Diverse Intersectional Visual Evaluation (DIVE)

• DIVE is a multimodal dataset designed to capture intersectional perceptions of harm and safety in text-to-image models using rigorously stratified demographic annotations.
• It compiles over 35,000 harm judgments from approximately 650 raters across 30 unique demographic intersections to assess explicit, violent, and bias-related safety concerns.
• Its innovative multi-objective alignment methodology leverages quantitative metrics and subgroup-specific signals to enhance model safety, steerability, and fairness.

Diverse Intersectional Visual Evaluation (DIVE) is a large-scale, multimodal dataset designed to facilitate the pluralistic alignment of text-to-image (T2I) models to diverse and intersectional human value systems, focusing on the nuanced and sometimes conflicting perceptions of safety, harm, and offensiveness. The DIVE dataset advances methodologies for deep alignment by explicitly addressing the underrepresentation of diverse demographics and leveraging quantitative, intersectional, and replicated human judgments over a challenging prompt set. By capturing structured feedback from a rigorously sampled, demographically stratified pool of annotators, DIVE enables both granular evaluation and targeted conditioning of T2I safety mechanisms (Rastogi et al., 15 Jul 2025).

1. Dataset Construction and Demographic Stratification

DIVE comprises 1000 prompt–image (PI) pairs selected from the Adversarial Nibbler challenge set, each representing adversarial or edge-case content for T2I harm assessment. These pairs were annotated through 31,980 usable harm-evaluation responses, resulting in 35,164 individual harm judgments—including direct Likert responses, “unsure” votes, and sub-questions—generated by a nominal pool of approximately 650 raters, with 617 retained after rigorous quality filtering for attention, response time, and self-consistency.

Intersectional stratification is achieved via a 2 (Gender) × 3 (Age) × 5 (Ethnicity) trisection, resulting in 30 unique demographic intersections, each with 23–25 independent raters. The group definitions are as follows:

Dimension	Values
Gender	Men, Women
Age	Gen Z (18–27), Millennial (28–43), Gen X (44+)
Ethnicity	White, Black, Latinx, South Asian, East Asian

Safety-violation types span explicit (sexual, n=300), violent (graphic violence, n=250), and bias (hate/stereotype, n=450) categories. Each type is subdivided into 12 balanced thematic topics, such as Religion, Gender, Age, and Political speech, with most violation–topic cells containing 50 PI pairs (except Violent/Other: 100).

2. Annotation Protocol and Harm Measurement Taxonomy

All raters receive an illustrated tutorial clarifying the tripartite harm taxonomy: Explicit (sexual), Violent (graphic violence), and Bias (hate/stereotype), with safe/unsafe examples for context and warnings about potential distress. Annotators interact with the PI pairs through a secure interface, with all stimuli hidden by default and visible only upon explicit request.

Multi-dimensional annotation for each PI pair involves:

1. Self-harm Likert: “How harmful do you find this?”—5-point scale (0 = Not at all harmful … 4 = Completely harmful), with “Unsure” available.
2. Others-harm Likert: “How harmful would others find this?”—same scale and options.
3. Harm typology: Multiple-choice on why the stimulus is harmful or offensive (Bias, Explicit, Violent, Other with free text).
4. Free-form explanation: Annotators can submit qualitative rationale.

“Unsure” selections require justification: poor image quality, lack of cultural context, or free-form (Other). Annotators are provided psychological resources and break recommendations.

3. Statistical Assessment: Reliability and Demographic Divergence

3.1 Inter-Rater Reliability

DIVE measures overall inter-rater reliability for harm scores using Krippendorff’s $\alpha$, yielding a global $\alpha \approx 0.24$. This dispersion is consistent with highly subjective evaluative tasks. The mathematical formulation is:

$$\alpha = 1 - \frac{\sum_{c} \sum_{i<j} w_{ij} \, p_{ij}}{\sum_{c} \sum_{i<j} w_{ij} \, e_{ij}}$$

where $p_{ij}$ and $e_{ij}$ represent observed and expected disagreement, with quadratic weights $w_{ij}$.

A Group Association Index (GAI) is defined to quantify intra-group reliability versus out-of-group reliability:

$GAI(G) = \frac{IRR(G)}{XRR(G, \neg G)}$

where $IRR(G)$ is in-group Krippendorff $\alpha$ and $XRR$ is cross-group reliability.

3.2 Statistical Testing of Demographic Effects

Demographic effect sizes are established using Mann-Whitney U (Gender) and Kruskal-Wallis H (Age, Ethnicity) tests, with all contrasts evaluated at $\alpha \approx 0.24$0 and stratified by topic confounds. Illustrative results include:

• $\alpha \approx 0.24$1, $\alpha \approx 0.24$2
• $\alpha \approx 0.24$3, $\alpha \approx 0.24$4
• $\alpha \approx 0.24$5, $\alpha \approx 0.24$6

Each statistic is stratified and prevalence-weighted to ensure topic-demographic balance.

4. Analysis of Intersectional Harm Perception

4.1 Self-Report Versus Perception of Harm to Others

Self-harm ratings are typically lower in aggregate than “harm to others” estimates, yet demographic disparities are consistent across both dimensions. For example, women’s mean Likert responses are approximately 0.2 higher than men’s on the 0–4 scale for both self- and others-harm, indicating robust sensitivity to demographic influence.

