Content uploaded by Francisco Olivos
Author content
All content in this area was uploaded by Francisco Olivos on Mar 27, 2024
Content may be subject to copyright.
1
Preprint
AI-Replicas as Ethical Practice: Introducing an Alternative to Traditional
Anonymization Techniques in Image-Based Research
Tobias Kamelskiab
Francisco Olivosa
aDepartment of Sociology and Social Policy, Lingnan University
February 11, 2024
Word Count: 8,226
bCorresponding author:
Tobias Kamelski, Lingnan University, 8 Castle Peak Road, Tuen Mun, Hong Kong.
Email: tobiaskamelski@ln.hk
2
Abstract
This paper introduces the use of AI-replicas as an alternative to traditional anonymization methods in image-based
qualitative research. It emphasizes the ethical and practical dilemmas posed by current anonymization methods,
such as distortion or loss of emotional and contextual information in images, and proposes the use of AI-replicas
to preserve the integrity and authenticity of visual data while ensuring participant anonymity. The paper outlines
the technological foundations of generative artificial intelligence (AI) and the practical application of Stable
Diffusion to generate AI-replicas for anonymization and fictionalization purposes. Furthermore, it discusses the
potential biases present in generative AI to suggest ways to mitigate these biases through careful prompt
engineering and participatory approaches. The introduced approach aims to enhance ethical practices in visual
research by providing a method that ensures participant anonymity without compromising the data's qualitative
richness and interpretative validity.
Keywords: Image Anonymization, Generative AI, Stable Diffusion, Research Methodology, Qualitative
Research, Visual Research, Ethics, Privacy
3
AI-Replicas as Ethical Practice: Introducing an Alternative to Traditional
Anonymization Techniques in Image-Based Research
1. Ethical and Research-Practical Challenges in Using Images for Qualitative Research
Over two decades ago, researchers in the field of visual research began discussing how ethical judgments arise
within their field in response to a call for situated visual ethics (e.g. Prosser 2000). At present, despite being
contested and unresolved (Wiles, Coffey, Robinson, et al. 2012), the debate has bourgeoned (Allen 2015; Hannes
and Parylo 2014; Miller 2015; Nutbrown 2010; Spencer 2020; Wiles, Coffey, Robinson, et al. 2012; Wiles,
Coffey, Robison, et al. 2012), and it is expected to grow due to the ‘blossoming of possibilities’ using found
images (Harris 2016; Rose 2012) from image-based online platforms, online archives, YouTube and online video
games, amongst others. More recently, the COVID-19 pandemic and the transition to remote platforms to
communicate across time and space have fostered the growth of visual research across fields (e.g. Polat, 2022;
MacEntee et al.; 2022; Burgos-Thorsen & Munk, 2023).
Against this background, one of the central debates in this pervasive ethical discussion in visual research
methods is anonymization as a default condition and a key concern of researchers (Wiles, Coffey, Robinson, et
al. 2012) and ethics committees (Allen 2015). The anonymity of sources is a condition of the ethical standard of
confidentiality: ‘Visual researchers do not disclose […] identifiable information concerning their research
participants’ (Papademas and IVSA 2009:254). Researchers who use research-generated, participant-generated or
found images are often mandated to ensure the anonymity of the research participants when navigating ethics
committees or dissemination gatekeepers (Miller 2015). The same requirement is also applied to researchers
conducting video analysis, who also ‘transcribe’ videos into images for reporting purposes (e.g. Gan et al., 2020;
Laurier, 2014). Recognizing the limitations and ethical dilemmas posed by traditional anonymization techniques
in visual research, this article aims to introduce and evaluate the potential of generative artificial intelligence (AI)
for creating AI-replicas. By leveraging contemporary technologies, we seek to enhance the process of
anonymizing visual data in qualitative research, demonstrating how novel generative AI technology can offer
novel approaches to image anonymization and fictionalization.
Visual researchers report that ethical reviewers require the distortion of images beyond recognition, even
after research participants consented to use their unmodified images (Spencer 2020). Similarly, collecting consent
or directly identifying picture producers are not always possible (Kozinets, 2020: 164; Tiidenberg and Baym,
2017). Image data, such as found footage or online data originating from platforms without a real-name policy,
are common in fields such as communication science or netnography (see Kozinets, 2020: 164). Such data often
cannot be traced back to producers who can be asked for permission; however, they may still contain personally
identifiable information about the producer or other individuals who need to be anonymized. These concerns can
be explained by the longevity of images in public circulation and the potential reutilization of these images as
found images in the future, which makes standards of anonymization even more salient (Wiles, Coffey, Robinson,
et al. 2012). These difficulties shape research practices because researchers may avoid engaging in image-based
research anticipating these challenges (Miller 2015; Nutbrown 2010), and participants in visual participatory
research may adapt their behaviour to prevent ethical breaches (Hannes and Parylo 2014).
