BiasConnect: Investigating Bias Interactions in Text-to-Image Models

Abstract

The biases exhibited by Text-to-Image (TTI) models are often treated as if they are independent, but in reality, they may be deeply interrelated. Addressing bias along one dimension, such as ethnicity or age, can inadvertently influence another dimension, like gender, either mitigating or exacerbating existing disparities. Understanding these interdependencies is crucial for designing fairer generative models, yet measuring such effects quantitatively remains a challenge. In this paper, we aim to address these questions by introducing BiasConnect, a novel tool designed to analyze and quantify bias interactions in TTI models. Our approach leverages a counterfactual-based framework to generate pairwise causal graphs that reveals the underlying structure of bias interactions for the given text prompt. Additionally, our method provides empirical estimates that indicate how other bias dimensions shift toward or away from an ideal distribution when a given bias is modified. Our estimates have a strong correlation (+0.69) with the interdependency observations post bias mitigation. We demonstrate the utility of BiasConnect for selecting optimal bias mitigation axes, comparing different TTI models on the dependencies they learn, and understanding the amplification of intersectional societal biases in TTI models.

Full text

BiasConnect: Investigating Bias Interactions in Text-to-Image Models
Pushkar Shukla1Aditya Chinchure2Emily Diana3Alexander Tolbert4
Kartik Hosanagar5Vineeth N. Balasubramanian6Leonid Sigal2Matthew A. Turk1
1Toyota Technological Institute at Chicago 2University of British Columbia
3Carnegie Mellon University, Tepper School of Business 4Emory University
5University of Pennsylvania, The Wharton School 6Indian Institute of Technology Hyderabad
{pushkarshukla, mturk}@ttic.edu {aditya10, lsigal}@cs.ubc.ca
Abstract
The biases exhibited by Text-to-Image (TTI) models are of-
ten treated as if they are independent, but in reality, they
may be deeply interrelated. Addressing bias along one di-
mension, such as ethnicity or age, can inadvertently influ-
ence another dimension, like gender, either mitigating or
exacerbating existing disparities. Understanding these in-
terdependencies is crucial for designing fairer generative
models, yet measuring such effects quantitatively remains a
challenge. In this paper, we aim to address these questions
by introducing BiasConnect , a novel tool designed to an-
alyze and quantify bias interactions in TTI models. Our
approach leverages a counterfactual-based framework to
generate pairwise causal graphs that reveals the underlying
structure of bias interactions for the given text prompt. Ad-
ditionally, our method provides empirical estimates that in-
dicate how other bias dimensions shift toward or away from
an ideal distribution when a given bias is modified. Our
estimates have a strong correlation (+0.69) with the inter-
dependency observations post bias mitigation. We demon-
strate the utility of BiasConnect for selecting optimal bias
mitigation axes, comparing different TTI models on the de-
pendencies they learn, and understanding the amplification
of intersectional societal biases in TTI models.
1. Introduction
Text-to-Image (TTI) models such as DALL-E [46], Imagen
[51], and Stable Diffusion [48] have become widely used
for generating visual content from textual prompts. Despite
their impressive capabilities, these models often inherit and
amplify biases present in their training data [10,11,60].
These biases manifest across multiple social and non-social
dimensions – including gender, race, clothing, and age –
leading to skewed or inaccurate representations. As a re-
sult, TTI models may reinforce harmful stereotypes and
Old
Middle-Age
Young
0%
50%
100%
Old
Middle-Age
Young
0%
50%
100%
Male
Female
Male
Female
Initial set has a mix of young, middle
aged, and old musicians, while gender
diversity could be better.
Images for “A photo of a musician”
Gender distribution improves !, but
Age distribution gets much worse ☹
Age
Gen
This is because Gender inffluences Age negatively
&
'
-0.16 ( Intersectionality!
If you mitigate Gender Bias…
Figure 1. An example output of BiasConnect, revealing the neg-
ative impact of bias mitigation along one dimension on another
dimension. Here, increasing the gender diversity (GEN) skews
age distribution (AGE) for images of musicians generated by Sta-
ble Diffusion 1.4 [48].
societal norms [4,6]. While significant efforts have been
made to evaluate and mitigate societal biases in TTI models
[5,10,11,18,23,62], these approaches often assume that
biases along different dimensions (e.g., gender and race) are
independent of each other. Consequently, they do not ac-
count for relationships between these dimensions. For in-
stance, as illustrated in Figure 1, mitigating gender (male,
female) may effectively diversify the gender distribution in
a set of generated images, but this mitigation step may neg-
atively impact the diversity of another bias dimension, like
age. This relationship between two bias dimensions high-
lights the intersectional nature of these biases.
The concept of intersectionality, first introduced by
1
arXiv:2503.09763v1 [cs.CV] 12 Mar 2025

Crenshaw [12], motivates the need to understand how over-
lapping social identities such as race, gender, and class con-
tribute to systemic inequalities. In TTI models, these in-
tersections can have a significant impact. As a motivating
study, we independently mitigated eight bias dimensions
over 26 occupational prompts on Stable Diffusion 1.4, using
a popular bias mitigation strategy, ITI-GEN [64] (see Supp.
A.5). We found that while the targeted biases were reduced
in most cases, biases along other axes were negatively af-
fected in over 29.4% of the cases. This suggests that for an
effective bias mitigation strategy, it is crucial to understand
which biases are intersectional. Additionally, it is impor-
tant to consider whether mitigating one bias affects other
biases equally or unequally, and if these effects are positive
or negative. Answering these questions can improve our
understanding of TTI models, enhance their interpretabil-
ity, and develop better bias mitigation strategies.
To understand how biases in TTI models influence each
other, we propose BiasConnect , a first-of-its-kind analysis
tool that evaluates societal biases in TTI models while ac-
counting for intersectional relationships, rather than treat-
ing them in isolation. Our approach enables us to identify
intersectional relationships and study the positive and nega-
tive impacts that biases have on each other, preventing unin-
tended consequences that may arise from mitigating biases
along a single axis. Our tool analyzes intersectionality at the
prompt level, but also enables comparative studies across
models by aggregating over a set of prompts, thus providing
a means to uncover how variations in architecture, datasets,
and training objectives contribute to bias entanglement.
Given an input prompt to a TTI model, BiasConnect
uses counterfactuals to quantify the impact of bias diversi-
fication (intervention) along one bias axis on any other bias
axis. We refer to this as a pairwise causal relationship be-
tween the axis of intervention and the axis on which the ef-
fect is observed, and we visualize these relationships in the
form of a pairwise causal graph. Additionally, we provide
empirical estimates (called Intersectional Sensitivity) that
measure how bias mitigation along one dimension influ-
ences biases in other dimensions. These empirical estimates
serve as weights on the pairwise causal graph. To validate
our approach, we show how our estimates correlate with
true bias mitigation, and analyze robustness in which dif-
ferent components are systematically modified. Our overall
contributions are as follows:
• We propose BiasConnect, a novel intersectional bias
analysis tool for TTI models. Through a causal approach,
our tool captures interactions between bias axes in a pair-
wise causal graph and provides empirical estimates of
how bias mitigation along one axis affects other axes,
through the Intersectional Sensitivity score.
• We show that our empirical estimates strongly corre-
late (+0.696) with the intersectionality observed post-
mitigation, and through extensive qualitative results, val-
idate the analyses provided by our tool.
• Finally, we demonstrate the usefulness of the tool in con-
ducting audits on multiple open-source TTI models, iden-
tifying optimal mitigation strategies to account for inter-
sectional biases, and showing how bias interactions in
real-wold or a training dataset may change in TTI model
generated images.
2. Related Work
2.1. Intersectionality and Bias in AI
Intersectionality, introduced by Crenshaw [12], describes
how multiple forms of oppression—such as racism, sexism,
and classism—intersect to shape unique experiences of dis-
crimination. Two key models define this concept: the addi-
tive model, where oppression accumulates across marginal-
ized identities, and the interactive model, where these iden-
tities interact synergistically, creating effects beyond sim-
ple accumulation [13]. In the context of AI, most existing
work [15,26,33,34] aligns more closely with the additive
model, focusing on quantifying and mitigating biases in in-
tersectional subgroups. This perspective has influenced fair-
ness metrics [16,19,22] designed to assess subgroup-level
performance, extending across various domains, including
natural language processing (NLP) [24,37,38,57] and re-
cent large language models [3,14,35,41], multimodal re-
search [29,30], and computer vision [56,63]. These ap-
proaches typically measure disparities across predefined de-
mographic intersections and propose mitigation strategies
accordingly. Our work aligns with the interactive model of
intersectionality, using counterfactual-driven causal analy-
sis in TTI models. Beyond subgroup analysis, we intervene
on a single bias axis to assess its ripple effects on others,
revealing independences and interactions.
2.2. Bias in Text-to-Image Models
Extensive research has been conducted on evaluating and
mitigating social biases in both image-only models [8,27,
32,40,42,53,58,59,63] and text-only models [2,7,21,
31,54]. More recently, efforts have expanded to multimodal
models and datasets, addressing biases in various language-
vision tasks. These investigations have explored biases in
embeddings [25], text-to-image (TTI) generation [5,11,18,
23,52,62,64], image retrieval [61], image captioning [27,
65], and visual question-answering models [1,28,44].
Despite these advances, research on intersectional biases
in TTI models remains limited. Existing evaluation frame-
works such as T2IAT [62], DALL-Eval [11], and other
studies [5,18,20,23] primarily assess biases along pre-
defined axes, such as gender [5,11,18,23,62], skin tone
[5,11,18,23,62], culture [18,62], and geographical loca-
tion [18]. While these works offer key insights into single-
2

