a target emotion, plans coordinated semantic and appearance instructions from it, and produces the edited image.

We sincerely hope these clarifications and results address your concerns, and we would greatly appreciate your reconsideration of the score.

Reviewer_q5XhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes EmoWeave, a unified affective image manipulation (U-AIM) model, that connects emotional intent inference, coordinated semantic and appearance planning, and image editing that supports emotion enhancement and migration. Their contributions include the model itself, a retrieval-augmented supervision construction (EGRA), and a dataset with benchmark (UAIM-14K and UAIMBench, respectively) for training and evaluation of both the planning process and final image edits.

Critical strengths

(1) Empirical Improvement & Method Unification. This work overcomes weakness in previous works by combining emotion inference, editing planning, and semantic and appearance editing in the same unified framework. Experimental results show gain over previous work. Additionally, this work provides all the components of the pipeline including the model, data, and benchmark, which will ease future benchmarking works.

(2) The ablation studies have been well considered to establish the effectiveness of the different components of the proposed framework and their contribution to the empirical performance.

Critical weaknesses

(1) Ambiguity in emotion inference. The emotion inference component seems harder to evaluate as ground truth might be subjective. Thus, an LLM as a judge setup seems insufficient to validate.

(2) Ablation studies show relatively smaller gains for executor & planner training. Given the ambiguity of evaluating emotion, there needs to be more analysis of whether this difference justifies the two stage training. Further, there is a need to show the effect of training for semantic edits versus appearance edits to establish the utility of using them together. In short, are the gains because there are more types of image training or because there is this specific type of image training?

(3) The construction of the dataset and benchmark is arguably a more significant contribution than the proposed model itself. While EmoWeave primarily adapts an existing unified multimodal architecture through task-specific supervision, the dataset and benchmark address an important gap in affective image manipulation by explicitly connecting emotional intent, coordinated editing instructions, and visual outcomes. However, the scale of the benchmark would need to be considered for future scalability and generalizability.

Questions for authors

[Clarification] How can emotion be inferred for emotion migration when the desired emotion might not be present in the source image?

3 comments
AuthorsOpenReview ↗

[Q] Response to q5Xh Questions

Q. When the desired emotion may not exist in the source image, how is the emotion for emotion migration inferred?

Thank you for this question. In Scenario B (Emotion Inferred), the inferred target is either an emotion already present in the image or a suitable emotion for a neutral image, rather than a purposeless random choice. This follows from how we construct the training data. For migration in Scenario B, we use only emotionally neutral source images, where the training target guides the recommended direction of emotion editing based on visual content; the remaining emotion-inference samples are emotion enhancement, where the target is the emotion the image already carries (main text, L220; Appendix, Fig. 8(c)). Trained on such data, which draws on photographers' own retouching, the planner learns emotional preferences closer to those of human photographers.

We will revise the wording around L220 in the main text to improve readability.

AuthorsOpenReview ↗

[W] Response to q5Xh Weaknesses

W1. Is the evaluation of subjective emotion inference reliable beyond an LLM judge?

Thank you for this point. We use multi-level cross-validation: GPT and Gemini as two verification models (L298), and EEmo-Logic and EmoCaliber as two VER models for emotion evaluation (Appendix, Tab. 11). On top of this, we ran a human evaluation of planner outputs on 100 benchmark samples with 3 annotators each (Tab. R1). Recommend Acc. (B) is computed on the 44 Scenario B samples, and annotator agreement is 81.34% with Fleiss' κ ≈ 0.558.

Tab. R1. Human evaluation of planner outputs.

MethodRecommend Acc. (B)State GroundingExecutabilityInstruction Coherence
Qwen3-VL-32B0.43180.70330.41000.7200
EmoWeave0.75000.78670.81330.7467

EmoWeave's planner is rated higher on all four aspects, markedly on recommendation accuracy and Executability. The end-to-end human evaluation is in our [W – Supplement] post.

W2. Do the gains come from our specific training rather than from more image training?

Thank you for this question. U-AIM aims to express the target emotion while making the image better, through coordinated content and appearance edits. A gain on emotion alone would therefore not meet the goal of the task, so we judge each training choice on emotion attainment (EmoACC, Emo-S) together with Semantic and Appearance quality. Tab. R2 reports all four, extending the ablation of Tab. 4 (L449–450) to 200 samples.

Tab. R2. Training ablation and unified vs. two-pass execution (200 samples).

VariantEmoACCEmo-SSemanticAppearance
Full Model0.65000.13546.55006.7600
w/o App.Exe Training0.63500.11376.47006.5750
w/o Sem.Exe Training0.63000.10276.42506.6575
w/o Executor Training0.62000.09016.40006.6400
w/o Planner & Executor Training0.60000.10236.40006.5675
Full Model, two-pass execution0.61000.08316.49506.6075

(a) Gains on all four metrics. The Full Model is best on every metric; the absolute gains are modest, but they appear on emotion, Semantic, and Appearance together.

(b) The type of executor training matters. Training on semantic edits only (w/o App.Exe) improves the emotion metrics over the untrained executor but lowers Appearance: emotion improves at the cost of image quality, which U-AIM aims to avoid. Training on both types improves Semantic and Appearance more than either type alone.

(c) Joint execution. Executing the two instructions in two passes instead of one lowers all four metrics (last row), so joint execution adds value beyond having both types of training data.

(d) Planner. Our trained planner is effective in State Grounding and Executability, and recommends more suitable target emotions in Scenario B, compared with its base model and other general models (Tab. 3, L399–415).