Nevertheless, apart from discouraging researchers and shaping research practices, anonymity
requirements are in tension with other aspects of visual research (Allen 2015). The analysis and scientific
communication of images usually rely on the formulation of ‘protocol sentences’ or ‘basic statements’ (Bohnsack,
2008: 3; Popper, 1959: 43, 59ff.) as textual descriptions of images. These texts represent a fundamental layer of
abstraction, as every analysis is presented based (fully or partially) on textual reductions and not the visual data
itself. This scenario means that a text-mediated analysis of images always involves a reduction and abstraction
that can approximate only the perceptive gaze or ‘seeing view’ (sehendes Sehen) of the interpreter (Bohnsack,
2008: 23). The perceptive gaze refers to the observer’s ability to engage with the meaning of a picture on a deeper
level and is conceptualized in opposition to the ‘recognising gaze’, in which the beholder simply acknowledges
what is objectively depicted by a picture (see Imdahl, 1996). Given this abstracting quality, preserving
4
representative elements and the argumentative plausibility of images as self-referential systems are inherent
problems in these protocol sentences (Luhmann, 1995). Protocol sentences are ideally presented alongside the
pictures they describe to verify their accuracy. This approach inevitably poses problems when reporting
anonymised image data, as anonymization can impede the verification of the presented analysis and possible
interpretations, as pivotal aspects might become unrecognizable (Nutbrown, 2011: 8). Therefore, anonymization
(e.g. adding black bars, blurring features or applying abstraction filters) must be applied to image features that are
irrelevant or at least less relevant to the research questions. For example, applying black bars over a subject’s eyes
is less relevant if the latter’s gaze is less important for the interpretation. Researchers and their audience, therefore,
engage to a certain degree in a ‘good-faith relationship’ in which the anonymized elements are accepted to be
irrelevant for interpretation or to be compensated for by the textual description.
Moreover, techniques that modify images present other challenges to image-based research. Firstly,
pixelation, blurring, black-out strips, cropping and other fictionalization techniques (Jordan 2014; Nutbrown
2010; Wiles, Coffey, Robinson, et al. 2012) silence research participants, which contradicts the visual research’s
potential ethical aim of giving voice and empowering often unheard groups, such as children or marginalized
populations. As explained by Allen (2015), blurring pictures in research on sexual and gender minorities can
undermine the politics of naming and violate their identity rights. Thus, anonymity might contradict the principles
of respect and dignity of research participants. The arguments against anonymization mainly rely on this tension
between researchers’ paternalism and respondents’ agency (for a summary, see Wiles, Coffey, Robinson, et al.
2012). Secondly, the anonymization techniques used by visual researchers could also bring epistemological
challenges. Although researchers must protect research participants at all costs, traditional fictionalization
techniques could move us away from the original truth (Nutbrown 2010) and introduce additional distortions to
what is already generated by and through the researcher’s interpretations and subsequent textual abstractions.
These techniques add additional layers of obscurity between what is analysed and what is reported; these scenarios
could negatively affect the integrity and validity of visual methods (Allen 2015; Nutbrown 2010). For instance,
emotions that are expressed through facial expressions or body language cannot be reported to the audience if the
images are blurred, which could hinder the validity of the interpretation because it is not possible to present the
evidence. In an extreme case, the concern for confidentiality may entail the anonymisation of locations, which is
more complex than what can be achieved with traditional methods. Researchers cannot insert black strips all over
an image to anonymise a school in image-based research with children; simply, they cannot report such images.
These concerns about the validity and reliability of the findings based on anonymized images not only affect what
can be or cannot be reported in a conference presentation but also make journal editors and peer reviewers hesitant
about publishing the studies.
An additional epistemological challenge is anticipating criticism from ethical review boards regarding
the use of images. Visual researchers not only engage in significant self-censorship in the dissemination of results
(Wiles, Coffey, Robison, et al. 2012) but also in avoiding research questions that could be more sensitive for
ethical review boards. Similar effects have been reported for photovoice participants who adopted different coping
strategies (Hannes and Parylo, 2014). Individuals could avoid certain conflict-ridden situations or affect their
behaviour by shooting landscapes but not by taking photos of themselves or by taking photos of friends even if
they are motivated to do otherwise. The tensions surrounding how data can be generated and reported have direct
epistemological consequences when researchers attempt to acquire knowledge through the voices of their
participants. Therefore, traditional anonymization techniques could bias the process of generating knowledge in
visual research as early as the definition of a research question or topic of interest.
Thirdly, anonymization techniques are rarely effective and can even produce undesired results
(MacIntosh 2006). As described by Allen (2015), in her experience with photo-anonymization in school-based
visual research, the cropping of pictures could direct attention to the bodies of students; thus, young people become
worried about sexualization and exploitation. Banks (2001) also suggested that facial modifications such as black
bars could invoke the association of anonymized subjects with criminal perpetrators due to common media
connotations. Similarly, the pixelation of faces could be associated with victimhood.
5
Beyond these potential unintended consequences and epistemological and ethical challenges, offering
anonymity to research participants is incontestable, and visual researchers must protect research participants even
if this entails compromising their research (Nutbrown 2010). In this context, this research leverages novel
computational methods and introduces generative AI to create AI-replicas as an alternative to traditional
fictionalization techniques for image anonymization. We aim to assist in research implementing this technique to
disseminate high-quality image data with strict observance of ethical principles. To the best of our knowledge,
this novel AI technique has not been identified amongst the ‘good ethical practices in relation to visual methods’
(Wiles, Coffey, Robinson, et al. 2012:42). This technique enables us to anonymize images whilst retaining
distinctive key elements for reporting meaningful data. Its use covers scenarios such as reporting on online data
in which gathering informed consent is difficult or practically impossible, reporting visual material that shows
vulnerable subjects, for example, political refugees or simply as an additional layer of anonymization.
Furthermore, this guide can facilitate the application of generative AI in participatory visual research where
subjects can undertake the creation of AI-replicas themselves. This can be imagined as an augmented photovoice
method (see also Wang and Burris, 1997), in which participants create and authorize AI-Replicas as vehicles to
express underlying structures of knowledge. As we will explain, AI can be used to create detailed and accurate
visual representations based on textual descriptions. This approach enables participants to generate visual support
for otherwise textual accounts.
