Pith. sign in

REVIEW 5 major objections 9 minor 57 references

EmoScene is a 1.2M-image dual-space dataset that makes emotional tone a controllable, continuous dial in text-to-image generation.

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

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

T0 review · deepseek-v4-flash

2026-08-04 05:33 UTC pith:47ZEAXN7

load-bearing objection Large, careful dataset; the abstract's controllability numbers are nowhere in the paper, so treat them as unverified. the 5 major comments →

arxiv 2604.00933 v2 pith:47ZEAXN7 submitted 2026-04-01 cs.CV

EmoScene: A Dual-space Dataset for Controllable Affective Image Generation

Li He , Longtai Zhang , Wenqiang Zhang , Yan Wang , Lizhe Qi This is my paper
classification cs.CV
keywords affective image generationvisual emotion datasetvalence-arousal-dominanceperceptual attributestext-to-image diffusioncontrollable generationhuman-in-the-loop annotationdual-space emotion representation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Text-to-image models can render scenes but not reliably render how the scene should feel, because emotion words in a prompt leave the visual carriers of affect unspecified. This paper tries to close that gap by building EmoScene, a 1.2M-image dataset in which every image carries discrete emotion labels, continuous valence–arousal–dominance scores, measurable perceptual attributes such as color, luminance, curvature, and complexity, and contextual captions. It then shows that a small modulation network, AffectCtrl, can inject these labels into a frozen diffusion generator as a residual control vector, shifting emotional tone continuously while preserving scene content. If the dataset labels are trustworthy, this makes emotional tone a tunable slider rather than a vague textual modifier, and gives the field a reusable resource for studying how emotion and perception interact.

Core claim

EmoScene claims to be a large-scale dual-space visual-emotion dataset that jointly encodes affective dimensions (eight discrete emotions plus continuous VAD) and perceptual attributes (color statistics, luminance, saturation, curvature, visual complexity) with contextual semantics, across 1.2M images and 345 scene categories. The paper argues that this unified representation is what allows fine-grained affective control: a frozen diffusion generator can be steered by a learned modulation vector that perturbs cross-attention, yielding 85.75% categorical emotion accuracy and Pearson correlations of 0.673–0.765 across five continuous control axes. The authors further report that after controlli

What carries the argument

The load-bearing object is the dual-space annotation: an affective space (discrete emotion plus VAD) tied to a perceptual space (HSV color statistics, luminance, curvature, edge density, texture complexity), with scene captions grounding both. The control mechanism is AffectCtrl's Affective–Perceptual Modulation Network, which encodes emotion, VAD, and HSV into a compact modulation vector via a query transformer, then injects the vector as learned offsets to the keys and values of selected cross-attention layers in a frozen diffusion model. This lets continuous affective targets translate into structured changes in lighting, color temperature, and contrast while leaving the base generator un

Load-bearing premise

The paper's central claims rest on the assumption that the 30,519 audited images (2.54% of the dataset) are representative of the remaining ~1.17M images, so the automated labels on those unseen images are as accurate as the labels humans checked.

What would settle it

Take a random sample stratified by all 345 scene categories from the unaudited portion of EmoScene, have independent human annotators label discrete emotion and VAD using the same protocol, and compare. If agreement falls substantially below the reported 91.22% or VAD error rises, the dataset's reliability claim fails; alternatively, re-running AffectCtrl with VAD targets shuffled while prompts are fixed should break the Pearson correlations if the control signal is real.

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

If this is right

  • Affective tone becomes continuously controllable: valence, arousal, dominance, and color can be dialed without retraining the base generator.
  • A reusable 1.2M-image resource with aligned affective, perceptual, and semantic labels becomes available for training and benchmarking emotion-aware systems.
  • The reported stable cross-source associations between emotion and perceptual statistics can serve as priors for generation and for psychology-style analyses of natural images.
  • An interface-aligned evaluation protocol allows future affect-control methods to be compared fairly on shared control channels.
  • If label quality holds, downstream tasks such as emotion-based visual question answering and retrieval gain a large-scale dual-space testbed.

Where Pith is reading between the lines

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

  • I would infer that the residual-conditioning design is not limited to color: the same mechanism could plausibly extend to other perceptual axes (depth, texture, composition) once datasets annotate them, making the approach a general recipe for perceptually grounded control.
  • The paper's own finding that dominance correlates weakly with low-level features suggests that dominance control will need to be conditioned on scene semantics rather than pixels; one testable extension is adding scene-category embeddings to the modulation vector and measuring dominance alignment.
  • Because the paper's results show the HSV branch produces only mild changes, a natural extension is a stronger perceptual branch with region-level color annotations to make color-emotion coupling more precise.
  • If the unaudited labels prove reliable, EmoScene could become a widely used benchmark for affective generation, but that hinges on an audit stratified by scene category, which the paper does not provide.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 9 minor

Summary. The paper introduces EmoScene, a large-scale dual-space affective image dataset with 1.2M images across 345 scene categories. Each image is annotated with discrete emotion labels (8 categories + neutral), continuous VAD scores, perceptual attributes (HSV color statistics, curvature, complexity), and contextual text descriptions. The annotation pipeline combines multiple MLLMs (Qwen2.5, InternVL3-8B) with human-in-the-loop quality control; a 30,519-image random audit yields 91.22% agreement and a multi-rater Fleiss' kappa of 0.85. The authors also present AffectCtrl/APMN, a lightweight modulation baseline built on frozen PixArt-α with cross-attention injection of a learned affective-perceptual vector, controlled by emotion label, scene, VAD, and HSV values. The abstract claims AffectCtrl achieves 85.75% categorical emotion accuracy, outperforms EmotiCrafter on valence/arousal control under a shared protocol, and obtains Pearson correlations of 0.673–0.765 across five continuous axes. The main text reports only CLIPScore, VAD MAE, and HSV color MAE comparisons against the unmodulated PixArt-α.