W3. The dataset and benchmark address an important gap; what about the benchmark's scale and generalizability?

We thank the reviewer for recognizing the value of our dataset and benchmark. The coverage and construction of UAIMBench are described in Appendix A.3, and we clarify them here. The benchmark covers diverse image content (portraits, multi-person images, natural scenery, cities, animals, still life, macro scenes) and all six emotions with a balanced distribution (Fig. 3, L282). Its construction pipeline is extensible: reference emotions are read by VER from the retouched pairs or taken from EmoVerse labels, migration targets follow a fixed sampling rule, and images are stratified-sampled, so more held-out images can be added under the same rules. Building on this, we doubled the end-to-end subset from 100 to 200 samples (A 56 → 111, B 44 → 89), and EmoWeave keeps its lead over all baselines on all four metrics, consistent with the original 100-sample results (Tab. R3), so the conclusions hold on a larger evaluation set.

Tab. R3. End-to-end results on 200 samples. EmoEdit and EmoStyle need a user-specified target, so they use the 111 Scenario A samples.

MethodnEmoACCEmo-SSemanticAppearance
EmoWeave2000.65000.13546.55006.7600
Qwen-Image-Edit-25112000.55000.04885.90506.2600
FireRed-Edit2000.54000.05985.99306.1038
BAGEL-Edit2000.53000.05175.84504.8100
VeraRetouch2000.55000.02374.97805.5930
EmoWeave111 (A-only)0.63960.21106.73876.9324
EmoEdit111 (A-only)0.47750.07224.79705.4140
EmoStyle111 (A-only)0.3694−0.03074.21203.8420

The Question is answered in a separate comment, [Q] Response to q5Xh Questions.

AuthorsOpenReview ↗

[W – Supplement] Response to q5Xh Weaknesses

W1 (supplement). Is the evaluation of subjective emotion inference reliable beyond an LLM judge?

We completed the end-to-end human evaluation of the edited images, and EmoWeave leads under it as well. It uses 100 benchmark samples with 3 annotators each and has two steps. (1) Without a target hint, annotators see only the edited image and choose the dominant emotion from Ekman-6 plus Neutral; Human EmoACC is the share of images whose chosen emotion matches the target. (2) In a blind pairwise comparison, annotators see the source image, the target emotion, and two randomly ordered results, and judge which conveys the target emotion better and which edit is more suitable overall.

Tab. R4. Human evaluation on U-AIM performance (100 samples; Tie: no preference).

MethodHuman EmoACCTarget-emotion preferenceOverall editing preference
EmoWeave0.690062%58%
FireRed-Edit0.580028%30%
Tie—10%12%

Together with the planner evaluation (Tab. R1 in our [W] reply), human judgments now cover both stages, planning and editing.

Thank you for your valuable feedback. We sincerely hope our replies resolve your concerns, and would appreciate your reconsideration of the score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Contributions, Common Concerns, and Guide to Our Replies

We thank all reviewers for their careful reading and constructive feedback. Reviewer q5Xh highlights the empirical improvement and method unification and notes that the model, data, and evaluation we provide will facilitate future research; Reviewer rVKc finds the work comprehensive and practically meaningful; and Reviewer E9j9 agrees that it addresses an important problem. Across all three reviews, the sub-scores are consistent: Presentation 3 (clear and easy to follow), Informativeness 2 (learned something interesting or useful), and Soundness 2 (only minor or fixable concerns).

Contributions & Innovations

  • New Unified Task. U-AIM, which infers or follows the emotional intent and plans and executes coordinated content and appearance edits toward it, covering emotion enhancement and migration.
  • Multi-Dimensional Data. UAIM-14K, built with EGRA, links source images, emotional intents, editing instructions, and results, and UAIMBench evaluates planning and end-to-end editing.
  • End-to-End Unified Model. EmoWeave, which performs emotion reasoning, instruction generation, and instruction-based editing in a single multimodal model.

Together, they provide all components of a pipeline for the new task (q5Xh), validated by comprehensive experiments (rVKc).

Common Concerns

  • Source of the gains: coordinated planning versus training data or teacher models (E9j9, q5Xh, rVKc).
  • Reliability of emotion evaluation beyond VER and VLM judges (q5Xh, rVKc).
  • Scale of the end-to-end evaluation (rVKc, q5Xh).

Guide to Our Replies

All additional experiments and clarifications are included in our replies:

  • Coordination and sources of gain: the executor ablation in Tab. 4, the semantic/appearance training ablation, unified versus two-pass execution, joint versus separate instruction generation, planners trained on data constructed with versus without EGRA, zero-shot planning of other large models including our base model in Tab. 3, the end-to-end comparison with FireRed-Edit, and the application of EGRA and CAST to BAGEL (E9j9-W2, E9j9-W3, E9j9-Q, q5Xh-W2, rVKc-Q2, rVKc-Q3).
  • Evaluation: human evaluation of planner outputs and of the edited images, and the end-to-end benchmark extended to 200 samples (q5Xh-W1, q5Xh-W3, rVKc-W1, rVKc-W2, rVKc-Q1).
  • Clarifications: difference from prior AIM methods including EmoAgent (E9j9-W1), EGRA design (E9j9-W2), and construction of Scenarios A and B (q5Xh-Q1, rVKc-W1).

Replies are organized by reviewer, with one post for the weaknesses [W] and one for the questions [Q], each answering the items in order. A revised PDF with changes marked in blue will be uploaded before the discussion ends.

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

Related papers

Discussion (0)

Sign in to comment.