Chapter 2 of this article further elaborates on the potential application of AI-replicas as an innovative
alternative to traditional anonymization methodologies whilst also elaborating on the technological foundations
of this method. Chapter 3 exemplifies the generative process and the ability to control the generation of AI-replicas
based on several example cases. Chapter 4 discusses AI-replicas and their effect on data richness and potential
bias embedded in the models that power generative AI.
2. Generative AI and AI-Replica as Ethical Solutions
2.1 Application Potential of AI-Replicas
The AI-replicas were inspired by Markham’s (2012) ethical fabrication method in qualitative research. In the
context of textual data, Markham suggested creating composite accounts by choosing ‘representative elements
from the dataset and composing a new original not traceable back to the originals’ (p. 7) and devising fictional
narratives that preserve the original data’s argumentative plausibility (p. 11). The AI-replicas contain computer-
generated versions of the original images, preserving crucial elements such as composition, colour and gestures
and essentially replacing images of subjects with realistic, fabricated replicas. These can be placed on a spectrum
between fictionalized abstractions and composite accounts on one end and more closely aligned anonymizations
on the other end. Similar to an artist reimagining a scene (e.g. courtroom sketches), AI-replicas retain distinctive
key elements, ensuring visual authenticity and anonymity. Utilizing deep learning and AI-powered methods is not
entirely new and has been tested in medical contexts (Yang et al., 2022) or when dealing with ‘computer vision
tasks requiring real-world data’ (Douglas, 2022), for example, for processing Google Street View Data.
Generative adversarial networks (GANs) offer sophisticated alternatives to classical ammonization methods by
generating realistic yet artificial faces and even bodies and superimposing them on a given picture (Hukkelås and
Lindseth, 2022; Maximov et al., 2020). The concept of AI-replicas introduced in this work is based on diffusion
models and differs from these anonymization methods in that they enable users to anonymize more than simply
human actors whilst enabling the creation of abstractions based on original data. Thus, whether AI-replicas are
employed as an advanced form of anonymization that aligns closely with the source material or as composite
accounts remains unclear. The latter can be of particular use when reporting typologies developed based on image
data or for the visualization of otherwise textual accounts.
AI-replicas are conceived as a means to report image data in a qualitative research context and must
therefore not be mistaken for a replacement of the latter throughout the initial analysis and interpretation. Similar
to the aforementioned protocol sentences, the AI-replicas are, to a certain degree, subject to situated interpretation
(Berger and Luckmann, 1967: 7) and the ‘interpretative authority’ of the research (Markham, 2012: 15). As such,
we argue that AI-replicas reflect the situated perceptions of the researcher whilst supporting the maintenance of
6
the data richness of the original data. The degree to which AI-replicas may manipulate or otherwise impact the
qualitative richness of images is unclear. We wish to address this question in Chapter 4 after we illustrate the
application potential of AI-replicas in Chapter 3.
AI-replicas offer numerous benefits that could contribute to overcoming some of the concerns
highlighted above. Firstly, AI-replicas preserve emotional and expressive content, which is crucial in visual data
where facial features are key biometric identifiers. (Smith et al., 2018; Smith and Miller, 2022). Standard
anonymization techniques often involve redactions such as adding black bars or blurring or pixelating faces. Such
alterations significantly alter the original material, impacting the reader’s understanding of the researcher’s
interpretation. AI-replicas can generate entirely new pictures whilst maintaining accurate facial expressions.
Secondly, AI-replicas provide additional customization and control over the reported data. These tools
enable researchers to modify aspects that are potentially irrelevant to specific research objectives, questions,
methodologies, and reporting requirements (e.g. race or location). Thirdly, in contrast to traditionally anonymized
images, AI-replicas exhibit enhanced legibility. The excessive application of filters and other alterations can
hinder the interpretability of visual materials. The AI-replicas can help maintain clear visual legibility by keeping
the images free of obscuring artifacts. These aspects can be implemented without compromising the privacy of
the data subjects whilst presenting a relatively low technological entry barrier.
2.2 Advancing Visual Research: An Introduction to Generative AI and Its Role in Crafting AI-Replicas
Users presently have a vast selection of publicly available generative AI tools for image generation, including
Adobe Firefly, Lexica, Dreamlike, Midjourney, DALL•E (Betker et al., 2023; Mishkin et al., 2022) and Stable
Diffusion (Stability-AI, 2023c). Whilst all these applications are chiefly designed for artistic purposes, a pivotal
difference between these services is that most of them are commercialized cloud-based process pipelines that offer
a streamlined experience for paying customers. Stable Diffusion is an open-source project that can be executed
on local computers. This approach provides greater degrees of control at the cost of user convenience. Although
suitable hardware is required to effectively run the software (i.e. a contemporary GPU with 10 GB VRAM), Stable
Diffusion avoids the drawbacks of purely web-based services. Firstly, considering local computations, Stable
Diffusion is constantly available and does not incur additional costs per image generation. Secondly, Stable
Diffusion allows the use of fine-tuned models (checkpoints) trained on custom datasets1. Thirdly, Stable Diffusion
ensures a higher degree of privacy because it is not a cloud-based service. A local installation of Stable Diffusion
avoids all possible concerns of losing control over images because all the data generated and used for generation
stay on a local computer. Finally, the open-source distribution and community support enable the use of additional
neural networks for greater control over the generated images.
Stable Diffusion is a diffusion model (Rombach et al., 2022; Weng, 2021) that utilizes ‘noise reversal’
or ‘denoising’ (see also Yang et al., 2024: 6). The modelling starts with a canvas of a random pattern of pixels
(visual noise), which is then refined across several sampling steps in which stable diffusion improves on the image
via iterative refinement (sampling steps). Throughout this process, new values are imposed on the pixels until
they resemble visual patterns associated with a text prompt. This process allows the model to create detailed and
accurate visual representations of the text (Rombach et al., 2022), which has enormous potential for applications
in text-based qualitative research.