Significance. If the dataset claims hold, EmoScene would be a substantial resource: it is nearly an order of magnitude larger than EmoSet (118k), spans diverse scene categories, and uniquely combines discrete emotions, continuous VAD, and structured perceptual attributes. The explicit dual-space annotation design and human-in-the-loop quality control are valuable methodological contributions. The dataset's perceived strengths include: the external audit statistics (91.22% agreement, κ=0.85), the inclusion of bias-mitigation stress tests, transparency in the annotation pipeline, and the intention to release code and annotations. However, the paper's central benchmark claim for the AffectCtrl baseline is not supported by the manuscript as written: the headline quantitative results in the abstract are absent from the body, and the appendix explicitly documents a limitation of the HSV control branch. The paper's contribution as a dataset paper is somewhat separable from the baseline claims, but the controllability evidence is too weak to establish the utility of the dual-space supervision for generation.

major comments (5)
  1. [Abstract vs. Sections 4–5 and Appendix F.3] The abstract claims AffectCtrl achieves 85.75% categorical emotion accuracy, outperforms EmotiCrafter on valence/arousal under a shared protocol, and obtains Pearson correlations of 0.673–0.765 across five continuous axes. These numbers appear nowhere in Sections 4–5, Table 3, or the appendix. Instead, Table 3 reports only CLIPScore, VAD MAE, and HSV color MAE against the unmodulated frozen baseline. Appendix E.3 defines an interface-aligned protocol for cross-method comparison but reports no cross-method results. Since these missing numbers are the primary evidence that dual-space supervision yields controllable generation, the central benchmarking claim is unverifiable as written. The authors should either add the experiment reporting these metrics and the EmotiCrafter comparison, or remove/qualify the abstract claims.
  2. [Appendix F.3, Fig. 21] The paper's own supplementary material states that varying HSV descriptors 'produces only very mild shifts in global color tone' and that the four outputs in Fig. 21 are 'almost visually identical,' calling the color branch 'a weak global tone adjustment rather than a strong control knob.' This directly contradicts the abstract's implication of successful continuous brightness/saturation control (Pearson correlations 0.673–0.765). The quantitative HSV MAE improvements in Table 3 are also minimal (e.g., Color MAE (S) 0.299 to 0.202). This limitation should be acknowledged in the main text and the abstract's controllability claim should be narrowed.
  3. [Section 5.2 / Table 3; Appendix E.2] The VAD evaluation uses a frozen VAD predictor chosen to be 'different from the MLLMs used for dataset annotation,' but the identity, architecture, and calibration of this evaluator to human judgment are never provided. The baseline is trained on EmoScene's MLLM-generated VAD labels and evaluated by an unknown automatic predictor; requested-vs-predicted MAE in Table 3 could reflect evaluator agreement with the training label distribution rather than true perceived affect. The authors should identify the evaluator, report its human-correlation, and show example predicted scores.
  4. [Appendix B.4] Section 5.1 says all experiments are on EmoScene, but Appendix B.4 states 'We do not ship predefined train/val/test subfolders; users may create splits as needed.' No train/test split is otherwise defined in the paper. Without a clear split or release of split indices, the Table 3 numbers are not reproducible and the risk of train/test leakage exists. Please specify the split procedure or provide split files.
  5. [Section 3.3 / Table 2] The audit covers 30,519 images (2.54% of the dataset) plus a 297-image bias test. The 91.22% agreement and κ=0.85 are credited to MLLM labels, but the paper does not state that the audit is a random sample of the full 1.2M across all 345 scene categories, nor does it stratify agreement by scene category, data source, or emotion class. Without evidence that the audited subset is representative, the label quality of the remaining ~97.5% remains unverified. This is not fatal for a dataset paper, but a stratified audit or per-category agreement table would strengthen the reliability claim.
minor comments (9)
  1. [Abstract] The abstract in the submission header differs from the arXiv abstract: the header version omits the 85.75% accuracy and Pearson correlation numbers and says the baseline serves as a 'reproducible probe.' The inconsistency between the two abstracts should be resolved.
  2. [Table 2] The row 'Multi-rater' reports VAD correlation r = 0.960/0.984/0.901 but the table caption calls it 'agreement' and the text says Fleiss' κ=0.85. Specify what r refers to (Pearson between human raters? between MLLM and humans?). Similarly, VAD MSE values are listed without a scale; state that VAD is on 1–9.
  3. [Table 3] VAD MAE (D) for PixArt-α is reported as 2.005 ± 1.416, identical to VAD MAE (A). This looks like a copy-paste error; please verify the dominance baseline.
  4. [Section 4.2 / Eq. (1)] The Q-Former formulation in Eq. (1) is vague: the output δ is said to aggregate conditions, but the dimensions and pooling mechanism are not specified. The appendix defines losses on δ but not its exact architecture; provide a diagram with tensor dimensions.
  5. [Appendix D, Eqs. (4)–(20)] The total loss in Eq. (4)/(5) contains many hyperparameters (λ_img, λ_eff, λ_pair, λ_dir, λ_con, α, β, γ, η, w_h, w_s, w_v, λ_mag, λ_inj, m, mfar, mnear, etc.). The paper says 'weight schedules' are in the supplementary but no exact values or schedules are given. Please include the numeric values for all loss weights in a table.
  6. [Section 5.2 / Fig. 8] The qualitative claim in Fig. 8 that APMV produces 'a colder and heavier atmosphere' is from a single example. It would help to include quantitative per-condition statistics (e.g., mean brightness, saturation) across the 600 generated samples per setting to support the observation of systematic shifts.
  7. [Appendix E.2] The VAD evaluator is described as 'different from the MLLMs used for dataset annotation,' but the name is not given. Merely being different does not guarantee non-circularity; the evaluator should be identified and ideally shown to have human-level agreement.
  8. [Appendix B.1, Step 6] The description generation step says HAM and MLLM jointly generate descriptions and that five randomly sampled descriptions are produced per image. It is unclear which description is used for training the baseline or reporting CLIPScore; specify which of the five is in the JSON and used in experiments.
  9. [General] There are minor typos and inconsistent notation: 'm1ore' in Section 3.2, 'Valenceseparates' in Section 3.2, and 'Emoset' vs 'EmoSet' in Table 1. Also, some references to 'Fig. 21' include 'Tab. 2' in the appendix caption but it should be Table 3.