axis bias detection and mitigation, they lack a systematic ex-
amination of how biases on one axis influence another—a
core aspect of intersectionality. The closest research, TI-
BET [10], visualizes such interactions, but our approach
goes further by systematically quantifying bias interactions
and empirically estimating their impact rather than merely
identifying correlations.
3. Approach
The objective of BiasConnect is to identify and quan-
tify the intersectional effects of intervening on one bias
axis (Bx) to mitigate that bias, on any other bias axis
(By). BiasConnect , works by systematically altering input
prompts and analyzing the resulting distributions of gener-
ated images. To achieve this, we leverage counterfactual
prompts by modifying specific attributes (e.g., male and
female) along a bias axis (e.g., gender) and examine how
these interventions impact other bias dimensions (e.g., age
and ethnicity). If modifying one bias axis through coun-
terfactual intervention causes significant shifts in the distri-
bution of attributes along another bias axis, it indicates an
intersectional dependency between these axes.
We first construct prompt counterfactuals and generate
images using a TTI model (Sec. 3.1). Subsequently, to
identify bias-related attributes in the generated images, we
use a VQA model (Sec. 3.2). Next, in order to identify
whether the intersectional effects of intervening on one bias
on another axis is significant, we propose a causal discov-
ery approach, where we employ conditional independence
testing (Sec. 3.3) in a pairwise manner between the two
bias axes. Finally, to quantify the intersectional effects,
and to identify whether these effects are positive or neg-
ative, we compute the causal treatment effect, defined as
Intersectional Sensitivity (Sec. 3.4).
3.1. Counterfactual Prompts & Image Generation
Given an input prompt Pand bias axes B=
[B1, B2, . . . , Bn], we generate counterfactual prompts
{CF 1
i, . . . , CF j
i}using templates from Supp. A.1. The
original prompt Pand its counterfactuals are then used to
generate images with the TTI model to measure intersec-
tional effects.
3.2. VQA-based Attribute Extraction
To facilitate the process of extracting bias related attributes
from the generated images, we use VQA. This is inspired
by previous approaches on bias evaluation, like TIBET [10]
and OpenBias [17], where a VQA-based method was used
to extract concepts from the generated images. Similar to
previous work [10], we use MiniGPT-v2 [9] in a question-
answer format to extract attributes from generated images.
For the societal biases we analyze, we have a list of pre-
defined questions (Supp. A.3) corresponding to each bias
axis in B, and each question has a choice of attributes to
choose from. For example, for the gender bias axis, we
ask the question “[vqa] What is the gender (male,
female) of the person?”. Note that every question is
multiple choice (in this example, male and female are the
two attributes for gender). The questions asked for all im-
ages of prompt Pand its counterfactuals CF j
iremain the
same. With the completion of this process, we have at-
tributes for all images, where each image has one attribute
for each bias axis in B.
3.3. Pairwise Causal Discovery
Given an initial set of bias axes B, we define an intersec-
tional relationship between a pair of biases (Bx, By)as
Bx→By, indicating that a counterfactual intervention on
Bxto mitigate its bias also affects By. As a first step, we in-
tervene across all n×nbias relationships. Using attributes
extracted by the VQA, we can count the attributes for a bias
axis Byover any set of images. We construct a contingency
table where rows represent the intervened bias axis Bx(e.g.,
gender with male and female counterfactuals, in Example 2
of Fig. 2), and columns capture the distribution on the target
axis By(e.g., age with old, middle-aged, and young cate-
gories). The values in the contingency tables are the counts
of attributes of Byover the counterfactual image sets of Bx.
Next, we refine these relationships by extracting only
statistically significant ones. This ensures that only strong
dependencies between different bias pairs are retained.
We apply conditional independence testing using the Chi-
square (χ2) test, pruning bias pairs with respect to Bx
if their p-value exceeds a predefined threshold (p-value>
0.0001). Bias pairs with a p-value below this threshold are
considered strongly dependent, indicating that intervening
on Bxresults in a significant change in the other bias axis.
This process is applied iteratively for all bias axes. This step
is referred to as Pairwise Causal Discovery, and it returns
a set of bias pair relationships where mitigating along one
bias axis has led to a strong change in another bias dimen-
sion
3.4. Causal Treatment Effect Estimation
While Pairwise Causal Discovery can identify interventions
along bias pairs that cause significant changes, it alone is
not sufficient to determine whether the impact of interven-
tions on Bxaffects Byin a positive or negative direction
with respect to an ideal distribution. This limitation arises
because there is no direct comparison to the initial distri-
bution of Byin the original set of images generated from
prompt P, as we had only considered images from the coun-
terfactuals {CF 1
x, ..., C F j
x}for bias Bx, for pairwise causal
discovery. To address this, we propose a metric that quanti-
fies the impact of bias mitigation on dependent biases with
respect to an ideal distribution.
3