Stable Diffusion can be compared to machine learning and neural networks and less so for generative
pretrained transformers such as ChatGPT.
Figure 1: Sampling and Denoising
7
Figure 1 illustrates this denoising process based on the following prompts with 1, 4, 6 and 35 sampling
steps:
Two men are squatting on the dirt floor, each holding a rooster. The men are focused intently on their
birds, preparing for a traditional cockfight. They are dressed casually denoting a warm, tropical climate.
Around them, a group of onlookers forms a semicircle, all standing and watching with keen interest,
photography.
Beyond text-to-image generation, Stable Diffusion enables image-to-image generation for guided
synthesis (Mao et al., 2023; Meng et al., 2022; Wang et al., 2021). This procedure allows for the control of the
generation of replicas based on input images, optionally under consideration of text prompts. The input images
will predominantly affect the composition and overall colour of the output images. This is achieved by generating
new images not only with random noise but based on input images with layers of added noise. Both methods and
their potential application scenarios will be further elaborated upon in Chapter 3.
The creation of replicas with Stable Diffusion is therefore understood as image generation based on the
composition of input images, notwithstanding that text-to-image generation also has application potential, as
illustrated below. The output quality and possible resolutions depend chiefly on the models used (Bie et al., 2023;
Podell et al., 2023: 16; Yang et al., 2024), which are trained on extensive image datasets (e.g. Beaumont, 20022;
Common Crawl, 2023; CompVis, 2022). Models are trained for image generation by adding noise to the training
data and effectively learning the difference between the original images and several rounds of additional noise.
During image generation, this process is reversed, and the images are generated by sequentially removing noise.
A plethora of free and paid models exist on the internet, with the option to train and fine-tune custom
models. The majority of these models are fine-tuned versions of the popular Stable Diffusion v1-5 checkpoint
(Rombach et al., 2022). Fine-tuning involves training existing models on specific types of images, such as human
faces, landscapes, animals or art. For the anonymization of human data, models based on the recently published,
new Stable Diffusion-XL 1.0 base model are recommended (Podell et al., 2023; Stability-AI, 2023b). It must be
noted that models used for generative AI are biased (Heikkilä, 2023), as the training data are vast but ultimately
limited. These biases must be further reduced in the long term by pruning and refining the base models. An
experiment by Nicoletti and Bass (2023, based on Stable Diffusion v1-5) revealed that Stable Diffusion displays
racial bias when prompted to generate images of high- and low-income jobs. These model-conditioned biases can
be avoided through more detailed prompts (e.g. the specification of skin colour). Nonetheless, such biases also
lead to inherent limitations when generating images of vulnerable individuals and thus, in training datasets, less
represented groups, for example, groups with physical disfigurations. In the following Chapter, we illustrate the
potential of Stable Diffusion to generate images in the form of semantically similar AI-replicas. This account, not
a comprehensive guide, aims to illustrate and inspire the application of AI replicas in image anonymization. These
AI-replicas should not replace the original material for analysis but serve as substitutes in scenarios such as (1)
circulating images in experiments, (2) supplementing field notes when visual documentation is unfeasible and (3)
reporting image data whilst safeguarding privacy and maintaining legibility.
3. Practical Application of Stable Diffusion for Generating AI-Replicas
The approach presented in this work is realised via the local installation of Stable Diffusion (Rombach et al., 2022;
Stability-AI, 2023c) and the Stable Diffusion web UI (AUTOMATIC1111, 2023), both of which are publicly
available via their GitHub repositories.2 Instead of passing parameters to Stable Diffusion via command-line
prompts, the latter allows the operation of Stable Diffusion via an intuitive browser interface. In the following
Chapter, we introduce four application examples of Stable Diffusion for the creation of AI-replicas to illustrate
its application potential.
3.1 Text-to-Image Generation
8
Figure 2 shows an example of a hypothetical case in which an image should be replaced with an AI-replica. Before
and throughout the generation process, which aspect of the images is relevant must be balanced and must be
reported. The focus may be on which aspects must be anonymized and what degree of abstraction and alteration
is acceptable; this remains, of course, relative to the research questions and the employed research methodology
and should be decided upon after the analysis of the original material.
Figure 2: First Example Case for Image Anonymization
Similar to chatbots such as ChatGPT, Bard or Poe, AI image generation tools operate in their most basic
application on text prompts. The text prompt communicates Stable Diffusion <what> to generate. Natural
language descriptions complemented with specific keywords usually work the best and ideally contain
descriptions of the scene, the composition, the perspective, the focal length or the medium. A possible text prompt
for the example picture could look as follows:
a Caucasian man with a blue jacket, black shirt and black jeans standing next to a tree in a park holding
a cell phone in his hand and looking at the camera with a serious look on his face, street and car in the
background, canon 50 mm, light bokeh.
When generating pictures via text prompts alone, the most pivotal variable next to the prompt itself and
the previously illustrated sampling steps (see Figure 1) is the classifier-free guidance (CFG) scale. The CFG scale
determines how closely the generated images follow the text prompt and thus regulates how much of the text
prompt finds its way into the final pictures and relies strongly on the model checkpoint used. A low CFG scale
will effectively generate a random image (underfitting), whilst an extremely high value will inevitably create
visual artifacts and potentially abstract outcomes (overfitting).
