Survey of Bias In Text-to-Image Generation: Definition, Evaluation, and Mitigation

1 Introduction

• T2I models can depict a skewed world, including white masculine CEOs and dark-skinned men as criminals.The introduction frames this as a question about what kind of world T2I models depict.Source: §1 · pp. 1–2
• Bias in generated images could contribute to representational and allocational harms, including reinforcing stereotypes and under-representing marginalized communities.The paper gives the example of images depicting more than 80% of “inmates” with dark skin despite people of color comprising less than half of the U.S. prison population.Source: §1 · p. 2
• The authors review 36 papers on gender presentation, skintone, and geo-cultural bias, examining how prior work defines, evaluates, and mitigates these biases.The review addresses three dimensions identified as primarily studied in prior work.Source: §1 · p. 2
• The review finds that geo-cultural bias, non-binary representation, and non-occupational aspects remain under-studied, while evaluation lacks a unified framework and mitigation is not comprehensive.Gender and skintone receive more attention than geo-cultural bias; occupational association dominates gender and skintone research.Source: §1 · p. 3

2 Bias Definitions

• The taxonomy groups prior definitions into gender, skintone, and geo-cultural bias, and its categories are not mutually exclusive.The authors stress that bias definitions vary and should be grounded in social harms to support clearer discussion.Source: §2 · p. 3

2.1 Gender Bias in T2I Models

• Reviewed studies characterize gender bias in text-to-image models through default-generation tendencies, occupational associations, characteristics and interests, stereotypical objects, and differences in generated image quality.Most studies treat gender as perceived presentation and roles rather than gender or sexual identity; only 6 of 36 papers explored bias beyond binary gender.Source: §2.1 · pp. 3–4
• Gender-bias definitions include stereotypical genders generated for powerful or powerless prompts, men for CEO and women for assistant, NSFW outputs for non-cisgender individuals, and disproportionate amplification of harmful input for feminine generations.Source: §2.1 · p. 4

2.2 Skintone bias in T2I Models

• Skintone bias is defined through default skintone depictions, occupational associations, and characteristics such as “poor,” “pleasant,” or “a criminal.”The surveyed work also describes a “white ideal,” depicting attractive people as white and poor people as people of color [9].Source: §2.2 · p. 4

2.3 Geo-Cultural bias in T2I Models

• Geo-cultural bias definitions cover over-representation of some cultures in default generations, under-representation of others, and stereotyped depictions of cultures, facial features, and language scripts.Examples include poverty-indicative images of Africa [9] and associations with culturally specific scripts [97].Source: §2.3 · p. 4

3 Bias Evaluation

• The surveyed evaluations lack unified datasets, frameworks, and metrics, with many studies using different handcrafted prompts.The paper reviews evaluation datasets and metrics to describe this scattered landscape.Source: §3 · p. 4

3.1 Evaluation Datasets

• Proposed evaluation resources include annotated or multilingual datasets, cultural and nationality benchmarks, and HEIM, while other studies use predefined datasets or prompt collections.Examples include MAGBIG [34], CCUB [59], ViSAGe, and HEIM [53]; Friedrich et al. use a subset of LAION-5B, while Vice et al. use 64 MS-COCO prompts.Source: §3.1 · p. 5

3.2 Evaluation Metrics

• Metrics are classified as classification-based, which directly infers characteristics, or embedding-based, which measures subtler characteristics in latent space.A majority of surveyed studies use classification-based metrics, though identity should not be determined only by appearance and current classifiers fail to represent transgender individuals.Source: §3.2 · p. 5

3.2.1 Classification-Based Metrics

• Classification-based evaluations use human annotations, classifiers, VQA models, and embedding-distance classification, often reporting demographic parity or other distribution measures.These methods include human ratings, CLIP and FairFace classifiers, VQA-based gender classification, and CLIP- or skin-pixel-based comparisons.Source: §3.2.1 · pp. 5–6

3.2.2 Embedding-Based Metrics

• Embedding-based measures include CLIP proximity comparisons, association scores, culturally explicit prompt comparisons, and image-quality measures such as FID and CLIPScore.Examples include T2IAT [105], MCAS, Relative Bias [97], and cluster-assignment shares [60].Source: §3.2.2 · pp. 6–7

4 Bias Mitigation

• Mitigation strategies fall into model-weight refinement and inference-time or data approaches that leave model weights unchanged.Most surveyed papers use prompt-based mitigation, which the review characterizes as low in robustness and controllability.Source: §4 · p. 7

4.1 Model Weight Refinement

• Model-weight refinement includes fine-tuning text encoders or models, fairness-oriented parameter-efficient training, and editing cross-attention layers.Examples include Fair Mapping [54], soft prompts, LoRA [94], ITI-GEN [112], and TIME [75].Source: §4.1 · p. 7

4.2 Inference-Time and Data Approaches

• Prompt-based mitigation uses fairness instructions or anti-stereotype prompts, but studies report overshooting, persistent bias, or degraded alignment and face-generation performance.Wan & Chang [104] report limited controllability and overshooting; Friedrich et al. [34] find gender-neutral prompts do not remove bias and degrade alignment and face generation.Source: §4.2 · pp. 7–8
• Other inference-time approaches guide outputs toward specified gender characteristics or distributions, sample diverse latent vectors, or fine-tune on diverse generated images.Fair Diffusion supports human-defined fairness instructions, while He et al. use multidirectional guidance with iterative distribution alignment.Source: §4.2 · p. 8