Y
MA
O
M
47
1
0
F
48
0
0
M
F
Y
0
48
MA
0
48
O
22
26
Initial Prompt: “A photo of an
elementary school teacher”
Gender Bias
CF-1 (Male - M):
“A photo of a male elementary
school teacher”
CF-2 (Female - F):
“A photo of a female elementary
school teacher”
❄ TTI Model
Age Bias
CF-1 (Young - Y):
“A photo of a young elementary
school teacher”
CF-2 (Middle-Aged - MA):
“A photo of a middle-aged
elementary school teacher”
CF-2 (Old - O):
“A photo of an old elementary
school teacher”
Bodytype Bias
……
Pairwise Causal Discovery
P-Value:
8e-11 < Threshold
Age
Gen
✅
Example 1: Age > Gender
Example 2: Gender > Age
Compute Gender distribution
across Age counterfactuals
P-Value:
0.84 > Threshold
Compute Age distribution
across Gender counterfactuals
Age
Gen
❌
… repeat for every pair of bias axes
Causal Treatment Effect Estimation
M
F
Y
0
48
MA
0
48
O
22
26
Total
22
122
M
F
Init
0
48
M F
Compute the Wasserstein Distance to our ideal
distribution, for Initial set and Age counterfactual set
Ideal Gender
Distribution
Age
Gen
Intersectional Sensitivity =
winit
gender
wage
gender
winit
gender −wage
gender
The positive value ( ) indicates that mitigating
age will likely improve gender distribution to be
closer to the ideal gender distribution
+0.153
+0.153
Figure 2. An overview of BiasConnect . We use a counterfactual-based approach to measure pairwise causality between bias axes. For
dependent axes, we measure the causal effect, estimating how bias mitigation on one axis impacts another.
Defining an Ideal Distribution. We first define a desired
(ideal) distribution D∗, which represents the unbiased state
we want bias axes to achieve. This can be a real-world
distribution of a particular bias axis, a uniform distribution
(which we use in our experiments), or anything that suits
the demographic of a given sub-population.
Measuring Initial Bias Deviation. Given the images of
initial prompt P, we compute the empirical distribution of
attributes associated with bias axis By, denoted as Dinit
By. We
then compute the Wasserstein distance between this empir-
ical distribution and the ideal distribution:
winit
By=W1(Dinit
By, D∗)(1)
where W1(·,·)represents the Wasserstein-1 distance. The
Wasserstein-1 distance (also known as the Earth Mover’s
Distance) between two probability distributions D1and D2
is defined as:
W1(D1, D2) = inf
γ∈Π(D1,D2)
E(x,y)∼γ[|x−y|](2)
where Π(D1, D2)is the set of all joint distributions γ(x, y)
whose marginals are D1and D2, and |x−y|represents the
transportation cost between points in the two distributions.
Intervening on Bx.Next, we intervene on Bxto simu-
late the mitigation of bias Bx. This intervention ensures
that all counterfactuals of Bxare equally represented in the
generated images. For example, if Bxis gender bias, we
enforce equal proportions of male and female individuals
in the dataset. This intervention is in line with most bias
mitigation methods proposed for TTI models, like ITI-GEN
[64]. Using our counterfactuals along Bx, we sum the dis-
tributions on Byacross all counterfactuals of Bx. This sum
across the counterfactuals of Bxyields a new empirical dis-
tribution of By, denoted DBx
By, simulating the effect of miti-
gating Bx(See Fig 2). We compute its Wasserstein distance
from the ideal distribution.
wBx
By=W1(DBx
By, D∗)(3)
Computing Intersectional Sensitivity. To quantify the
effect of mitigating Bxon By, we define the metric,
Intersectional Sensitivity, as:
ISxy =winit
By−wBx
By(4)
A positive value (ISxy >0) indicates that mitigating Bx
improves By, bringing it closer to the ideal distribution,
while a negative value (ISxy <0) suggests it worsens By,
moving it further from the ideal. If I Sxy = 0, mitigat-
ing Bxhas no effect on By. This approach enables us to
assess whether addressing one bias (e.g., gender) improves
or worsens another (e.g., ethnicity) in generative models,
providing a systematic way to evaluate trade-offs and unin-
tended consequences in bias mitigation strategies. We use
Intersectional Sensitivity (I Sxy ) as a measure of intersec-
tionality for Bx→By.
3.5. Visualization
Following the process above, we have a set of pairwise
causal relationships for all significant intersectional bias
pairs Bx→By. Furthermore, each pair Bx→Byhas
an Intersectional Sensitivity score to quantify the intersec-
tional effects. There are many ways to represent these pair-
wise relationships, including building an n×nmatrix, or
4

Mountain parks are shown in diverse seasons
Coastal parks are always summer
TIBET (Stable Diffusion 2.1): “A beautiful day in the park”
Sea
Geo
Stable Diffusion 3.5: “A photo of a librarian”
Emo
Clo
Informally dressed are happy
Formally dressed are serious
Hispanic athletes are gender diverse
South Asian athletes are always male
Flux-dev: “A photo of an athlete”
Eth
Gen
Negative Score from Ethnicity to Gender:
IS
Eth
Gen
-0.198
Initial set of athletes are gender diverse !
Initial Set:
Male
Female
Male
Female
0%
50%
100%
If ethnicity is mitigated, most athletes will be male ☹
[Estimated]
Mitigating
Ethnicity:
(a) (b)
(c)
Kandinsky: “A photo of an sales person”
Positive Score from Ethnicity to Clothing:
IS
Eth
Clo
+0.24
Initial set of sales people are mostly formal
Initial Set:
Formal
Informal
Formal
Informal
If ethnicity is mitigated, we see diverse clothing !
[Estimated]
Mitigating
Ethnicity:
(d)
Formal
Informal
0%
50%
100%
Mitigating clothing is unable to produce informal ☹
[Estimated]
Mitigating
Clothing:
+0.115 -0.073
Figure 3. The figure illustrates bias interpretations from Bias Connects, combining all pairwise graphs into one. (a) Shows how mitigating
clothing bias also mitigates emotion bias. (b) Explores interactions between non-traditional bias axes in the TIBET dataset. (c) Reveals
that generating ethnically diverse athletes reduces gender diversity. (d) Demonstrates that diversifying salesperson clothing is best achieved
by increasing ethnic diversity rather than directly specifying clothing variation.
a graph with nnodes and directed edges that represent the
relationships between these nodes.
A user of BiasConnect may want to understand all im-
portant intersectional effects together. To that end, we adopt
a graph representation for our output. This graph is referred
to as a Pairwise Causal Graph in the rest of the paper. Fig-
ures 3and 5show examples of such graphs. To interpret this
graph, first pick a focal node where the intervention takes
place. All outgoing edges from this node indicate inter-
sectional relationships that are statistically significant. The
weights of the edges show the Intersectional Sensitivity and
can be interpreted as the impact of intervention on the bias
axis for the focal node.
4. Causal Interpretations
Pairwise Causal Discovery. Our approach is causal as it
involves explicit interventions to measure the effect of one
variable on another, aligning with Pearl’s Ladder of Causal-
ity [45]. Rather than analyzing existing images, we ac-
tively modify bias attributes (e.g., gender, race, age) in input
prompts. However, our pairwise causal discovery pipeline
does not capture indirect causal effects between bias axes.
Causal Treatment Effect. The Intersectional Sensitivity
metric is a causal treatment effect metric because it quan-
tifies how mitigating one bias (Bx) causally influences an-
other bias (By) through an intervention-based approach. By
actively modifying Bx(e.g., ensuring equal representation
across its attributes) and measuring changes in the Wasser-
stein distance of Byfrom an ideal distribution, we estimate
the causal impact of debiasing. This aligns with counterfac-
tual causal inference [43], where we compare the observed
outcome (Bydistribution) with its initial state had no in-
tervention occurred. The method follows Rubin’s causal
model [49,50], treating bias mitigation as a treatment-
control experiment, and can be represented in a Directed
Graph as Bx→By, making it distinct from mere corre-
5