Figure 3: CFG Value Comparison
Using the previous prompt, Figure 3 illustrates how a higher CFG value will inevitably underscore all
the elements of the prompt, with the CFG 7 variations resembling the original compositions the best. All the
variations default to a tree with yellow leaves when not closely specified. These details can be adjusted using the
text prompt, as illustrated in the adjusted version CFG 7* (i.e. ‘next to a green tree’). These adjusted examples
show that the colour of the leaves can easily be adjusted but also reveal that the similarity of CFG 7 with the
composition of the original pictures was random, as CFG 7* deviates further in terms of composition3. Depending
9
on the case, one can argue that either of these variations already suffices as abstract representations of the original
image but with a clear potential to improve the alignment with the original image.
These basic text-to-image generations demonstrate some of the inherent limitations of this technology.
For example, whilst CFG 7 resembles the original picture superficially the most, the man was rendered with two
mobile phones instead of one by chance. This phenomenon is related to the stochastic nature of Stable Diffusion
and can be compensated for by continuous generations with different seed values4, as illustrated by CFG 7*.
However, the latter renders the man with a green jacket instead of a blue jacket and concentrates on the phone
instead of the camera. Here, we observe an association error that can occur in complex prompts, in which different
colours are assigned to different objects. These phenomena are related to the employed model and the complexity
of the data on which it was trained. Similar to when multiple objects are used, these errors can be circumvented
by generating the same images with different seed values. Whilst these examples illustrate the basic potential of
generative AI, they also highlight its shortcomings concerning the predictability of the outcome that must be
addressed before the generated images can be used as functional substitutes.
Text-to-image generations benefit from highly detailed prompts and, if possible, fine-tuned models that
specialize in the desired outcome. This type of AI-replica is suitable when the pictures are supposed to be more
abstract representations of findings, such as typologies or aggregated findings, and for the support of field notes
or narratives that had no visual documentation in the first place. Similarly, visual support can be provided to
textual descriptions of visual data recalled by participants (cf. Ward, 2017: 1653). Text-to-image generations are
arguably less effective if one wishes to report on more specific aspects of an image that require stronger adherence
to the original composition. For this sake, greater accuracy can be achieved via image-to-image generation.
3.2 Image-to-Image Generation
Whilst text-to-image generations begin with complete random noise, image-to-image generations start
with an input image as the basis. In contrast to text-to-image generation, new images are not calculated purely by
approximating the text prompt through the denoising of complete random noise. Instead, noise is applied to
various degrees to the input image or a particular section of it (Meng et al., 2022). One can simply imagine
applying clay to a sculpture to mould a new face along the lines of the underlying original. The intensity to which
noise is applied to the input image is regulated by the denoising strength. The denoising strength determines to
what degree the original image will be reimagined toward the desired outcome as specified in the text prompt (the
CFG scale still applies). A reduced denoising strength adds less noise, keeping the output closer to the input,
whilst high values lead to more abstract output.
Figure 4: Denoising Strength Comparison
Figure 4 shows the image-to-image generation results with the original prompt by using three different
denoising strengths (CFG value 7). The range of 0.35 to 0.51 strongly approximates the original pictures whilst
displaying varying levels of alteration. 0.35 appears as a light alteration, 0.51 gives the impression of an
enactment, whilst 0.91 appears as a reinterpretation of the original. Although quite similar, one can observe a
subtle difference in the 0.51 variation, for example, an additional car in the background, the absence of a light
post or a shift in the mobile phone’s position. Based on our experience, the AI-replicas, in the presented 0.30–
0.60 range, can be used as a substitute for the original picture for reporting purposes to retain pivotal compositional
and expressive information whilst altering the environment and actor. More research is needed to systematically
10
identify an optimal range. However, the most suitable value is strongly related to the picture in question and the
desired outcome. If necessary, a prompt can be used to further alter the picture.
Figure 5: Prompt Alteration Comparison
Figure 5 illustrates how ‘Caucasian male’ in the original prompt was replaced with ‘Asian male’ and
‘Caucasian female’. Whilst this example could be considered a critical alteration in the pictures, it serves the sole
purpose of illustrating the flexibility and fidelity with which the AI-replicas can be altered to suit the situated
needs of the researcher. This example shows that the text prompt is a secondary factor when working with image-
to-image generations, which align closely with the original composition. However, a supporting text prompt
allows us to alter critical aspects of the image that can further support the anonymization of the input data.
Image-to-image generation enables stronger alignment of the AI-replicas with the composition of the
input image. These algorithms are suitable for AI-replicas, which should communicate the nuances of a particular
composition that could otherwise not be captured by text-to-image generation. The denoising procedure that
generates the AI-replicas ‘on top’ of the original image can essentially replace everything within the picture with
a semantic double, leading to full anonymization. This procedure is our recommended default approach when
generating AI-replicas. Furthermore, high denoising values allow for more abstracted AI-replicas whilst still
having high control over the output image based on the original data material. Higher degrees of abstraction can
be useful when the risk of exposing an informant is not rooted in biometric information alone but rather in iconic
compositions that would give away the identity of the informant due to the alleged iconic singularity of a given
picture. Whilst abstraction can be desirable in certain scenarios, as described above, it leaves the concern that
images may lose the details and qualitative nuance that are deemed essential for the analysis. For even greater
control over image generation, Stable Diffusion allows the use of additional conditions in the form of ControlNet
models.