Circularity Check

0 steps flagged

No significant circularity found: the main control evaluation uses a frozen evaluator explicitly chosen to differ from the annotating MLLMs, and no load-bearing self-citation chain is present.

full rationale

I examined the claimed derivation chain: EmoScene labels are produced by two MLLMs (Qwen2.5 and InternVL3-8B) with human-in-the-loop audit; AffectCtrl is trained on those labels; evaluation uses a frozen VAD predictor that the paper explicitly says is chosen to be different from the annotating MLLMs, plus a deterministic HSV extractor. The abstract's specific numbers (85.75% categorical accuracy, EmotiCrafter outperformance, Pearson 0.673–0.765) are absent from the main text and Appendix F.3 admits that HSV-based control is visually weak, but absence/contradiction is a correctness or evidence problem, not a circularity. No equation was found in which a reported prediction reduces by construction to a fitted parameter: the alignment losses supervise the modulation vector's geometry (e.g., L_VAD-align), not the evaluator, and Appendix E.2 explicitly states an anti-circularity choice of evaluator. There is no load-bearing self-citation, imported uniqueness theorem, or ansatz smuggled in via citation. Therefore, under the stated rules for flagging circularity, the paper does not exhibit circular derivation.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The APMV is a learned control vector inside the model, not a postulated entity with independent falsifiable handles. The central claims rest on unverified assumptions about label generalization and the validity of an unnamed automatic evaluator.

free parameters (3)
  • Loss weights in Eq. (4)/(5): λ_img, λ_eff, λ_pair, λ_dir, λ_con, α, β, γ, η, w_h, w_s, w_v, λ_mag, λ_inj, m, mfar, mnear
    Chosen by hand/tuning; exact values said to be in supplementary but not present in the provided text. Controllability results depend on them.
  • VAD aggregation weights
    Step 8: final VAD is a weighted average of Qwen2.5 and InternVL3-8B predictions; weights are not specified.
  • Quality filtering thresholds (aesthetic/sharpness/CLIP)
    Steps 2–3: cutoffs for image filtering are not given; they determine the final 1.2M composition.
axioms (4)
  • domain assumption Automated MLLM annotations on the un-audited 97.5% of images are as reliable as the audited subset.
    Section 3.3 audits 2.54% + 297 images; remaining images accepted without per-image human verification.
  • domain assumption The frozen VAD evaluator provides a valid, bias-free proxy for human affective judgment.
    Appendix E.2: evaluator is unnamed and not calibrated to human ratings.
  • domain assumption Eight discrete emotions plus neutral and continuous VAD adequately span affective expression for this task.
    Section 3.1 uses this vocabulary from the annotation prompt (Fig. 11).
  • domain assumption Perceptual attributes (HSV statistics, curvature, entropy, edge density) are the relevant low-level carriers of affective expression.
    Section 3.2 and Steps 9–10; this is the premise of the dual-space design.

pith-pipeline@v1.3.0-alltime-deepseek · 23299 in / 15857 out tokens · 165578 ms · 2026-08-04T05:33:41.630127+00:00 · methodology