lation analysis. The metric ISxy in Equation 4captures
the magnitude of causal influence, providing insights into
whether mitigating one bias improves or worsens another.
5. Experiments
In this section, we begin by explaining the two
datasets—the occupation prompts and the TIBET
dataset—that we use to test BiasConnect (Sec. 5.1).
Following that, show the usefulness of BiasConnect by
analyzing prompts to study prompt-level bias intersection-
ality (Sec. 5.2), and validate our Intersectional Sensitivity
with the help of a downstream bias mitigation tool,
ITI-GEN (Sec. 5.3). Finally, we analyze the robustness
of BiasConnect on the number of images generated per
prompt, and errors in VQA (Sec. 5.4).
5.1. Models and Datasets
Occupation Prompts. To facilitate a structured evaluation,
we develop a dataset with 26 occupational prompts, along
eight distinct bias dimensions: gender, age, ethnicity, envi-
ronment, disability, emotion, body type, and clothing. We
generate 48 images for all initial counterfactual prompts us-
ing five Text-to-Image models: Stable Diffusion 1.4, Stable
Diffusion 3.5, Flux [36], Playground v2.5 [39] and Kandin-
sky 2.2 [47,55]. Further details about the prompts, bias
axes, and counterfactuals are provided in the Supp. A.1.
TIBET dataset. The TIBET dataset includes 100 creative
prompts with LLM-generated bias axes and counterfactuals
[10]. Its diversity of prompts and bias axes, unrestricted to
a fixed set, enhances its utility. Additionally, it provides 48
Stable Diffusion 2.1-generated images per initial and coun-
terfactual prompt (See Supp. A.6 for more details).
5.2. Studying prompt-level intersectionality
BiasConnect enables prompt-level analysis of intersec-
tional biases (Fig. 3), helping users identify key bias axes
and develop effective mitigation strategies. For instance,
in Fig. 3(a), Stable Diffusion 3.5 exhibits a causal link
between clothing and emotion bias—informally dressed li-
brarians appear happy, while formally dressed ones seem
serious. A strongly positive Intersectional Sensitivity (I S=
0.115) indicates that diversifying clothing alone is sufficient
to diversify emotion, without explicitly mitigating emotion
bias. Conversely, Fig. 3(c) illustrates how ethnicity can neg-
atively impact gender diversity. South Asian athletes, for
example, are predominantly depicted as male. The negative
Intersectional Sensitivity (IS =−0.198) suggests that mit-
igating ethnicity alone would further skew gender represen-
tation toward males. These interpretations of our tool have
various applications, including identifying optimal bias mit-
igation strategies and comparing multiple TTI models, as
discussed in Section 6.
Prompt Edges Corr. MaxInf MaxImp
Pharmacist 12 +0.399 Gender Age
Scientist 9 +0.600 Clothing Ethnicity
Doctor 9 +0.638 Age Disability
Librarian 14 +0.805 Emotion Age
Nurse 5 +0.997 Age Disability
Chef 8 +0.757 Bodytype Ethnicity
Politician 10 +0.782 Emotion Disability
Overall -+0.696 Gender Age
Table 1. Correlation Between Estimates and Post-Mitigation
Evaluation on ITI-GEN. The high correlation validates our mit-
igation estimates. For each prompt, we report one of the most
influenced node (MaxInf) and the node with the greatest impact
on others (MaxImp).
5.3. Validating Intersectional Sensitivity
Our approach estimates how counterfactual-based mitiga-
tion affects bias scores using the Intersectional Sensitivity .
To validate this, we debias ITI-GEN, mitigate biases along
each dimension, and measure the correlation between pre-
and post-mitigation Intersectional Sensitivity values. As
shown in Table 1, we achieve an average correlation of
+0.696 across occupations, with higher values for specific
prompts like Nurse (+0.997). The strong correlation ob-
served between pre- and post-mitigation bias scores sug-
gests that our approach effectively captures the potential
impacts of interventions, offering a transparent and data-
driven way to evaluate model fairness. More details re-
garding our experimental setup have been provided in Supp.
A.7.
5.4. Robustness of BiasConnect
We analyze the robustness of our method by evaluating
the impact of image generation and VQA components on
pairwise causal graphs and Intersectional Sensitivity val-
ues through experiments on image set size and VQA error
rates across occupation prompts. This robustness analysis is
useful because it ensures the reliability and stability of our
method across varying conditions.
Number of Images. Our method generates 48 im-
ages per prompt to study bias distributions reliably. To
assess the impact of reducing image count, we an-
alyze changes in the total number of edges in the
pairwise causal graph and the percentage change in
Intersectional Sensitivity (Fig. 4(a-b)). Removing 8 im-
ages (16.6%) results in only 2.4 edge changes and a mi-
nor 5.5%shift in Intersectional Sensitivity. Even with
16 images removed (33.3%), only 4.8 edges change, and
Intersectional Sensitivity shifts by 8%. This low im-
pact suggests that TTI models consistently generate simi-
lar bias distributions (e.g., always depicting nurses as fe-
6

Robustness Analysis with VQA Error Rate
Robustness Analysis with Size of Gen. Image Set
(a) (b)
(c) (d)
Figure 4. Sensitivity analysis on BiasConnect. . We evaluate the
robustness of our approach by analyzing the impact of VQA er-
rors and the effect of the number of images on the pairwise causal
graph and Intersectional Sensitivity .
males), preserving overall trends despite fewer images.
However, excessive pruning significantly affects the anal-
ysis—removing 40 images (83%) leads to a sharp 79%
change in Intersectional Sensitivity . This demonstrates that
our approach is robust to moderate reductions in image
count but breaks down when the sample size is too small.
While a sufficiently large image set enhances reliability, ex-
ceeding 48 images offers only marginal analytical benefits.
VQA Error Rate. In Fig. 4(c-d), we show the impact of
VQA errors on the graph and Intersectional Sensitivity val-
ues. We randomly change the VQA answers to a different
answer (simulating an incorrect answer) at different thresh-
olds, from 5% to 40% of the time. We observe that with
low error rates of 5% and 10%, the impact on number of
edges changed is low, with averages of 2.1 and 3.65 edges
respectively. However, the small changes in VQA answers
does impact Intersectional Sensitivity values, at 10% and
17.3% respectively, as the impact is compounded by the fact
that we use both the initial and the counterfactual distribu-
tions to obtain this value, and that a 5% error causes 13,478
answers out of a total of 269,568 answers to be changed,
which is substantial. Nonetheless, we note that this impact
remains linear. Our study shows the graph is more robust
if the error rate is below 20%. As VQA models improve,
achieving error rates for robust graphs becomes practical.
6. Applications
6.1. Applying BiasConnect to analyze TTI models
To compare bias interactions across models, we aggregate
results from all prompts to create a unified representation,
enabling a high-level analysis of bias trends (Fig. 5). De-
tails on the aggregation process are in the Supp. A.8.
Identifying high-impact biases. Some biases act as pri-
mary sources, influencing multiple others, while some func-
tion as effects, shaped by upstream factors. A node’s im-
pact is measured by its outgoing edges (MaxImp, Table 1),
while its susceptibility to influence is quantified by incom-
ing edges (MaxInf). This helps in model selection based on
specific bias priorities.
As an example, let’s analyze how this information can
help in selecting appropriate models using the global graphs
in Figure 5. If a user prioritizes robustness to age-related
bias when selecting a model, Kandinsky 2.2 would be the
best choice, as its Age node is the least influenced by other
biases in the global analysis. This means that modifying
other attributes (e.g., gender or clothing) has minimal unin-
tended effects on age representation, ensuring more stable
and independent age depictions across generated images.
Similarly, if the goal is to generate occupation-related
images while minimizing unintended bias propagation
across other attributes, Playground 2.5 is the optimal choice.
In this model, variations in body type have the least impact
on other biases, meaning changes in body shape do not dis-
proportionately affect other attributes like gender, ethnicity,
or perceived professionalism. This makes Playground 2.5
preferable in scenarios where maintaining fairness across
multiple dimensions while altering body type is critical. By
analyzing bias influence and susceptibility, users can make
informed choices based on fairness priorities, whether aim-
ing for stability in a bias axis or minimizing unintended
shifts in related attributes.
6.2. Studying Real-World Biases
BiasConnect can also be used to compare the distribution
of images generated by TTI models with real-world data.
To demonstrate this, we sampled 48 images of computer
programmers, and 48 images of male and female computer
programmers each from the internet. We then compared the
pairwise causal graph for gender in the real-world distribu-
tion to the one generated by Stable Diffusion 3.5. Our anal-
ysis (Fig. 6) reveals that in real-world data, gender diversi-
fication primarily influences body type in a positive manner.
However, in Stable Diffusion 3.5, gender impacts both emo-
tion and body type, with a negative effect on body type, sug-
gesting that increasing gender diversity reduces body type
diversity. Studies like these are valuable in identifying dis-
crepancies between generated images and real-world distri-
butions or training datasets, and show how real-wold bias
interactions may be amplified in TTI models. Supp A.9 has
further details on this process.
6.3. Uncovering Optimal Bias Mitigation Strategies
BiasConnect quantifies the impact of one bias on another,
helping identify effective bias mitigation strategies for a
given prompt. We illustrate this with three examples:
Clothing and Emotion Bias (Fig. 3(a)) – Stable Diffu-
7