3.3 ControlNet
ControlNet (Zhang et al., 2023) enables conditions for image generation via a vast array of models and can be
used by itself or in combination with image-to-image generation to exert more precise control over image
generation. Given the considerable diversity among ControlNet models, a comprehensive introduction to all of
them at this juncture is impractical. For image anonymization, the two most important models are OpenPose for
body poses and facial expressions and Depth for spaces. ControlNet preprocessors process input images and
extract information that ControlNet models use to guide the generation of new images. These parameters can be
used for text-to-image and image-to-image generation. The benefit of using ControlNet is that it allows further
abstraction when used in text-to-image generation and greater alignment with the input image in image-to-image
generation.
11
Figure 6: Second Example Case for Image Anonymization OpenPose Preprocessor Skeleton Frame
Figure 6 exemplifies the application of the OpenPose model and preprocessor. The input image is on the
left-hand side, whilst the abstraction on the right represents the body key points identified by the OpenPose model.
The latter will be used to guide the image generation of Stable Diffusion. A prompt to describe the input image
could be:
Two men engaged in a friendly discussion whilst sitting at a round table in a modern library. The person
on the left is a young Caucasian man with short brown hair wearing a black jacket and black pants. He
is in a pensive pose with his hands clasped together, looking attentively at his interlocutor. The person
on the right is a young Hispanic man with short black hair, a warm smile and a light stubble wearing a
navy-blue and white sports jacket and black pants. They are surrounded by bookshelves filled with
various books. On the table, there is a piece of paper with notes, a textbook and a blue aluminium can.
The overall atmosphere is casual and intellectual, with a 35 mm eye-level angle.
Figure 7: Image Generation with the OpenPose Model Comparison
Figure 7 illustrates the possible outcomes of image generation. Variation (1) is a replica purely based on
the text prompt, whilst variation (2) uses the OpenPose model to align the actors in the replica with the original
image. Variation (3) uses image-to-image generation with a denoising strength of 0.51 plus the OpenPose model.
Variation (3) closely follows the original composition whilst emulating the facial expressions of the original actors.
This procedure helps to further align facial expressions and the environment with the original. Some details, such
as the beverage, are ignored when only the text prompt and the preprocessor are used, but they are included when
image-to-image generation is utilised.
12
Figure 8: Illustration of Pose Transfer from Original Images to AI-Replicas
Image-to-image generations continue to provide results for the creation of replicas, especially when
layered with models such as OpenPose. However, this approach does not undermine the application potential of
ControlNet. Figure 8 illustrates how OpenPose can be used to transfer compositional elements to new settings,
allowing us to retain the choreography between the actors whilst completely replacing the latter or the environment.
Figure 8 was generated based on the same prompt as related to Figure 8 but by changing the gender to female, the
location to a restaurant and the removal of bookshelves. As stressed above, the exact application scenario depends
on what aspects of an image should be reported.
Figure 9: Illustration of Depth ControlNet Using a Preprocessor Depth Mask
The functionality of ControlNet is limited not only to human actors and their poses. For instance,
fictionalization may also be required to anonymize locations whilst keeping information concerning the
perspective of the depicted scene. Figure 9 illustrates how the depth model uses depth maps to control the
perspective and depth of the generated images. In addition to the introduced models, a variety of other ControlNet
models can be used to control the output of image generation with Stable Diffusion (see Zhang, 2023). All these
models are situational in their application and can be combined.
3.4 Inpainting
Another key function for the creation of AI-replicas is inpainting. ‘Inpainting’ can be conceived as ‘filling masked
regions of an image with new content either because parts of the image are corrupted or to replace existing but
undesired content within the image’ (Rombach et al., 2022: 8). This method is essentially a localized image-to-
image generation that allows the AI-replication of specific image sections. When applied effectively, this
technology allows pivotal areas of an image to be rendered anew and merged seamlessly with the existing image.
This technique can be applied to both AI-generated images and original pictures. The application potential of this
function is threefold. Firstly, inpainting allows the occasional generation problems of the methods introduced
above, such as two mobile phones or the colour of the jacket, to be fixed (Figure 3). Secondly, inpainting enables
the restoration of previously anonymized elements (e.g. faces) and their replacement with AI-replicas. The latter
is of critical importance when reporting visual data that have been stored in anonymized form only, whether for
ethical or legal reasons (cf. Regulation (EU) 2016/679, 2016). Thirdly, inpainting allows the AI-replication of a
particular information (e.g. faces) and thus serves as the most immediate alternative to traditional anonymization
methods (i.e. black bars, blur or pixelation) that retain the visual legibility of the original image. This application
13
of AI-replicas can be especially useful if one wishes to prepare material to be rotated beyond the confidential
relationship between the researcher and the data subject, for example, as stimulus material in group discussions.
Figure 10: Inpainting for Restoration of Blurred Faces
Figure 10 illustrates the potential of the inpainting function to fill in previously anonymized pictures. On
the basis of our previous example, we created a blurred version of Figure 2, which we subsequently processed
with inpainting to create an AI-replica that emulates a face similar to the original image. This particular AI-replica
was generated with a denoising strength of 0.51 and a CFG value of 7 (see Figure 21 in Rombach et al., 2022: 32
for further examples). This process is subject to the same limitations as full-image generation (i.e. parameters and
the text prompt) and may require several iterative steps to achieve the desired output. Whilst showing the same
anonymizing potential as illustrated in Figure 5, this step also underlines the potential of AI-replicas to enhance
anonymized material in a way that retains the legibility of non-anonymized material.
4. Discussion
The previous Chapter illustrated the application potential of AI-replicas as an alternative to traditional
anonymization methods when reporting visual data. However, this possibility also invites a set of questions
concerning the application potential, pitfalls and ethical concerns that we intend to address in this Chapter.