0 comments
read the original abstract

Text-to-image diffusion models achieve high visual fidelity, yet fine-grained affective control remains difficult because textual emotion cues often fail to specify the visual perceptual factors underlying affective expression. Existing visual-affect datasets are likewise often limited to discrete labels, specific domains, or limited supervision of perceptual attributes. We introduce EmoScene, a large-scale dual-space dataset for controllable affective image generation, containing 1.2M images across more than 300 scene categories. Its affective space jointly represents discrete emotions and continuous valence--arousal--dominance (VAD), its perceptual space records measurable appearance attributes, and contextual descriptions ground both in scene semantics. EmoScene combines multi-model annotation with human-in-the-loop quality control. A random audit of 30,519 images yields 91.22\% agreement on discrete emotion labels, while an independent multi-rater evaluation yields Fleiss' $\kappa=0.85$. After controlling for source and scene composition, affective dimensions and perceptual attributes exhibit stable associations across data sources, reflecting statistical tendencies rather than deterministic visual rules. To demonstrate the dataset's utility, we further develop AffectCtrl, which learns residuals in the conditioning space of frozen diffusion models to support categorical emotion generation and continuous control over VAD, brightness, and saturation. AffectCtrl achieves 85.75\% categorical emotion accuracy, outperforms EmotiCrafter in valence and arousal control under a shared evaluation protocol, and obtains Pearson correlations of 0.673--0.765 across all five continuous axes. These results demonstrate that EmoScene provides a scalable data foundation for analyzing and controlling affective expression in visual generation.

Figures

Figures reproduced from arXiv: 2604.00933 by Li He, Lizhe Qi, Longtai Zhang, Wenqiang Zhang, Yan Wang.

Figure 1
Figure 1. Figure 1: Multi-space attributes in the EmoScene Dataset: The proposed EmoScene dataset provides a dual-space emotion representation, jointly modeling affective, perceptual, and contex￾tual dimensions.The total count of annotated attributes is shown in ✄ ✂ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dual-Space Annotation Pipeline. The pipeline proceeds from Step 1 to Step 11. Step 1 collects raw images from open photographic, artistic platforms. Step 2 filters low￾quality samples using aesthetic and sharpness scores, and Step 3 verifies or corrects scene labels and image–text consistency. Steps 4 to 5 detect human subjects and objects and record their attributes and interactions. Step 6 generates shor… view at source ↗
Figure 3
Figure 3. Figure 3: Representative examples of EmoScene’s dual-space annotations [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Structure of the affective space represented in the Valence–Arousal plane. Different emotions occupy distinct yet continuous regions, reflecting the organization of the affective space. shown in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Perceptual color composition across emotion categories. Top: proportions of achromatic and chromatic components.Bottom: distributions over eight chromatic hues [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Interplay between affective and perceptual spaces. Left:relationships between VAD and HSV color statistics. Right: correlations between emotion dimensions and perceptual structural features. priors for controllable affective image generation. Interplay between Affective and Perceptual Spaces. To quantify the link between emotion and low-level appearance, we compute Pearson correlations between VAD and HSV … view at source ↗
Figure 7
Figure 7. Figure 7: Overview of the proposed simple dual-space modulation baseline. Given a text prompt and continuous affective–perceptual controls derived from EmoScene (e.g., VAD scores and HSV), APMN first encodes the affective–perceptual signals using a separate condition encoder. The resulting condition embeddings are combined with text embeddings in a Q-Former module to generate the Affective–Perceptual Modulation Vect… view at source ↗
Figure 8
Figure 8. Figure 8: Effect of APMV modulation. Under identical prompts and seeds, the baseline PixArt￾α output (left) largely keeps a neutral mood, whereas our APMV-modulated PixArt-α (right) produces a colder and heavier atmosphere that is more aligned with the target negative emotion, while maintaining structural fidelity. given our focus on controllability with a fixed backbone. Additional quantitative ablations and extend… view at source ↗
Figure 9
Figure 9. Figure 9: Continuous Control along Valence and Arousal. Image generations under different VA settings. Higher valence (left to right) generally leads to brighter and less oppressive scenes, while higher arousal (top to bottom) increases contrast and perceived energy. The trajectories illustrate that the APMV-modulated model responds smoothly to changes in affective targets. the VA dimensions. For a fixed prompt and … view at source ↗
Figure 10
Figure 10. Figure 10: Motivation for our work. (a)Limitations of prompt-based discrete control.Simply rewriting the prompt, such as changing happy to extremely happy, rarely leads to a clear change in how the image feels. (b)Dataset-grounded continuous control. In contrast, EmoScene provides VAD and perceptual labels that we turn into simple control sliders for a lightweight modulation baseline on a frozen generator, enabling … view at source ↗
Figure 11
Figure 11. Figure 11: Prompt used for CoT-style VAD and discrete emotion annotation. [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Human-in-the-loop verification UI. Left: the current image. Right: machine an￾notations to be verified: two discrete-emotion suggestions (e.g., from different annotators/models) and three V/A/D scores. The reviewer only selects Yes/No for each field. Selecting No requires a brief rationale in the comment box and moves the sample to the re-check queue; Yes accepts the value as final. ments are made along w… view at source ↗
Figure 13
Figure 13. Figure 13: Pairing example. In the beach folder, the image (.jpg) and its annotation (.json) share the same <stem>; the bottom block shows a truncated JSON corresponding to the image. We will distribute the dataset in an annotation-first manner (image file + paired JSON metadata) and only release subsets that satisfy the corresponding licensing constraints. Human subjects and privacy. EmoScene contains diverse scene… view at source ↗
Figure 14
Figure 14. Figure 14: Valence–Dominance distributions across different discrete emotions. Each panel shows the per-emotion density on the VD plane (higher intensity = higher density) [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Arousal–Dominance distributions across different discrete emotions. Each panel shows the per-emotion density on the AD plane. C Extended Statistics and Analyses C.1 Distribution on VD and AD Planes by Discrete Emotions We also visualize the joint distributions on Valence–Dominance(VD) and Arousal–Dominance(AD). These per-emotion density maps reveal how perceived dominance co-varies with valence and arousa… view at source ↗
Figure 16
Figure 16. Figure 16: Per–emotion HSV statistics. It clarifies how color carriers co-vary with affect cate￾gories by displaying the central tendency and spread of HSV for each emotion. categories show higher saturation; dominance shows weaker coupling to low–level color than va￾lence/arousal. We include these summaries to aid reproducibility and to serve as priors for color￾aware modulation in our dual-space controls. C.3 Visu… view at source ↗
Figure 17
Figure 17. Figure 17: Visual proofs of bias mitigation in EmoScene. We showcase eight representative counter-intuitive samples from our dataset, complete with their VAD scores and annotation ratio￾nales. The top row includes scenes with exceptionally low luminance that correctly receive positive affective labels (high valence; e.g., fireworks, starry sky, candle-lit dinner). The bottom row displays scenes with bright lighting … view at source ↗
Figure 18
Figure 18. Figure 18: Interactive demo. Dataset annotations (Emotion, VAD, HSV) are exposed as simple controls for continuous affect–perception navigation; sampling settings are kept explicit for repro￾ducibility. 33 [PITH_FULL_IMAGE:figures/full_fig_p033_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Effect of APMV modulation under identical text prompts. Left: forest scene. Right: mountain scene. For each scene, the left image is generated by the baseline PixArt-α model without affective modulation, and the right image is generated by our APMV-modulated PixArt-α. APMV mainly alters global lighting, color temperature, and contrast while preserving the underlying scene structure. to low and high valenc… view at source ↗
Figure 20
Figure 20. Figure 20: Varying valence and arousal across two scenes. We fix dominance to a neutral value (D = 5.0) and sample four points on the VA plane for two scenes using the same text prompt and random seed. Left: urban street viewed from a crosswalk. Right: stone-paved town square with a central fountain. Rows correspond to low and high valence (V = 2 and V = 8), and columns to low and high arousal (A = 2 and A = 8). As … view at source ↗
Figure 21
Figure 21. Figure 21: HSV-based color control yields only subtle changes. We fix the text prompt and VAD values for a forest trail scene and vary the HSV color descriptors while injecting APMV. Rows correspond to lower and higher saturation, and columns to lower and higher value (brightness). The four outputs are almost visually identical, indicating that our current HSV-based design provides only weak control over global tone… view at source ↗
Figure 22
Figure 22. Figure 22: EmoScene Dataset Gallery (Part 1/3). Intra-class diversity for Amusement, Anger, and Awe. The samples span a wide range of social interactions, urban conflicts, and vast natural landscapes. Notice how the dominance score dynamically varies with the semantic context (e.g., remarkably low dominance in Awe versus varied dominance in Anger ). 37 [PITH_FULL_IMAGE:figures/full_fig_p037_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: EmoScene Dataset Gallery (Part 2/3). Intra-class diversity for Contentment, Disgust, and Excitement. These variations highlight the dataset’s ability to ground distinct ac￾tivation levels and valences across entirely different scenes—from the absolute tranquility and low arousal of Contentment to the peak dynamic energy of Excitement and the multifaceted environ￾mental revulsion of Disgust. 38 [PITH_FULL… view at source ↗
Figure 24
Figure 24. Figure 24: EmoScene Dataset Gallery (Part 3/3). Intra-class diversity for Fear, Sadness, and Neutral. While Fear and Sadness capture intensely negative affective states with varying arousal, the Neutral category serves as a crucial baseline. Characterized by mid-range valence, arousal, and dominance (VAD ≈ 5/5/5), these neutral samples ground the dataset in everyday, purely functional, and emotionally detached seman… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 9 linked inside Pith