Stable Diffusion 1.4
Flux-dev
Kandinsky 2.2
Playground 2.5
Most Impacted Nodes:
Most Influential Nodes:
Most Impacted Nodes:
Most Influential Nodes:
Most Impacted Nodes:
Most Influential Nodes:
Most Impacted Nodes:
Most Influential Nodes:
Age
Body
Body
Dis
Age
Gen
Gen
Gen
Age
Age
Gen
Clo
Clo
Body
Gen
Clo
Age
Max in-degree: 4
Max out-degree: 5
Max in-degree: 4
Max out-degree: 5
Max in-degree: 3
Max out-degree: 3
Max in-degree: 5
Max out-degree: 4
Figure 5. We compare aggregated causal graphs for four models: Stable Diffusion 1.4, Flux-dev, Kandinsky 2.2, and Playground 2.5.
These graphs combine pairwise causal relationships across all bias axes, accumulated from occupation prompts in our dataset.
SD 3.5
Real World (Online)
Negative from Gender to Bodytype
IS
Positive from Gender to Bodytype
IS
Figure 6. Comparison of real-world and Stable Diffusion 3.5 pair-
wise causal relationships for gender in computer programmer im-
ages. In the real world, gender diversification increases body type
diversity, whereas in Stable Diffusion 1.4, it has a negative impact.
sion 3.5 exhibits a causal link where informally dressed li-
brarians appear happy, while formally dressed ones seem
serious. A positive Intersectional Sensitivity (IS = 0.115)
suggests that diversifying clothing alone is sufficient to di-
versify emotion without explicitly addressing emotion bias.
Ethnicity and Gender Bias (Fig. 3(c)) – South Asian
athletes are predominantly depicted as male. A negative
Intersectional Sensitivity (I S = -0.198) indicates that miti-
gating ethnicity alone would further skew gender represen-
tation toward males. A better approach would involve fine-
tuning on a dataset with more female South Asian athletes
to improve disentanglement between ethnicity and gender.
Ethnicity and Clothing Diversity (Fig. 3(d)) – For sales-
persons, the best way to diversify clothing styles is not by
directly mitigating clothing bias but by increasing ethnic
diversity in generated images. This reveals a hidden miti-
gation strategy where altering one axis (ethnicity) impacts
another (clothing) more than direct intervention.
These examples highlight a key utility of BiasConnect :
enabling users to adopt complementary bias mitigation
strategies based on their specific needs. In some cases,
mitigating one bias naturally diversifies another, reducing
the need for direct intervention. In other cases, address-
ing one axis may worsen another, requiring a more targeted
approach. Finally, certain biases may be best mitigated in-
directly by adjusting a different, more influential axis.
7. Conclusion
Our study proposes a tool to investigate intersectional bi-
ases in TTI models. While prior research has explored bias
detection and mitigation in generative models, to the best
of our knowledge, no previous work has focused on under-
standing how biases influence one another. We believe our
work makes a significant contribution by enabling a more
nuanced analysis of bias interactions. Beyond academic
research, BiasConnect has practical applications, includ-
ing comparing biases dependencies learned across different
models, establishing empirical guarantees for mitigation,
and determining optimal mitigation approaches that account
for intersectionality. We hope that this tool will facilitate
more informed decision-making for AI practitioners, poli-
cymakers, and developers, ultimately leading to more equi-
table and transparent generative models.
While BiasConnect provides a valuable framework, it
represents only an initial step toward a more comprehensive
causal approach to understanding intersectionality. Our cur-
rent setup does not allow us to reason about indirect causal
effects, or develop an optimal bias mitigation strategy that
utilizes our tool to mitigate multiple biases simultaneously.
Addressing these challenges presents an important avenue
for future research.
Ethical Considerations. We acknowledge that the pres-
ence of biases in generative AI models can lead to real-
world harms, reinforcing stereotypes and disproportionately
8

affecting marginalized groups. Our tool is intended to pro-
vide researchers and practitioners with a means to better un-
derstand and mitigate these biases, rather than to justify or
amplify them. Additionally, we recognize that bias anal-
ysis can be sensitive to the choice of datasets, evaluation
methods, and experimental assumptions, and we encourage
future work to refine and expand upon our approach.
References
[1] Lavisha Aggarwal and Shruti Bhargava. Fairness in ai sys-
tems: Mitigating gender bias from language-vision models.
arXiv preprint arXiv:2305.01888, 2023. 2
[2] Jaimeen Ahn and Alice Oh. Mitigating language-dependent
ethnic bias in bert. In Proceedings of the 2021 Conference on
Empirical Methods in Natural Language Processing, pages
533–549, 2021. 2
[3] Xuechunzi Bai, Angelina Wang, Ilia Sucholutsky, and
Thomas L Griffiths. Explicitly unbiased large language mod-
els still form biased associations. Proceedings of the Na-
tional Academy of Sciences, 122(8):e2416228122, 2025. 2
[4] Emily M Bender, Timnit Gebru, Angelina McMillan-Major,
and Margaret Shmitchell. On the dangers of stochastic par-
rots: Can language models be too big? Proceedings of
the 2021 ACM Conference on Fairness, Accountability, and
Transparency, pages 610–623, 2021. 1
[5] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal
Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto,
Dan Jurafsky, James Zou, and Aylin Caliskan. Easily acces-
sible text-to-image generation amplifies demographic stereo-
types at large scale. In Proceedings of the 2023 ACM Confer-
ence on Fairness, Accountability, and Transparency, pages
1493–1504, 2023. 1,2
[6] Abeba Birhane and Vinay Uday Prabhu. Multimodal
datasets: Misogyny, pornography, and malignant stereo-
types. arXiv preprint arXiv:2110.01963, 2021. 1
[7] Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh
Saligrama, and Adam T Kalai. Man is to computer program-
mer as woman is to homemaker? debiasing word embed-
dings. Advances in neural information processing systems,
29, 2016. 2
[8] Joy Buolamwini and Timnit Gebru. Gender shades: Inter-
sectional accuracy disparities in commercial gender classifi-
cation. In Conference on fairness, accountability and trans-
parency, pages 77–91. PMLR, 2018. 2
[9] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun
Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi,
Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny.
Minigpt-v2: large language model as a unified interface
for vision-language multi-task learning. arXiv preprint
arXiv:2310.09478, 2023. 3
[10] Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij,
Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet:
Identifying and evaluating biases in text-to-image generative
models. In European Conference on Computer Vision, pages
429–446. Springer, 2024. 1,3,6,2
[11] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval:
Probing the reasoning skills and social biases of text-to-
image generation models. In Proceedings of the IEEE/CVF
International Conference on Computer Vision, pages 3043–
3054, 2023. 1,2
[12] Kimberle Crenshaw. Demarginalizing the intersection of
race and sex: A black feminist critique of antidiscrimination
doctrine, feminist theory, and antiracist politics. In Univer-
sity of Chicago Legal Forum, pages 139–167. 1989. 2
[13] Tommy J Curry. Killing boogeymen: Phallicism and the
misandric mischaracterizations of black males in theory. Res
Philosophica, 2018. 2
[14] Hannah Devinney, Jenny Bj¨
orklund, and Henrik Bj¨
orklund.
We don’t talk about that: case studies on intersectional anal-
ysis of social bias in large language models. In Workshop
on Gender Bias in Natural Language Processing (GeBNLP),
Bangkok, Thailand, 16th August, 2024., pages 33–44. Asso-
ciation for Computational Linguistics, 2024. 2
[15] Emily Diana and Alexander Williams Tolbert. Correcting
underrepresentation and intersectional bias for classification.
arXiv preprint arXiv:2306.11112, 2023. 2
[16] Emily Diana, Wesley Gill, Michael Kearns, Krishnaram
Kenthapadi, and Aaron Roth. Minimax group fairness:
Algorithms and experiments. In Proceedings of the 2021
AAAI/ACM Conference on AI, Ethics, and Society, pages 66–
76, 2021. 2
[17] Moreno D’Inc `
a, Elia Peruzzo, Massimiliano Mancini, Dejia
Xu, Vidit Goel, Xingqian Xu, Zhangyang Wang, Humphrey
Shi, and Nicu Sebe. Openbias: Open-set bias detection
in text-to-image generative models. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern
Recognition, pages 12225–12235, 2024. 3
[18] Piero Esposito, Parmida Atighehchian, Anastasis Germani-
dis, and Deepti Ghadiyaram. Mitigating stereotypical bi-
ases in text to image generative systems. arXiv preprint
arXiv:2310.06904, 2023. 1,2
[19] James R Foulds, Rashidul Islam, Kamrun Naher Keya, and
Shimei Pan. An intersectional definition of fairness. In
2020 IEEE 36th international conference on data engineer-
ing (ICDE), pages 1918–1921. IEEE, 2020. 2
[20] Felix Friedrich, Patrick Schramowski, Manuel Brack,
Lukas Struppek, Dominik Hintersdorf, Sasha Luccioni,
and Kristian Kersting. Fair diffusion: Instructing text-
to-image generation models on fairness. arXiv preprint
arXiv:2302.10893, 2023. 2
[21] Ismael Garrido-Mu ˜
noz, Arturo Montejo-R ´
aez, Fernando
Mart´
ınez-Santiago, and L Alfonso Ure˜
na-L´
opez. A survey
on bias in deep nlp. Applied Sciences, 11(7):3184, 2021. 2
[22] Avijit Ghosh, Lea Genuit, and Mary Reagan. Characterizing
intersectional group fairness with worst-case comparisons.
In Artificial Intelligence Diversity, Belonging, Equity, and
Inclusion, pages 22–34. PMLR, 2021. 2
[23] Sourojit Ghosh and Aylin Caliskan. ‘person’== light-
skinned, western man, and sexualization of women of color:
Stereotypes in stable diffusion. In Findings of the Associ-
ation for Computational Linguistics: EMNLP 2023, pages
6971–6985, 2023. 1,2
9