4.1 AI-Replicas and Data Richness
A major concern in image anonymization and data fictionalization is that removing information from an image
must be carefully balanced against the ethical and research-practical necessities of a given usage scenario. The
question concerning the data richness is pivotal for the previously addressed concern of validity and reliability of
the presented interpretation (see Chapter 1). When applying traditional anonymization methods, it is assumed that
the application of black bars or pixelation of the face does not diminish the richness of the data and thus still
allows the reader to comprehend the provided interpretation. However, this is relative to a given research objective.
If the particular gaze and expressions on a given face are pivotal factors for a given interpretation, then
anonymizing them would render the interpretation impossible.
The examples presented in Chapter 3 show that guided image synthesis can help create AI replicas that
can retain expressive nuances whilst completely obscuring the identity of the data subjects whilst retaining the
compositional complexity of the originals. However, providing a definitive answer to the question of whether AI-
replicas distort or manipulate the data to a critical degree is impossible because this approach is relevant to a given
research objective. For example, one could argue that the additional car and missing lamppost in Figure 5 (0.51)
critically altered the essence of the original picture. Similarly, one could argue that these changes are too subtle
to have any meaningful impact on any potential research questions and instead contribute to the fictionalization
effect of AI-replicas. If an extremely holistic application of the proposed method seems undesirable, then
retreating to more local applications is always an option, as illustrated in Figure 10. On this level, AI-replicas
generated with Stable Diffusion would fulfil the same function as contemporary automated full-body
anonymization methods (Hukkelås and Lindseth, 2022) but allow the user fundamentally greater control over the
outcome. This approach utilises AI-replicas and is particularly desirable if we wish to retain selective information,
such as facial expressions.
14
The results of qualitative data analysis, particularly interpretative approaches that go beyond simple
codification of written text, as is commonly used in qualitative content analysis (cf. Graneheim and Lundman,
2004), are the outcome of a deep engagement of researchers with the data material. The richness of the data, or
rather what is considered relevant for the research question at hand, plays a critical role in the identification and
explication of meaning as represented by the data material. The situatedness of the researcher also plays a critical
role in this process (Berger, 2015; Berger and Luckmann, 1967: 7). Therefore, AI-replicas will echo the selectivity
and validation of the research-relevant information to a certain degree, similar to written protocol sentences.
However, in contrast to protocol sentences, AI-replicas can align closely with the original material by using image-
to-image generation with lower denoising strength. In such a sense, AI-replicas are a form of display ‘to provide
evidence for claims in a format readers can easily access’ (Watt, 2015). Creating AI-replicas must not be
‘something automated delegated to the machine’ but rather a delicate process that is carefully administered by the
researcher and based on their original analysis. Therefore, similar to qualitative analysis itself (Elo and Kyngäs,
2008: 113), it is recommended that fellow researchers validate AI-replicas.
In addition to the anonymization of original data, we propose the use of AI-replicas for reporting
aggregated data that represent and accentuate particular aspects of the analysed image data but without referring
to any image in particular (e.g. for the development of typologies; visualization of prototypes) or supporting data
that are not visual in nature (e.g. field notes). For instance, social psychologists and anthropologists have long
used the idea of prototypes in reference to the collection of quintessential elements of a concept (see MacLaury,
1991; cf. ‘ideal type’ in Schütz, 1972; Weber, 1949). A concept has features with various levels of centrality that
define a prototype, and people use them to judge what an object is and how it can be categorised. Here, more
abstracted AI-replicas can find an application to convey the overall features of a particular prototype but without
the need to adhere to the iconic singularity of any analysed particular image.
When used complementarily to field notes or mind protocols, AI-replicas can fulfil a supplementary
function. Similar to interpretations or interviews regarding this matter, field notes are reductionist in nature and
bound the situated perception of the researcher (Emerson et al., 1995; Mulhall, 2003). Here, the AI-replicas
function complementarily to field notes, similar to visual sketches (Phillippi and Lauderdale, 2018), and can
provide a pivotal visual dimension where visual documentation would otherwise (for which reason whatsoever)
not be possible. This process requires the same critical reflection concerning the results of the image generations
as would be for the creation of field notes (Watt, 2015) and qualitative research in general (Berger, 2015). This
call for reflexivity applies not only to the generation of completely synthetic AI-replicas but also to the types
discussed above.
4.2 Handling Bias
A common concern when addressing any technology that falls under the broader label of AI is how biased the
results are. As previously stressed, various sources have pointed out how susceptible technologies such as Stable
Diffusion are to biased results (Heikkilä, 2023; Jenka, 2023; Nicoletti and Bass, 2023). Stable Diffusion uses
contrastive language–image pretraining (CLIP) (OpenAI, 2020; Radford et al., 2021) to establish the connection
between real language input (text prompts) and the image material it was trained on, effectively guiding the
generation process. For example, CLIP connects the word ‘house’ with generations it learnt to represent ‘houses’.
The particular models of Stable Diffusion (e.g. SD 1.5 or SD-XL 1.0) guide the creative capacity of image
generation. Metaphorically speaking, CLIP operates as a teacher who provides Stable Diffusion with the
association of words and semantic concepts, whilst Stable Diffusion’s creative capacity translates these text-
concept combinations into a visual representation. Potentially biased generations are therefore found in the
datasets on which a CLIP or Stable Diffusion model was trained. If a Stable Diffusion model is not trained
sufficiently on a given object (e.g. physical features associated with Down syndrome), then the model will not
know how to generate it (underfitting) or how to approximate it, in the best case, via association with other objects
in the dataset. Similarly, as illustrated in Figure 3, Stable Diffusion tends to represent certain concepts in a given
way (e.g. the yellow tree). This situation can be regarded as the result of the dataset and an association in and
through the text prompt. For example, ‘a tree during the autumn season’ will most likely also lead to a yellow tree,
despite not being prompted as such. Similarly, the prompt used for Figure 1 did not specify the skin colour of the
15
two men or the bystanders. However, the model relies on data associating cockfight, clothing and climate with a
skin colour that suits the cultural background of cockfights (see Geertz, 1972). The semantic association of
concepts can therefore also be beneficial for the generation of AI-replicas. The concern described above is most
evident in text-to-image generations that can suffer from insufficient prompts, as illustrated by the tree. Using
image-to-image generation is therefore a potent strategy to counteract any bias in the data.