5.1 Human-Centric Approaches for Bias Definitions

• Researchers should clarify which generation tasks and stereotypes they assess, distinguish visual presentation from identity, and ground bias concepts in social inequality and power differences.The survey warns against conflating skintone with race or ethnicity and culture with nationality.Source: §5.1 · p. 8
• Bias research should extend beyond occupational associations and examine under-studied dimensions such as disability, LGBTQIA+, and geo-cultural bias.The survey notes that few studies examine geo-cultural bias and that gender research often centers on occupations.Source: §5.1 · p. 8

5.2 Human-Centric Approaches for Bias Evaluation

• Evaluations should include community perspectives and disclose researcher and annotator positionality, since binary-focused strategies can overlook transgender and non-binary experiences.The survey links narrow evaluation strategies to reliance on binary gender stereotypes.Source: §5.2 · p. 8
• Automated gender evaluation needs more reliable VQA models, which can misidentify gender in complex images and may carry their own biases.The cited example is an image containing two people.Source: §5.2 · p. 9

5.3 Human-Centric Approaches for Bias Mitigation

• Mitigation should account for user preferences because diverse outputs alone do not ensure users feel represented or included.Few surveyed methods considered user preferences; Ungless et al. [101] studied how non-binary users prefer to be represented.Source: §5.3 · p. 9
• Training-based mitigation can require data and resources and risk catastrophic forgetting or degraded generation quality, while prompt-based methods can be unstable and difficult to control.The survey calls for methods that are robust, controllable, and resource-friendly.Source: §5.3 · p. 9
• One-size-fits-all mitigation may produce inaccurate or offensive images, while alignment to human preferences raises challenges including proxy objectives, reward hacking, and overoptimization.Prompt blocking may also contribute to erasing non-cisnormative identities; T2I alignment has mostly targeted image quality and prompt matching rather than societal bias.Source: §5.3 · p. 9
• Because bias concepts and social norms change, the survey encourages adaptive, lifelong mitigation that responds to continuous feedback and community needs.Real-time monitoring and automatic parameter adjustment are offered as possible approaches.Source: §5.3 · p. 9

6 Conclusion

• The authors reviewed definitions, evaluation metrics, and mitigation methods across 36 papers on gender presentation, skintone, and geo-cultural bias.They describe this as the first survey of biases in T2I models.Source: §6 · p. 9
• The survey finds that gender and skintone receive more attention than geo-cultural bias, occupational associations dominate gender and skintone research, and non-binary representation is often overlooked.It also reports limited study of power dynamics and explicit content generation.Source: §6 · p. 10
• Evaluation metrics vary across studies, no unified framework has been consolidated, and current mitigation methods lack holistic and effective solutions.The survey highlights human-centric directions for bias definitions, evaluation, and mitigation.Source: §6 · p. 10

7 Ethics Statement

• The survey treats gender, skintone, and geo-culture as visual presentations, not internal identities, and cautions against inferring race or ethnicity from images.It also distinguishes culture from nationality and uses “gender presentation,” “skintone,” and “geo-cultural” for its dimensions.Source: §7 · p. 10
• Human annotation and automated classification can both carry bias, risking its propagation into T2I bias evaluation results.The surveyed approaches include human annotation, model-based classifiers, and embedding-distance classification.Source: §7 · p. 10
• The survey's definitions and methods should not be used to legitimize generative AI applications involving surveillance, worker harm or suppression, or weaponization.The authors warn against ethics-washing such applications through bias evaluation or mitigation.Source: §7 · p. 10

A Literature Selection Methodology

• The authors began by examining highly cited T2I papers on bias and fairness indexed by Google Scholar and arXiv, then followed works cited by and citing those papers.This approach was intended to capture a wide range of perspectives and methodologies.Source: A Literature Selection Methodology · p. 19
• The review included both peer-reviewed works and preprints because the field is relatively new and many papers are recent.Source: A Literature Selection Methodology · p. 19
• The authors acknowledge that relevant papers may have been overlooked despite efforts to make the selection thorough.Source: A Literature Selection Methodology · p. 19

• 18 of 36 reviewed works used classifiers to identify demographic characteristics in generated images.Source: B Summarization of Evaluation Metrics · p. 19
• Only a few studies explored embedding-based measures such as generation diversity and image quality.Source: B Summarization of Evaluation Metrics · p. 19
• Table 2 stratifies prior bias-evaluation metrics by metric types and categories.Source: Table 2 · p. 19

• Most reviewed mitigation studies explored Parameter-Efficient Finetuning and prompt-based methods.Source: C Summarization of Mitigation Metrics · p. 19, Table 3 · p. 20
• Model editing, inference-time guidance and sampling, and data augmentation remain under-explored mitigation approaches.Source: C Summarization of Mitigation Metrics · p. 19
• Table 3 stratifies reviewed T2I bias-mitigation approaches by mitigation types and methods.Source: Table 3 · p. 20