  1. [1]

    Univ of California Press, 1954

    Rudolf Arnheim.Art and visual perception: A psychology of the creative eye. Univ of California Press, 1954

    1954

  2. [2]

    Univ of California Press, 2023

    Rudolf Arnheim.Visual thinking. Univ of California Press, 2023

    2023

  3. [3]

    A comprehensive survey on affective computing: Challenges, trends, applications, and future directions.IEEE access, 12:96150–96168, 2024

    Sitara Afzal, Haseeb Ali Khan, Md Jalil Piran, and Jong Weon Lee. A comprehensive survey on affective computing: Challenges, trends, applications, and future directions.IEEE access, 12:96150–96168, 2024

    2024

  4. [4]

    A systematic review on affective computing: Emotion models, databases, and recent advances.Information Fusion, 83:19–52, 2022

    Yan Wang, Wei Song, Wei Tao, Antonio Liotta, Dawei Yang, Xinlei Li, Shuyong Gao, Yixuan Sun, Weifeng Ge, Wei Zhang, et al. A systematic review on affective computing: Emotion models, databases, and recent advances.Information Fusion, 83:19–52, 2022

    2022

  5. [5]

    Emotions in the loop: A survey of affective computing for emotional support.arXiv preprint arXiv:2505.01542, 2025

    Karishma Hegde and Hemadri Jayalath. Emotions in the loop: A survey of affective computing for emotional support.arXiv preprint arXiv:2505.01542, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    A review of human emotion synthesis based on generative technology