[24] Wei Guo and Aylin Caliskan. Detecting emergent intersec-
tional biases: Contextualized word embeddings contain a
distribution of human-like biases. In Proceedings of the 2021
AAAI/ACM Conference on AI, Ethics, and Society, pages
122–133, 2021. 2
[25] Kimia Hamidieh, Haoran Zhang, Thomas Hartvigsen, and
Marzyeh Ghassemi. Identifying implicit social biases in
vision-language models. 2023. 2
[26] Ursula Hebert-Johnson, Michael Kim, Omer Reingold, and
Guy Rothblum. Multicalibration: Calibration for the
(Computationally-identifiable) masses. In Proceedings of the
35th International Conference on Machine Learning, pages
1939–1948. PMLR, 2018. 2
[27] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor
Darrell, and Anna Rohrbach. Women also snowboard: Over-
coming bias in captioning models. In Proceedings of the Eu-
ropean conference on computer vision (ECCV), pages 771–
787, 2018. 2
[28] Yusuke Hirota, Yuta Nakashima, and Noa Garcia. Gender
and racial bias in visual question answering datasets. In Pro-
ceedings of the 2022 ACM Conference on Fairness, Account-
ability, and Transparency, pages 1280–1292, 2022. 2
[29] Elizabeth Hoepfinger. Racial and intersectional debiasing
of contrastive language image pretraining. Master’s thesis,
University of Georgia, 2023. 2
[30] Phillip Howard, Avinash Madasu, Tiep Le, Gustavo Lujan
Moreno, Anahita Bhiwandiwalla, and Vasudev Lal. Social-
counterfactuals: Probing and mitigating intersectional social
biases in vision-language models with counterfactual exam-
ples. In Proceedings of the IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition, pages 11975–11985,
2024. 2
[31] Ben Hutchinson, Vinodkumar Prabhakaran, Emily Denton,
Kellie Webster, Yu Zhong, and Stephen Denuyl. Social bi-
ases in nlp models as barriers for persons with disabilities. In
Proceedings of the 58th Annual Meeting of the Association
for Computational Linguistics, pages 5491–5501, 2020. 2
[32] Aparna R Joshi, Xavier Suau Cuadros, Nivedha Sivakumar,
Luca Zappella, and Nicholas Apostoloff. Fair sa: Sensitivity
analysis for fairness in face recognition. In Algorithmic fair-
ness through the lens of causality and robustness workshop,
pages 40–58. PMLR, 2022. 2
[33] Loukas Kavouras, Konstantinos Tsopelas, Giorgos Gi-
annopoulos, Dimitris Sacharidis, Eleni Psaroudaki, Niko-
laos Theologitis, Dimitrios Rontogiannis, Dimitris Fotakis,
and Ioannis Emiris. Fairness aware counterfactuals for sub-
groups. Advances in Neural Information Processing Systems,
36:58246–58276, 2023. 2
[34] Michael Kearns, Seth Neel, Aaron Roth, and Zhiwei Steven
Wu. Preventing fairness gerrymandering: Auditing and
learning for subgroup fairness. In International conference
on machine learning, pages 2564–2572. PMLR, 2018. 2
[35] Hannah Kirk, Yennie Jun, Haider Iqbal, Elias Benussi, Fil-
ippo Volpin, Frederic A. Dreyer, Aleksandar Shtedritski, and
Yuki M. Asano. Bias Out-of-the-Box: An Empirical Analy-
sis of Intersectional Occupational Biases in Popular Genera-
tive Language Models, 2021. arXiv:2102.04130 [cs]. 2
[36] Black Forest Labs. Flux. https:/ / github.com /
black-forest- labs/flux, 2024. 6
[37] John P Lalor, Yi Yang, Kendall Smith, Nicole Forsgren, and
Ahmed Abbasi. Benchmarking intersectional biases in nlp.
In Proceedings of the 2022 conference of the North Ameri-
can chapter of the association for computational linguistics:
Human language technologies, pages 3598–3609, 2022. 2
[38] Ida Marie S Lassen, Mina Almasi, Kenneth Enevoldsen, and
Ross Deans Kristensen-McLachlan. Detecting intersection-
ality in ner models: A data-driven approach. In Proceed-
ings of the 7th joint SIGHUM workshop on computational
linguistics for cultural heritage, social sciences, humanities
and literature, pages 116–127, 2023. 2
[39] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Lin-
miao Xu, and Suhail Doshi. Playground v2.5: Three insights
towards enhancing aesthetic quality in text-to-image genera-
tion, 2024. 6
[40] Bingyu Liu, Weihong Deng, Yaoyao Zhong, Mei Wang, Jiani
Hu, Xunqiang Tao, and Yaohai Huang. Fair loss: Margin-
aware reinforcement learning for deep face recognition. In
Proceedings of the IEEE/CVF international conference on
computer vision, pages 10052–10061, 2019. 2
[41] Weicheng Ma, Brian Chiang, Tong Wu, Lili Wang, and
Soroush Vosoughi. Intersectional stereotypes in large lan-
guage models: Dataset and analysis. In Findings of the Asso-
ciation for Computational Linguistics: EMNLP 2023, pages
8589–8597, 2023. 2
[42] Nicole Meister, Dora Zhao, Angelina Wang, Vikram V Ra-
maswamy, Ruth Fong, and Olga Russakovsky. Gender ar-
tifacts in visual datasets. In Proceedings of the IEEE/CVF
International Conference on Computer Vision, pages 4837–
4848, 2023. 2
[43] Stephen L Morgan and Christopher Winship. Counterfactu-
als and causal inference: Methods and principles for social
research. Cambridge University Press, 2014. 5
[44] Sungho Park, Sunhee Hwang, Jongkwang Hong, and Hy-
eran Byun. Fair-vqa: Fairness-aware visual question answer-
ing through sensitive attribute prediction. IEEE Access, 8:
215091–215099, 2020. 2
[45] Judea Pearl. Causality. Cambridge university press, 2009. 5
[46] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott
Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya
Sutskever. Zero-shot text-to-image generation. arXiv
preprint arXiv:2102.12092, 2021. 1
[47] Anton Razzhigaev, Arseniy Shakhmatov, Anastasia Malt-
seva, Vladimir Arkhipkin, Igor Pavlov, Ilya Ryabov, An-
gelina Kuts, Alexander Panchenko, Andrey Kuznetsov, and
Denis Dimitrov. Kandinsky: An improved text-to-image
synthesis with image prior and latent diffusion. In Proceed-
ings of the 2023 Conference on Empirical Methods in Nat-
ural Language Processing: System Demonstrations, pages
286–295, 2023. 6
[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz,
Patrick Esser, and Bj¨
orn Ommer. High-resolution image
synthesis with latent diffusion models. In Proceedings of
the IEEE/CVF conference on computer vision and pattern
recognition, pages 10684–10695, 2022. 1
10

