How well can Text-to-Image Generative Models understand Ethical Natural Language Interventions?

1 Introduction

• Neutral prompts can produce images biased toward white males, while such bias may amplify dataset bias and harm underrepresented groups.The paper frames the question across gender, skin color, and culture.Source: §1 · p. 1
• The authors test whether equitable-judgment knowledge added to the text prompt can diversify outputs without giving visual cues toward a particular group.Their example adds that all individuals can be a lawyer irrespective of gender.Source: §1 · p. 2
• ENTIGEN evaluates intervention effects across gender, skin color, and culture using minDALL·E, DALL·E-mini, and Stable Diffusion with CLIP and human judgments.The paper treats gender as gender expression rather than identity and avoids inferring racial associations from appearance.Source: §1 · pp. 1–2
• A few ethical interventions diversify generated groups while preserving image quality, and some cases flip bias toward originally underrepresented groups.Phrases such as “irrespective of gender” and “culture” tend to trigger large changes in generations.Source: §1 · p. 2

2.1 ENTIGEN Benchmark Construction

• ENTIGEN pairs neutral prompts with ethical interventions across gender, skin color, and cultural wedding scenarios, while excluding interventions unrelated to those axes from expected diversification.Its interventions are designed not to give away visual cues; the benchmark contains 246 prompts.Source: §2.1 · p. 2

2.2 Image Generation.

• Each prompt generates nine images from each of three models, repeated nine times: minDALL·E, DALL·E-mini, and Stable Diffusion.The models were selected for efficient generation of high-quality images.Source: §2.2 · pp. 2–3

2.3 Evaluation Metrics.

• Diversity is measured from the gap between group counts, while a normalized difference score captures bias direction and magnitude across axes and prompt conditions.Smaller diversity scores indicate more diverse outputs; uncertain gender or skin-color judgments are excluded, and normalization permits comparison across prompts with different group counts.Source: §2.3 · p. 3
• Human annotators evaluate professions and objects, while CLIP evaluates the full image set; CLIP is used only for gender because its agreement with humans is stronger there.Agreement is 78–85% for gender recognition and 67–78% for skin-color recognition.Source: §2.3 · p. 3

3.1 CLIP-based Results

• Gender-specific interventions increase CLIP-based gender diversity for all models, with ‘irrespective of their gender’ improving scores more than the simpler instruction.Skin-color interventions do not significantly affect gender diversity, while irrelevant interventions can still change scores.Source: §3.1 · p. 3, Table 1 · p. 4

3.2 Human Evaluation Results

• Human evaluations find better diversity scores with axis-specific ‘irrespective of’ instructions, while irrelevant instructions usually do not improve scores.The reported comparison covers gender and skin-color interventions for professions and objects.Source: §3.2 · pp. 3–4
• Cultural interventions make generations from all models more diverse, whereas irrelevant instructions do not influence cultural diversity.Source: §3.2 · p. 4
• Ethical interventions may flip bias direction, including more men and dark-skinned individuals in DALL·E-mini makeup images and more women than men in Stable Diffusion police images.Source: §3.2 · p. 4
• Stable Diffusion sometimes generates multiple people whose gender or skin color annotators cannot assign, making diversity estimates from scores alone unreliable.Source: §3.2 · p. 4

3.3 Quality of Image Generation

• Compared with original prompts, ethical interventions slightly reduced the number of good-quality generations by 0–1.5 images per attribute, except for DALL·E-mini and Stable Diffusion in the profession category.Source: §3.3 · p. 4

Gender phrase analysis

• Removing ‘irrespective of their gender’ restores the original gender-skewed generations, although why the intervention can produce anti-stereotype images remains unresolved.In pretraining captions, ‘irrespective of’ appears 37 times in equitable contexts, but the full phrase ‘irrespective of their gender’ appears once; the authors note that the effect warrants further exploration.Source: §4 · pp. 4–5

Cultural phrase analysis

• The authors report that caption co-occurrences of ‘culture’ with ‘Western’ (75), ‘Indian’ (394), and ‘Chinese’ (322) help explain generations depicting these Non-Western cultures after cultural interventions.Source: §4 · p. 5

• ENTIGEN is a framework and associated dataset for evaluating changes in text-to-image generation diversity when ethical interventions are added.Source: §5 · p. 5
• Without fine-tuning, models can generate diverse groups with ethical interventions, and preliminary analyses identify ‘irrespective of gender’ and ‘culture’ as keyphrases associated with large changes.Source: §5 · p. 5