    Fei Ma, Yifan Xie, Yukan Li, Ying He, Yi Zhang, Hongwei Ren, Zhou Liu, Wei Yao, Fuji Ren, Fei Richard Yu, et al. A review of human emotion synthesis based on generative technology. IEEE Transactions on Affective Computing, 2025

    2025

  7. [7]

    Gpt-4v with emotion: A zero-shot benchmark for generalized emotion recognition

    Zheng Lian, Licai Sun, Haiyang Sun, Kang Chen, Zhuofan Wen, Hao Gu, Bin Liu, and Jian- hua Tao. Gpt-4v with emotion: A zero-shot benchmark for generalized emotion recognition. Information Fusion, 108:102367, 2024

    2024

  8. [8]

    Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. InICLR, 2024

    2024

  9. [9]

    ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers.arXiv preprint arXiv:2211.01324, 2022

    Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers.arXiv preprint arXiv:2211.01324, 2022

    Pith/arXiv arXiv 2022

  10. [10]

    Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    2022

  11. [11]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  12. [12]

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

    Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischinski, Danny Cohen-Or, and Hui Huang. Emoset: A large-scale visual emotion dataset with rich attributes. InICCV, pages 20383–20394, 2023. 14

    2023

  13. [13]

    Emogen: Emotional image content generation with text-to-image diffusion models

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

    2024

  14. [14]

    Emoti- crafter: Text-to-emotional-image generation based on valence-arousal model.arXiv preprint arXiv:2501.05710, 2025

    Shengqi Dang, Yi He, Long Ling, Ziqing Qian, Nanxuan Zhao, and Nan Cao. Emoti- crafter: Text-to-emotional-image generation based on valence-arousal model.arXiv preprint arXiv:2501.05710, 2025

    arXiv 2025

  15. [15]

    Emoprefer: Can large language models understand human emotion preferences?arXiv preprint arXiv:2507.04278, 2025

    Zheng Lian, Licai Sun, Lan Chen, Haoyu Chen, Zebang Cheng, Fan Zhang, Ziyu Jia, Ziyang Ma, Fei Ma, Xiaojiang Peng, et al. Emoprefer: Can large language models understand human emotion preferences?arXiv preprint arXiv:2507.04278, 2025

    arXiv 2025

  16. [16]

    Com- prehensive exploration of diffusion models in image generation: a survey.Artificial Intelligence Review, 58(4):99, 2025

    Hang Chen, Qian Xiang, Jiaxin Hu, Meilin Ye, Chao Yu, Hao Cheng, and Lei Zhang. Com- prehensive exploration of diffusion models in image generation: a survey.Artificial Intelligence Review, 58(4):99, 2025

    2025

  17. [17]

    Affective computing in the era of large language models: A survey from the nlp perspective.arXiv preprint arXiv:2408.04638, 2024

    Yiqun Zhang, Xiaocui Yang, Xingle Xu, Zeran Gao, Yijie Huang, Shiyi Mu, Shi Feng, Daling Wang, Yifei Zhang, Kaisong Song, et al. Affective computing in the era of large language models: A survey from the nlp perspective.arXiv preprint arXiv:2408.04638, 2024

    Pith/arXiv arXiv 2024

  18. [19]

    A mixed bag of emotions: Model, predict, and transfer emotion distributions

    Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 860–868, 2015

    2015

  19. [20]

    Building a large scale dataset for image emotion recognition: The fine print and the benchmark

    Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. Building a large scale dataset for image emotion recognition: The fine print and the benchmark. InProceedings of the AAAI conference on artificial intelligence, volume 30, 2016

    2016

  20. [21]

    Predicting person- alized image emotion perceptions in social networks.IEEE transactions on affective computing, 9(4):526–540, 2016

    Sicheng Zhao, Hongxun Yao, Yue Gao, Guiguang Ding, and Tat-Seng Chua. Predicting person- alized image emotion perceptions in social networks.IEEE transactions on affective computing, 9(4):526–540, 2016

    2016

  21. [22]

    Emotic: Emotions in context dataset

    Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Emotic: Emotions in context dataset. InProceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 61–69, 2017

    2017

  22. [23]

    Affectnet: A database for facial expression, valence, and arousal computing in the wild.IEEE Transactions on Affective Computing, 10(1):18–31, 2017

    Ali Mollahosseini, Behzad Hasani, and Mohammad H Mahoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild.IEEE Transactions on Affective Computing, 10(1):18–31, 2017

    2017

  23. [24]

    Introducing the open affective stan- dardized image set (oasis).Behavior research methods, 49(2):457–470, 2017

    Benedek Kurdi, Shayn Lozano, and Mahzarin R Banaji. Introducing the open affective stan- dardized image set (oasis).Behavior research methods, 49(2):457–470, 2017

    2017

  24. [25]

    Artemis: Affectivelanguageforvisualart

    Panos Achlioptas, Maks Ovsjanikov, Kilichbek Haydarov, Mohamed Elhoseiny, and Leonidas J Guibas. Artemis: Affectivelanguageforvisualart. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11569–11579, 2021. 15

    2021

  25. [26]

    Emotion recognition for multiple context awareness

    Dingkang Yang, Shuai Huang, Shunli Wang, Yang Liu, Peng Zhai, Liuzhen Su, Mingcheng Li, and Lihua Zhang. Emotion recognition for multiple context awareness. InEuropean conference on computer vision, pages 144–162. Springer, 2022