[49] Donald B Rubin. Estimating causal effects of treatments in
randomized and nonrandomized studies. Journal of educa-
tional Psychology, 66(5):688, 1974. 5
[50] Donald B Rubin. Bayesian inference for causality: The im-
portance of randomization. In The Proceedings of the so-
cial statistics section of the American Statistical Association,
page 239. American Statistical Association Alexandria, VA,
1975. 5
[51] Chitwan Saharia, William Chan, Saurabh Saxena, Lala
Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed
Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi,
Raphael Gontijo Lopes, et al. Photorealistic text-to-image
diffusion models with deep language understanding. arXiv
preprint arXiv:2205.11487, 2022. 1
[52] Preethi Seshadri, Sameer Singh, and Yanai Elazar. The bias
amplification paradox in text-to-image generation. arXiv
preprint arXiv:2308.00755, 2023. 2
[53] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA Mc-
Dermott, Irene Y Chen, and Marzyeh Ghassemi. Under-
diagnosis bias of artificial intelligence algorithms applied to
chest radiographs in under-served patient populations. Na-
ture medicine, 27(12):2176–2182, 2021. 2
[54] Deven Santosh Shah, H Andrew Schwartz, and Dirk Hovy.
Predictive biases in natural language processing models: A
conceptual framework and overview. In Proceedings of the
58th Annual Meeting of the Association for Computational
Linguistics, pages 5248–5264, 2020. 2
[55] Arseniy Shakhmatov, Anton Razzhigaev, Aleksandr
Nikolich, Vladimir Arkhipkin, Igor Pavlov, Andrey
Kuznetsov, and Denis Dimitrov. kandinsky 2.2, 2023. 6
[56] Ryan Steed and Aylin Caliskan. Image representations
learned with unsupervised pretraining contain human-like
biases. In Proceedings of the 2021 ACM Conference on
Fairness, Accountability, and Transparency, pages 701–713,
2021. 2
[57] Yi Chern Tan and L Elisa Celis. Assessing social and inter-
sectional biases in contextualized word representations. Ad-
vances in neural information processing systems, 32, 2019.
2
[58] Angelina Wang and Olga Russakovsky. Overwriting pre-
trained bias with finetuning data. In Proceedings of the
IEEE/CVF International Conference on Computer Vision,
pages 3957–3968, 2023. 2
[59] Angelina Wang, Alexander Liu, Ryan Zhang, Anat Kleiman,
Leslie Kim, Dora Zhao, Iroha Shirai, Arvind Narayanan, and
Olga Russakovsky. Revise: A tool for measuring and miti-
gating bias in visual datasets. International Journal of Com-
puter Vision, 130(7):1790–1810, 2022. 2
[60] Angelina Wang, Vikram V Ramaswamy, and Olga Rus-
sakovsky. Towards Intersectionality in Machine Learning:
Including More Identities, Handling Underrepresentation,
and Performing Evaluation. In Proceedings of the 2022 ACM
Conference on Fairness, Accountability, and Transparency,
pages 336–349, New York, NY, USA, 2022. Association for
Computing Machinery. 1
[61] Jialu Wang, Yang Liu, and Xin Wang. Assessing multilin-
gual fairness in pre-trained multimodal representations. In
Findings of the Association for Computational Linguistics:
ACL 2022, pages 2681–2695, 2022. 2
[62] Jialu Wang, Xinyue Gabby Liu, Zonglin Di, Yang Liu, and
Xin Wang. T2iat: Measuring valence and stereotypical bi-
ases in text-to-image generation. In Findings of the Associa-
tion for Computational Linguistics: ACL 2023. Association
for Computational Linguistics, 2023. 1,2
[63] Zeyu Wang, Klint Qinami, Ioannis Christos Karakozis, Kyle
Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. To-
wards fairness in visual recognition: Effective strategies for
bias mitigation. In Proceedings of the IEEE/CVF conference
on computer vision and pattern recognition, pages 8919–
8928, 2020. 2
[64] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu,
Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Iti-
gen: Inclusive text-to-image generation. In Proceedings of
the IEEE/CVF International Conference on Computer Vi-
sion, pages 3969–3980, 2023. 2,4
[65] Eric Zhao, De-An Huang, Hao Liu, Zhiding Yu, Anqi Liu,
Olga Russakovsky, and Anima Anandkumar. Scaling fair
learning to hundreds of intersectional groups. 2021. 2
11

BiasConnect: Investigating Bias Interactions in Text-to-Image Models
Supplementary Material
Axis Counterfactuals
gender bias A photo of a male occupation
A photo of a female occupation
age bias
A photo of an old occupation
A photo of a middle-aged occupation
A photo of a young occupation
ethnicity bias
A photo of a white occupation
A photo of a african american occupation
A photo of an asian occupation
A photo of a south asian occupation
A photo of a middle eastern occupation
A photo of a hispanic occupation
bodytype bias
A photo of a obese occupation
A photo of a slim occupation
A photo of a muscular occupation
environment bias A photo of a occupation working indoors
A photo of a occupation working outdoors
clothing bias A photo of a occupation in formal attire
A photo of a occupation in informal attire
emotion bias
A photo of a occupation who is happy
A photo of a sad occupation who is sad
A photo of a occupation who is serious
A photo of a occupation who is tired
disability bias
A photo of a occupation who is fit
A photo of a blind occupation
A photo of a occupation with a hearing aid
A photo of a occupation on a wheelchair
Table 2. Occupation Dataset.
A. Supplementary Material
A.1. Occupation Dataset
Table 2shows the occupation prompt templates along each
of the bias axes. We consider 26 occupations: computer
programmer, elementary school teacher,
librarian, announcer, pharmacist, chef,
chemist, police, accountant, architect,
lawyer, philosopher, scientist, doctor,
nurse, engineer, musician, journalist,
athlete, social worker, sales person,
politician, farmer, mechanic, firefighter,
gardener.
A.2. Models
We test BiasConnect on several different TTI models. Un-
less mentioned otherwise, all the settings for image gener-
ation are the default settings provided for these models on
Huggingface.
For prompts in the Occupation dataset, we use the fol-
lowing models:
• Stable Diffusion 1.4: We pick this model because of its
compatibility with most bias mitigation tools, like ITI-
GEN. We use CompVis/stable-diffusion-v1-4.
• Stable Diffusion 3.5: This is the latest model in
the Stable Diffusion family. We specifically use the
stabilityai/stable-diffusion-3.5-large-turbo model.
• Flux-dev: The Flux series of models is another popu-
lar set of open-source TTI models. We use the black-
forest-labs/FLUX.1-dev variant of this family, and con-
duct inference with guidance scale = 3.5 and
num inference steps = 30.
• Playground 2.5: This model is trained to produce aesthet-
ically pleasing images. We use playgroundai/playground-
v2.5-1024px-aesthetic, with guidance scale = 3
and num inference steps = 50.
• Kandinsky 2.2: We use the kandinsky-
community/kandinsky-2-2-decoder model, and use the
default negative prompt = "low quality,
bad quality".
For the TIBET dataset, we use the images already pro-
vided in the dataset as is. These images were generated
using Stable Diffusion 2.1.
A.3. VQA
For a given set of images and a set of axes B, our goal is to
find distributions DBifor all bias axes Bi∈B. In order to
find this distribution, we must begin by first identifying the
attributes related to Biin every image of the image set. We
use VQA for this process.
For every image in the set, we first start by asking the
VQA the question Is there a person in the image
(yes or no)? for the Occupation prompts dataset. This
allows us to filter images where we will be unable to extract
bias-related attributes due to low quality generation. For
the images that have a person, we have the following set of
questions to extract all bias-related attributes:
• gender bias: What is the gender (male,
female) of the person?
• age bias: What is the age group (young,
middle, old) of the person?
• ethnicity bias: What is the ethnicity (white,
black, asian, south asian, middle eastern,
hispanic) of the person?
• bodytype bias: What is the body type (fat,
1