Limitations

• Ethical interventions may not reliably produce diverse outputs, and the study does not claim they reduce bias in practice.Its stated aim is to analyze current systems’ responses to interventions.Source: Limitations · pp. 5–6
• Creating many ethical interventions and evaluating them with people remains a current limitation and an important future direction.Source: Limitations · p. 6
• Human evaluations may reflect perceptual biases of annotators from the US and UK, while automatic evaluations may reflect CLIP’s bias.The authors plan to involve annotators from diverse regions and less biased vision models.Source: Limitations · p. 6

Ethics Statement

• The analysis uses selected categories and binary man/woman and light-/dark-skinned groupings, assigning gender from perceived appearance rather than identity and examining generated images, not real people.Source: Ethics Statement · p. 6
• Skin color is also reduced to light- or dark-skinned labels, with annotators relying on their own perception rather than the Fitzpatrick scale.Source: Ethics Statement · p. 6
• The authors warn that missing data modes may compound training-data biases and that drastic prompt-driven changes can produce objectionable content; they recommend sophisticated NSFW filters.They describe Stable Diffusion’s CLIP-based filter as a positive step.Source: Ethics Statement · p. 6
• Future extensions could represent more groups and generate diverse people from neutral prompts; annotators were recruited through MTurk and paid $10 per hour, totaling about $60.Source: Ethics Statement · p. 6

A Related Work

• Prior work reports societal bias in text-to-image models, while this paper studies natural-language ethical interventions, building on intervention work for language-model predictions.Source: A Related Work · p. 8

B Image Generation Details

• minDALL·E and DALL·E-mini were selected for image quality and efficiency, generating an image in 10 seconds on an RTX1080Ti, versus 20 minutes for Disco Diffusion.Source: B Image Generation Details · p. 8
• Each ENTIGEN prompt was used to generate nine images from each selected model nine times.Source: B Image Generation Details · p. 8
• Table 4 contains the texts used for CLIP-based classification.Source: Table 4 · p. 8

C Details of CLIP-based and Human Evaluation

• CLIP first filters images by person-versus-object compatibility, then estimates gender by comparing similarity to man, woman, and uncertain text prompts.Source: C Details of CLIP-based and Human… · p. 8
• Human evaluators check whether an image depicts a person and matches its prompt, then judge gender and skin color; three annotators assess each image, with majority views determining results.Source: C Details of CLIP-based and Human… · p. 8

D List of Studied Professions, Objects and Cultural Scenarios

• ENTIGEN includes 25 professions, 8 objects, and 3 cultural scenarios, selected for prominent and representative visual features.Source: D List of Studied Professions, Objects… · p. 8

C Details of CLIP-based and Human Evaluation

• Figure 2 shows the annotation interface used to collect human evaluation results.Source: C Details of CLIP-based and Human… · p. 8, Figure 2 · p. 9

E Prompts and Ethical Interventions in ENTIGEN

• Each original prompt has 4–7 ethical-intervention variants, including 1–2 irrelevant interventions and others addressing gender, skin color, or culture.Source: E Prompts and Ethical Interventions in… · p. 9

F More on Bias Results

• Ethical interventions reduce the absolute bias score in most cases, according to human evaluations.Figure 3 reports the bias results; the bias formulation is given in Equation 2.Source: F More on Bias Results · p. 9
• For DALL·E-mini images of people wearing makeup, the bias direction flips toward people who look like men.Source: F More on Bias Results · p. 9
• The paper presents bias results from human evaluations alongside a formulation of bias for social axis g.Source: F More on Bias Results · p. 9

G Case Study

• Bride prompts originally generated only Western weddings, while the intervention “from diverse cultures” diversified the generations.Source: G Case Study · p. 9
• Table 5 lists attributes by category for CLIP evaluation and marks those also evaluated by human annotators.Source: Table 5 · p. 9
• Table 6 reports CLIP-based and human evaluation results for the objects category.Source: Table 6 · p. 10
• A positive bias score denotes a bias toward people categorized as men, light-skinned, and Western by human annotators.Source: Figure 3 · p. 10
• Figures 4–8 show prompt-conditioned generations for doctor, police officer, make, and bride attributes across Stable Diffusion, DALL·E-mini, and minDALL·E.Source: Figure 4 · p. 11, Figure 5 · p. 11, Figure 6 · p. 12
• Table 7 lists pre-training-data contexts in which the phrase “irrespective of” occurs.Source: Table 7 · p. 13