    2022

  26. [27]

    Findingemo: An image dataset for emotion recognition in the wild.Advances in Neural Infor- mation Processing Systems, 37:4956–4996, 2024

    Laurent Mertens, Elahe Yargholi, Hans Op de Beeck, Jan Van den Stock, and Joost Vennekens. Findingemo: An image dataset for emotion recognition in the wild.Advances in Neural Infor- mation Processing Systems, 37:4956–4996, 2024

    2024

  27. [28]

    Affective image classification using features inspired by psychology and art theory

    Jana Machajdik and Allan Hanbury. Affective image classification using features inspired by psychology and art theory. InProceedings of the 18th ACM international conference on Multimedia, pages 83–92, 2010

    2010

  28. [29]

    Ferv39k: A large-scale multi-scene dataset for facial expression recognition in videos

    Yan Wang, Yixuan Sun, Yiwen Huang, Zhongying Liu, Shuyong Gao, Wei Zhang, Weifeng Ge, and Wenqiang Zhang. Ferv39k: A large-scale multi-scene dataset for facial expression recognition in videos. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20922–20931, 2022

    2022

  29. [30]

    Ctnet: Conversational transformer network for emotion recognition.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:985– 1000, 2021

    Zheng Lian, Bin Liu, and Jianhua Tao. Ctnet: Conversational transformer network for emotion recognition.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:985– 1000, 2021

    2021

  30. [31]

    Emotion-llama: Multimodal emotion recognition and rea- soningwithinstructiontuning.Advances in Neural Information Processing Systems, 37:110805– 110853, 2024

    Zebang Cheng, Zhi-Qi Cheng, Jun-Yan He, Kai Wang, Yuxiang Lin, Zheng Lian, Xiaojiang Peng, and Alexander Hauptmann. Emotion-llama: Multimodal emotion recognition and rea- soningwithinstructiontuning.Advances in Neural Information Processing Systems, 37:110805– 110853, 2024

    2024

  31. [32]

    Multimodal cross-and self-attention network for speech emotion recognition

    Licai Sun, Bin Liu, Jianhua Tao, and Zheng Lian. Multimodal cross-and self-attention network for speech emotion recognition. InICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4275–4279. IEEE, 2021

    2021

  32. [33]

    Dpcnet: Dual path multi-excitation collaborative network for facial expression representation learning in videos

    Yan Wang, Yixuan Sun, Wei Song, Shuyong Gao, Yiwen Huang, Zhaoyu Chen, Weifeng Ge, and Wenqiang Zhang. Dpcnet: Dual path multi-excitation collaborative network for facial expression representation learning in videos. InProceedings of the 30th ACM international conference on multimedia, pages 101–110, 2022

    2022

  33. [34]

    Emotion recognition from multiple modalities: Fundamentals and methodologies.IEEE Signal Processing Magazine, 38(6):59–73, 2021

    Sicheng Zhao, Guoli Jia, Jufeng Yang, Guiguang Ding, and Kurt Keutzer. Emotion recognition from multiple modalities: Fundamentals and methodologies.IEEE Signal Processing Magazine, 38(6):59–73, 2021

    2021

  34. [35]

    Merbench: A unified evaluation benchmark for multimodal emotion recognition.arXiv preprint arXiv:2401.03429, 2024

    Zheng Lian, Licai Sun, Yong Ren, Hao Gu, Haiyang Sun, Lan Chen, Bin Liu, and Jianhua Tao. Merbench: A unified evaluation benchmark for multimodal emotion recognition.arXiv preprint arXiv:2401.03429, 2024

    Pith/arXiv arXiv 2024

  35. [36]

    Emoe: Modality- specific enhanced dynamic emotion experts

    Yiyang Fang, Wenke Huang, Guancheng Wan, Kehua Su, and Mang Ye. Emoe: Modality- specific enhanced dynamic emotion experts. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 14314–14324, 2025

    2025

  36. [37]

    Knowledge-based emotion recognition using large language models

    Bin Han, Cleo Yau, Su Lei, and Jonathan Gratch. Knowledge-based emotion recognition using large language models. In2024 12th International Conference on Affective Computing and Intelligent Interaction (ACII), pages 1–9. IEEE, 2024. 16

    2024

  37. [38]

    Towards robust multimodal emotion recognition under missing modalities and distribution shifts.arXiv preprint arXiv:2506.10452, 2025

    Guowei Zhong, Ruohong Huan, Mingzhen Wu, Ronghua Liang, and Peng Chen. Towards robust multimodal emotion recognition under missing modalities and distribution shifts.arXiv preprint arXiv:2506.10452, 2025

    Pith/arXiv arXiv 2025

  38. [39]

    Multimodal emotion recognition and sentiment analysis in multi-party conversation contexts.arXiv preprint arXiv:2503.06805, 2025

    Aref Farhadipour, Hossein Ranjbar, Masoumeh Chapariniya, Teodora Vukovic, Sarah Ebling, and Volker Dellwo. Multimodal emotion recognition and sentiment analysis in multi-party conversation contexts.arXiv preprint arXiv:2503.06805, 2025

    Pith/arXiv arXiv 2025

  39. [40]