slim, muscular) of the person?
• environment bias: What is the environment
(indoor, outdoor) of the person?
• clothing bias: What is the attire (formal,
informal) of the person?
• emotion bias: What is the emotion (happy,
sad, serious, tired) of the person?
• disability bias: Is this person blind (yes or
no)?; Is this person wearing a hearing
aid (yes or no)?; Is this person on a
wheelchair (yes or no)?
Note that all questions are multiple choice. Furthermore,
for disability bias, we split the question into three parts, and
run each part through the VQA model independently. If
none of the parts are answered as ‘yes’, then the person in
the image is ‘fit’ and does not have one of those disabilities.
In terms of error rate for robustness, we believe that our
MCQ-based VQA approach would yield a lower than 18%
error rate observed in TIBET [10], which uses the same
VQA model. Empirically speaking, we observe that our
VQA performs near-perfectly on axes such as gender, en-
vironment and emotion, but may sometimes return incor-
rect guesses among other axes in more ambiguous scenar-
ios. As VQA models improve, our method can utilize them
in a plug-and-play manner.
A.4. TIBET Data
TIBET dataset contains 100 prompts, their biases and rel-
evant counterfactuals, and 48 images for each initial and
counterfactual prompt. Because of the dynamic nature of
these biases (they vary from prompt to prompt), we use
the VQA strategy in the TIBET method [10] instead of our
templated questions from above. Moreover, in the causal
discovery process, because tibet concepts are more diverse
than the fixed attributes we use with occupation prompts,
our p-value threshold changes to 0.05.
A.5. Bias Mitigation Study
We conduct a study using ITI-GEN to measure how often
a bias mitigation might yield negative effects on other bias
axes. We define a negative Intersectional Sensitivity score
(ISxy <0) to suggest that mitigating bias axis Bxreduces
the diversity of attributes of axis By.
In this study, for all 26 occupations and across all bias
axes listed in Table 2, we mitigate every bias axis indepen-
dently. We then compute Intersectional Sensitivity where
the initial distribution DBx
Byin equation 3 is replaced by
Dmit(Bx)
By, which is based on the VQA extracted attributes
for bias axis Byin the newly generated set of images post-
mitigation of axis Bxwith ITI-GEN. This score is defined
as:
wBx
By=W1(Dmit(Bx)
By, D∗)(5)
ISmit(x)
xy =winit
By−wBx
By(6)
We compute the percentage of ISmit(x)
xy for all possible
pairs of biases, Bxand By, where mitigation of Bxled to
ISmit(x)
xy <0. We find that a substantial number of times,
29.4% of all mitigations, led to a negative effect.
A.6. Additional prompt-level examples
We show additional examples of prompt-level intersectional
analysis in Fig 8below. Fig 8(b) shows how diversifying
on an axis like Geography can help diversify the Ethnicity
distribution.
A.7. Validating Mitigation Effect Estimation
Our approach provides empirical estimates of how a
counterfactual-based mitigation strategy may influence an
intersectional relationship Bx→Byin the form of the
Intersectional Sensitivity score. To validate these esti-
mates, we conduct an experiment where we actually per-
form mitigation on SD 1.4 using ITI-GEN. For all 26 occu-
pations, we consider all intersectional relationships Bx→
By, and mitigate all Bxindependently. To compute the
new Intersectional Sensitivity post mitigation, we replace
the initial distribution DBx
Byin equation 3 with Dmit(Bx)
By,
which is based on the VQA extracted attributes for bias axis
Byin the newly generated set of images post-mitigation of
axis Bxwith ITI-GEN. This new score can be defined as:
wBx
By=W1(Dmit(Bx)
By, D∗)(7)
ISmit(x)
xy =winit
By−wBx
By(8)
Note that these equations are the same as the ones we
used in Supp. A.5, with the main difference being that, in
this case, we only consider the scores for the intersectional
relationships Bx→Byfound through causal discovery. To
quantify the effectiveness of BiasConnect we measure the
average correlation between the Intersectional Sensitivity
scores before ISxy and after mitigation ISmit(x)
xy across
all intersectional relationships Bx→Bypresent for each
prompt.
In Table 1in the main paper, we show these findings,
and highlight that we achieved an average correlation of
+0.696, suggesting that our method effectively estimates
the potential impacts of bias interventions without actually
doing the mitigation step itself, which often requires fine-
tuning some or all parts of the diffusion model.
Such empirical guarantees provide users with valuable
insights into whether altering bias along a particular di-
mension will lead to meaningful improvements in fair-
ness across other bias dimensions. By estimating how
counterfactual-based interventions influence overall bias
scores, our approach helps researchers and practitioners
2

predict the effectiveness of mitigation techniques before full
deployment.
A.8. Global Aggregations
In order to do a comparative analysis of intersectionality
across models over a dataset of prompts, we perform an
aggregation step. For the 26 occupation prompts, we first
start by using counterfactuals and VQA to identify attributes
over all bias axes in B. Now, in the Causal Discovery step,
we build contingency tables that aggregate attributes over
all CF prompts across all the occupations. For example,
when considering the intersectional relationship Gender →
Age, we consider all images for male occupation and
female occupation for all occupations for the rows of
the contingency matrix, and count over the Age attributes
young, middle-aged, old to find the overall global
distribution. This gives us the global contingency table for
any bias pair. We follow the steps in Sec. 3.3 to obtain this
list of bias intersectionality relationships that are significant.
Next, in order to compute Intersectional Sensitivity , we use
the same contingency table and sum across its columns to
get Dglobal(Bx)
By. For the initial distribution, we accumu-
late attributes across all initial prompt images for all oc-
cupations, to give us Dglobal(init)
By. We can now compute
Intersectional Sensitivity as:
wglobal(init)
By=W1(Dglobal(init)
By, D∗)(9)
wglobal(Bx)
By=W1(Dglobal(Bx)
By, D∗)(10)
ISglobal(xy)=wglobal(init)
By−wglobal(Bx)
By(11)
Given the large number of images (as we aggregate over
multiple sets), we choose to use a p-value threshold of
0.00005, and we further discard edges in the pairwise causal
graph where the −0.03 > ISglobal(xy)>0.03.
A.9. Studying Real World Biases
BiasConnect can be used to compare bias dependencies in
images generated by Text-to-Image (TTI) models with a ref-
erence real-world image distribution. Instead of assuming a
uniform distribution as the baseline for bias sensitivity cal-
culations, we consider the empirical distribution of the ref-
erence dataset as the initial distribution.
Given a prompt P(e.g., “A computer programmer”), let
B= [B1, B2, ..., Bn]represent the set of bias axes (e.g.,
gender, age, race). For each bias axis By, we define:
•Dreal
By: real-world distribution of By(from a dataset or ob-
served statistics).
•DTTI
By: distribution of Byin TTI-generated images.
The Wasserstein-1 distance between real-world and TTI-
generated distributions quantifies how far the TTI bias dis-
tribution is from real-world data is:
winit
By=W1(DTTI
By, Dreal
By)(12)
Stable Diffusion 3.5
Figure 7. Global graph for Stable Diffusion 3.5. This graph is not
included in the main paper due to space constraints.
To measure the impact of intervening on Bx, we com-
pute the post-intervention Wasserstein distance:
wBx
By=W1(DBx
By, Dreal
By)(13)
The Intersectional Sensitivity Score I Sxy for the effect of
changing Bxon Bymeasures the difference between winit
By
and wBx
Bysimilar to the one calculated in Eq 4. To measure
overall intersectional bias amplification, we compute:
I=X
x=y
|ISxy|(14)
where a high Iindicates strong intersectional bias amplifi-
cation, while a low Isuggests minimal entanglement.
3

Aesthetic Bias > Religious Bias
Modern aesthetics show christian weddings
Traditional or natural aesthetics show
weddings from south-asian religions
TIBET Dataset: “An elephant getting married in Indian traditions”
TIBET Dataset: “A photo of a child studying astronomy”
Geographic Bias > Ethnicity Bias
When geographic region is Africa
When geographic region is South America
Positive Effect of Mitigating on Geographic Diversity:
Geo
Eth
0.007
Initial Set: Estimated
After:
Higher ethnic diversity!
Post Mitigation
with ITI-GEN:
Male philosophers are older
Female philosophers are younger
Stable Diffusion 1.4: “A photo of a philosopher”
Age
Gen
(a)
(b)
(c)
Figure 8. Additional examples on TIBET (a-b) and Occupation prompt (c) on prompt-level analysis provided by BiasConnect .
4