Auditing Gender Presentation Differences in Text-to-Image Models

Zhang, Yanzhe; J, Lu; Turk, Greg; Yang, Diyi

doi:10.48550/arxiv.2302.03675

Cited by 5 publications

(4 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These systems were chosen because they are popular implementations of different state-of-the-art text-to-image generative AI techniques using diffusion models [12] and CLIP image embeddings [13] that were developed primarily by researchers and academics. These models have also been studied in academic research for the content nature, and cultural and social biases of their generated outputs [14,15,16,17,18,19,20]. For this experiment, Stable Diffusion's v1-4 and v2-1 pretrained weights will each be used independently in conjunction with Stable Diffusion Web UI.…”

Section: Methodology For Interviewsmentioning

confidence: 99%

From Development to Dissemination: Social and Ethical Issues with Text-to-Image AI-Generated Art

2023

Proceedings of the Canadian Conference on Artificial Intelligence

View full text Add to dashboard Cite

Text-to-image generative artificial intelligence (AI) have made global news headlines for not only having the ability to generate high-fidelity artworks, but also for causing increased discussion on the ethicality of its impact on living artists, the automation and commodification of art production, the frequent non-consensual collection and usage of sensitive and copyrighted images as training data, and the routinely exhibited cultural and social biases in their generated outputs. In addition, there are concerns that open-sourced text-to-image generative AI models, such as Stable Diffusion, and techniques like Textual Inversion, allow for technical restrictions on the content subject matter to be removed and for generated images to be subject specific, which could be utilized as a new medium for disinformation and sexual or targeted abuse. Because ethical discussions on AI-generated art using text-to-image generative AI models have only come to light in the last quarter of 2022, academic research on the social and ethical implications of this technology have yet to be thoroughly explored. Therefore, it is imperative for research to be done on these implications with regards to the technological development, evaluation, perception, creation, and moderation of AI-generated artworks while text-to-image generative AI systems are still in the preliminary stages of public dissemination and adoption.

show abstract

Section: Methodology For Interviewsmentioning

confidence: 99%

From Development to Dissemination: Social and Ethical Issues with Text-to-Image AI-Generated Art

2023

Proceedings of the Canadian Conference on Artificial Intelligence

View full text Add to dashboard Cite

show abstract

“…Secondly, Stable-Diffusion V2.1 (Rombach et al 2022) is utilized to generate ten images per prompt resulting in 5k images. To address any potential bias (Zhang et al 2023;Bakr et al 2023) in the generated images, a human evaluation was conducted to filter out the non-agnostic images based on two simple questions; 1) Do you recognize a human in the scene? 2) If yes, Are the gender and race anonymous?…”

Section: Which Metric Is Better?mentioning

confidence: 99%

“…Recent efforts focus on estimating model bias, driven by the fact that more than balanced data is needed to create unbiased models (Wang et al 2019). Bias (Zhang et al 2023;Bolukbasi et al 2016;Caliskan, Bryson, and Narayanan 2017) is characterized by the model's representation of different subgroups when generating the supergroup, such as assessing whether it equally depicts men and women in images of people. The primary cause of bias stems from spurious correlations captured during training.…”

Section: Introductionmentioning

confidence: 99%

ImageCaptioner2: Image Captioner for Image Captioning Bias Amplification Assessment

Abdelrahman,

Sun,

et al. 2024

AAAI

View full text Add to dashboard Cite

Most pre-trained learning systems are known to suffer from bias, which typically emerges from the data, the model, or both. Measuring and quantifying bias and its sources is a challenging task and has been extensively studied in image captioning. Despite the significant effort in this direction, we observed that existing metrics lack consistency in the inclusion of the visual signal. In this paper, we introduce a new bias assessment metric, dubbed ImageCaptioner2, for image captioning. Instead of measuring the absolute bias in the model or the data, ImageCaptioner2pay more attention to the bias introduced by the model w.r.t the data bias, termed bias amplification. Unlike the existing methods, which only evaluate the image captioning algorithms based on the generated captions only, ImageCaptioner2incorporates the image while measuring the bias. In addition, we design a formulation for measuring the bias of generated captions as prompt-based image captioning instead of using language classifiers. Finally, we apply our ImageCaptioner2metric across 11 different image captioning architectures on three different datasets, i.e., MS-COCO caption dataset, Artemis V1, and Artemis V2, and on three different protected attributes, i.e., gender, race, and emotions. Consequently, we verify the effectiveness of our ImageCaptioner2metric by proposing Anonymous-Bench, which is a novel human evaluation paradigm for bias metrics. Our metric shows significant superiority over the recent bias metric; LIC, in terms of human alignment, where the correlation scores are 80% and 54% for our metric and LIC, respectively. The code and more details are available at https://eslambakr.github.io/imagecaptioner2.github.io/.

show abstract

“…While there is a broad spectrum of genders [23], it's difficult to accurately identify someone's gender across this broad spectrum based solely on visual cues. Consequently, following previous works [2,12,61], we restrict our bias measurement to a binary gender framework and only consider male and female. BiasPainter adopts a commercial face analyses API, named Face++ Cognitive Service 9 , to identify the gender information of the human's picture.…”

Section: Properties Assessmentmentioning

confidence: 99%

Characterization of Two BAHD Acetyltransferases Highly Expressed in the Flowers of Jasminum sambac (L.) Aiton

Wang

Zhang

Chen

et al. 2021

Plants

View full text Add to dashboard Cite

Volatile benzenoid compounds are found in diverse aromatic bouquets emitted by most moth-pollinated flowers. The night-blooming Jasminum sambac is widely cultivated worldwide in the tropics and subtropics for ornamental and industrial purposes owing to its fragrant flowers. Benzylacetate is a characteristic constituent in jasmine scent which makes up to approximately 20–30% of the total emission in the headspace or extract, but the biosynthesis enzymes and the encoding genes have not yet been described. Here, we identify two cytosolic BAHD acyltransferases specifically expressed in the petals with a positive correlation closely to the emission pattern of the volatile benzenoids. Both JsBEAT1 and JsBEAT2 could use benzylalcohol and acetate-CoA as substrates to make benzylacetate in vitro. The recombinant GST-JsBEAT1 has an estimated apparent Km of 447.3 μM for benzylalcohol and 546.0 μM for acetate-CoA, whereas in the instance of the His-JsBEAT2, the Km values are marginally lower, being 278.7 and 317.3 μM, respectively. However, the catalytic reactions by the GST-JsBEAT1 are more efficient than that by the His-JsBEAT2, based on the steady-state kcat parameters. Furthermore, ectopic expression of JsBEAT1 and JsBEAT2 in the transgenic P. hybrida plants, driven by a flower-specific promotor, significantly enhances the biosynthesis of benzylbenzoate and benzylacetate, as well as the total VOCs.

show abstract

Auditing Gender Presentation Differences in Text-to-Image Models

Cited by 5 publications

References 18 publications

From Development to Dissemination: Social and Ethical Issues with Text-to-Image AI-Generated Art

From Development to Dissemination: Social and Ethical Issues with Text-to-Image AI-Generated Art

ImageCaptioner2: Image Captioner for Image Captioning Bias Amplification Assessment

Characterization of Two BAHD Acetyltransferases Highly Expressed in the Flowers of Jasminum sambac (L.) Aiton

Contact Info

Product

Resources

About