    Affective image editing: Shaping emotional factors via text descriptions.arXiv preprint arXiv:2505.18699, 2025

    Peixuan Zhang, Shuchen Weng, Chengxuan Zhu, Binghao Tang, Zijian Jia, Si Li, and Boxin Shi. Affective image editing: Shaping emotional factors via text descriptions.arXiv preprint arXiv:2505.18699, 2025

    arXiv 2025

  40. [41]

    A unified and interpretable emotion representation and expression generation

    Reni Paskaleva, Mykyta Holubakha, Andela Ilic, Saman Motamed, Luc Van Gool, and Danda Paudel. A unified and interpretable emotion representation and expression generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2447–2456, 2024

    2024

  41. [42]

    Schuller, and Kurt Keutzer

    Sicheng Zhao, Guiguang Ding, Qingming Huang, Tat-Seng Chua, Björn W. Schuller, and Kurt Keutzer. Affective image content analysis: A comprehensive survey. InProceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 5534–5541. International Joint Conferences on Artificial Intelligence Organization, 7 2018

    2018

  42. [43]

    A survey on facial expression recognition of static and dynamic emotions, 2024

    Yan Wang, Shaoqi Yan, Yang Liu, Wei Song, Jing Liu, Yang Chang, Xinji Mai, Xiping Hu, Wenqiang Zhang, and Zhongxue Gan. A survey on facial expression recognition of static and dynamic emotions, 2024

    2024

  43. [44]

    Affectgan: Affect-based gen- erative art driven by semantics

    Theodoros Galanos, Antonios Liapis, and Georgios N Yannakakis. Affectgan: Affect-based gen- erative art driven by semantics. In2021 9th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW), pages 01–07. IEEE, 2021

    2021

  44. [45]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. InProceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023

    2023

  45. [46]

    Emoedit: Evoking emotions through image manipulation

    Jingyuan Yang, Jiawei Feng, Weibin Luo, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Emoedit: Evoking emotions through image manipulation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 24690–24699, 2025

    2025

  46. [47]

    Emoctrl: Controllable emotional image content generation.arXiv preprint arXiv:2512.22437, 2025

    Jingyuan Yang, Weibin Luo, and Hui Huang. Emoctrl: Controllable emotional image content generation.arXiv preprint arXiv:2512.22437, 2025

    Pith/arXiv arXiv 2025

  47. [48]

    Emospace: Fine-grained emotion prototype learning for immersive affective content generation.arXiv preprint arXiv:2602.11658, 2026

    Bingyuan Wang, Xingbei Chen, Zongyang Qiu, Linping Yuan, and Zeyu Wang. Emospace: Fine-grained emotion prototype learning for immersive affective content generation.arXiv preprint arXiv:2602.11658, 2026

    Pith/arXiv arXiv 2026

  48. [49]

    Attrictrl: Fine-grained control of aesthetic attribute intensity in diffusion models.arXiv preprint arXiv:2508.02151, 2025

    Die Chen, Zhongjie Duan, Zhiwen Li, Cen Chen, Daoyuan Chen, Yaliang Li, and Yingda Chen. Attrictrl: Fine-grained control of aesthetic attribute intensity in diffusion models.arXiv preprint arXiv:2508.02151, 2025

    arXiv 2025

  49. [50]

    Jrdb-social: A multi- faceted robotic dataset for understanding of context and dynamics of human interactions within social groups

    Simindokht Jahangard, Zhixi Cai, Shiki Wen, and Hamid Rezatofighi. Jrdb-social: A multi- faceted robotic dataset for understanding of context and dynamics of human interactions within social groups. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22087–22097, 2024. 17

    2024

  50. [51]

    Mod- eling thousands of human annotators for generalizable text-to-image person re-identification

    Jiayu Jiang, Changxing Ding, Wentao Tan, Junhong Wang, Jin Tao, and Xiangmin Xu. Mod- eling thousands of human annotators for generalizable text-to-image person re-identification. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 9220–9230, 2025

    2025

  51. [52]

    comforting,

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479, 2025. 18 Appendix for EmoScene: A Dual-space Dataset for Controllable Affective Image Generation A Motiva...

    Pith/arXiv arXiv 2025

  52. [53]

    Scene Description: Briefly describe the main content and scenario in the image

  53. [54]

    Determine whether it appears active or passive, powerful or weak

    Subject Analysis: Analyze the state, posture, expression, and interaction with the environment of the main subject(s) (person, animal, or object). Determine whether it appears active or passive, powerful or weak. Part Two: Three Dimensions and Discrete Emotion Assessment (Text and Final Scores Only) You need to synthesize all visual cues in the image and ...

  54. [55]

    Valence (Pleasure): 1=Extremely Negative, 5=Neutral, 9=Extremely Pleasant

  55. [56]

    Arousal (Activation): 1=Extremely Calm, 5=Normal, 9=Extremely Exciting

  56. [57]

    Dominance (Control): 1=Completely Powerless, 5=Neutral, 9=Complete Control

  57. [58]

    arousal inconsistent with calm scene

    Primary Emotion: amusement, anger, awe, contentment, disgust, excitement, fear, sadness. Use neutral if no significant emotion is present. Use unknown if it cannot be determined. Scoring Rationale (VAD): Based on your analysis, explain why you gave these three specific VAD scores. Final Output Format Requirement: Please strictly output your analysis resul...

    2081