4.2 Intra-Group Reliability and Cohesion

Intersectional groups show greater internal coherence compared to single-axis groups. For example:

Intersection	GAI	Significance
Gen Z–Black	1.38	**
Millennial–Black	1.30	**
White–Woman	1.14	*
South Asian–Woman	1.15	*

($\alpha \approx 0.24$7: $\alpha \approx 0.24$8; $\alpha \approx 0.24$9: $$\alpha = 1 - \frac{\sum_{c} \sum_{i<j} w_{ij} \, p_{ij}}{\sum_{c} \sum_{i<j} w_{ij} \, e_{ij}}$$0 post–BH correction)

Single-dimension GAIs range narrowly (Gender: 0.98–1.04, Ethnicity: 0.98–1.12, Age: 0.97–1.05), underscoring added value of intersectional analysis.

4.3 Groupwise Divergences via Content-Type Simulation

Simulated subgroup splits (repeated 100×) track divergences in case-flagging. Women identified 9–14 more unsafe items than men across violation types. South Asian, Latinx, and Black raters each flagged 24–28 “Bias” cases missed by other raters; White raters missed 17 non-White-flagged Bias cases. Gen Z flagged more bias violations than Millennials or Gen X.

5. Pluralistic Alignment and Model Supervision Methodologies

5.1 Multi-Objective Alignment Loss

DIVE’s detailed, group-conditioned scores enable a multi-objective alignment criterion for T2I model $$\alpha = 1 - \frac{\sum_{c} \sum_{i<j} w_{ij} \, p_{ij}}{\sum_{c} \sum_{i<j} w_{ij} \, e_{ij}}$$1:

$$\alpha = 1 - \frac{\sum_{c} \sum_{i<j} w_{ij} \, p_{ij}}{\sum_{c} \sum_{i<j} w_{ij} \, e_{ij}}$$2

where:

• $$\alpha = 1 - \frac{\sum_{c} \sum_{i<j} w_{ij} \, p_{ij}}{\sum_{c} \sum_{i<j} w_{ij} \, e_{ij}}$$3 is a conventional diffusion loss;
• $$\alpha = 1 - \frac{\sum_{c} \sum_{i<j} w_{ij} \, p_{ij}}{\sum_{c} \sum_{i<j} w_{ij} \, e_{ij}}$$4 is a safety-score head conditioned on group $$\alpha = 1 - \frac{\sum_{c} \sum_{i<j} w_{ij} \, p_{ij}}{\sum_{c} \sum_{i<j} w_{ij} \, e_{ij}}$$5;
• $$\alpha = 1 - \frac{\sum_{c} \sum_{i<j} w_{ij} \, p_{ij}}{\sum_{c} \sum_{i<j} w_{ij} \, e_{ij}}$$6 is the group-mean human DIVE score for PI $$\alpha = 1 - \frac{\sum_{c} \sum_{i<j} w_{ij} \, p_{ij}}{\sum_{c} \sum_{i<j} w_{ij} \, e_{ij}}$$7;
• $$\alpha = 1 - \frac{\sum_{c} \sum_{i<j} w_{ij} \, p_{ij}}{\sum_{c} \sum_{i<j} w_{ij} \, e_{ij}}$$8 are group weights (e.g., uniform or demography-weighted);
• $$\alpha = 1 - \frac{\sum_{c} \sum_{i<j} w_{ij} \, p_{ij}}{\sum_{c} \sum_{i<j} w_{ij} \, e_{ij}}$$9 is a fidelity loss (e.g., CLIP, LPIPS).

5.2 Application in Training, Evaluation, and LLM-Judge Distillation

• Supervised Safety Filter: Use DIVE group scores as targets to learn demographic-conditional safety filters; optionally, as RL rewards for policy-gradient algorithms.
• Subgroup Evaluation: Report safety accuracy and subgroup AUCs per demographic; analyze false negatives per group, especially those highlighted in simulated splits.
• LLM Rater Distillation: Fine-tune vision-LLaMA models (e.g., Gemma) to approximate DIVE group scores at scale for unannotated data.

6. Data Collection Procedures and Annotation Best Practices

DIVE’s 2 × 3 × 5 trisectional recruiter paradigm is budget-efficient for scaling intersectional data, consistently delivering 20–30 replicates per intersection with 23–25 unique raters per group for robust plurality estimates and direct estimation of within-group variance $p_{ij}$0.

Multiple annotation modalities (Likert, categorical, free-form) are maintained for both explainability and robustness. Extensive instructions and welfare mechanisms (breaks, support resources, pre-exposure tutorial) mitigate psychological stress and maximize data quality.

7. Extrapolation, Model Steerability, and Future Research

Zero-shot LLM raters (e.g., Gemma 4B) achieve poor agreement with human subgroup consensus (Kendall’s $p_{ij}$1). Explicitly prompting with demographic context increases $p_{ij}$2 to $p_{ij}$3, averaged across 30 groups. Few-shot exemplar promptings offer marginal further gains. The data suggests that fine-tuning LLM judges directly on DIVE or equipping them with demographic adapters are promising directions for enabling scalable, subgroup-aware evaluation and safety enforcement.

In summary, DIVE is the first dataset to provide large-scale, highly replicated, and demographically deep multimodal safety annotations for T2I model alignment. It enables systematic subgroup-level evaluation, reveals substantive context-dependent divergences in harm perception, and informs both supervised and reinforcement-based pluralistic alignment strategies, supporting the empirical foundation that the critical question for safety in generative models is not solely “what is safe,” but “safe for whom?” (Rastogi et al., 15 Jul 2025).