This circumstance highlights the importance of understanding proper prompt engineering as a pivotal
skill that can actively help counteract biases in the models used. Generating AI-replicas will require the researcher
to cultivate a high conceptual clarity of what should be generated. For example, when generating ‘a man next to
a tree’, we must actively decide what age or skin colour Stable Diffusion should render, but this must remain
relative to the situated need of the AI-replica. Moreover, this level of detail must always be anticipated by the
researchers during generation. Generating AI-replicas is therefore an active and reflective process and not
necessarily a one-click automatization. As such, the convenience of simple pixelations might be lost, but the
approach could yield potentially more authentic representations.
The aforementioned limitation of the models used will prevent Stable Diffusion from generating
everything conceivable, as some classes of objects will inevitably be insufficiently represented. These limitations
are usually addressed using fine-tuned models that specialize in particular outputs. The extent of this limitation,
or rather the actual creative potential, requires further investigation. However, due to the rapid improvement of
the models, this is difficult to realize. Furthermore, a more practical application of AI replicas for research
purposes is needed to identify field- and discipline-specific needs to train custom checkpoints.
5. Conclusion
Offering anonymity to research participants is an incontestable principle in image-based research (Nutbrown
2010). Anonymity is a recurrent requirement of ethical review boards and outlets (Allen 2015). Nevertheless,
anonymization poses several potential challenges to the research process: silencing the voice of research
participants, shaping the research interests of researchers, reducing agency, increasing paternalism and limiting
the possibility of reporting evidence that affects the validity of the findings. Moreover, current image modification
and fictionalization techniques (Jordan 2014; Nutbrown 2010; Wiles, Coffey, Robinson, et al. 2012) do not fully
account for or overcome these limitations.
This article introduces AI-replicas as a novel approach for ethically reporting visual data, adeptly
balancing participant anonymity and data integrity. We have demonstrated that the AI-replica has the potential to
effectively preserve emotional content and image authenticity, overcoming the limitations inherent in traditional
anonymisation. This article proposes the use of AI-generated replicas to report visual data as complementary
material to otherwise purely textual protocol sentences (Bohnsack, 2008: 3; Popper, 1959: 59ff.) or as a novel
reverse photovoice procedure where the voice of participants can be utilized as an input to generate visual
representations.
Certain concerns about the applicability of AI in qualitative research are legitimate. Firstly, qualitative
researchers could be concerned about the possible distortions of images that could affect studies’ conclusions.
Nevertheless, our proposal only entails the use of AI-replicas for reporting purposes. Secondly, qualitative
researchers could be concerned about well-known biases of AI (Heikkilä, 2023; Nicoletti and Bass, 2023).
Nevertheless, traditional anonymization methods can also be subject to bias. Moreover, text prompts provide
flexibility to adjust the output image to counteract biased generations. In this sense, the reflexibility of researchers
and critical evaluation of outputs are key elements for identifying biases and correcting them. Furthermore, if
possible, research participants can be involved in the generation of AI-replicas to give them creative authority
over the end product. Thirdly, qualitative research could involve data documentation and transparency. This
situation is another limitation of traditional fictionalization methods. It is therefore pivotal to adhere to good
practices of generative AI usage and document its use as diligently as possible (i.e. prompts, parameters and
ControlNet usage).
An inherent benefit of traditional anonymization methods is that the degree of their impact is often self-
evident. When pixelating a face, researchers can easily judge whether this technique disables the audience from
comprehending and potentially validating a presented interpretation. This aspect becomes more nuanced with the
16
proposed AI-replicas, as the AI-replicas can ideally not be distinguished from the original ones in terms of
perceived authenticity. The practical application of AI-replicas must be accompanied by a declaration of how and
to what extent this technology has been used. The type of image generation employed (i.e. text-to-image, image-
to-image or inpainting) and the type of ControlNet applied should therefore be described in a transparent manner.
Reporting the guiding prompt used and the specific parameters for the generation (i.e. denoising strength and CFG
scale) is also recommended. This approach will help the reader understand how to assess the degree and form to
which an AI-replica is an alteration of the original. This practice can be complemented by other developing
standards of AI usage in academic work, for example, AI usage cards (Wahle et al., 2023). A similar practice of
standardized reporting of generation data is already in place amongst online communities that use Stable Diffusion
for artistic purposes and must be only slightly streamlined for academic purposes and variations such as image-
to-image and inpainting.
The use of AI-replicas necessitates cautious consideration of AI biases and broader applicability in
various research scenarios. Future work should aim to further evaluate the application potential and quality of this
method concerning the limitations of the employed models to improve the utilization of AI-replicas in various
application scenarios. Furthermore, future studies can investigate the application potential of AI-replicas for the
fictionalization of video material, which can be made publicly available complementarily to the respective main
publication.5 As we recognize the current early stages of this methodology, a fundamental consideration is the
ongoing development of the underlying technology, datasets and models, which implies continuous refinements
and advancements. This article marks an initial step in leveraging generative AI’s potential for anonymization and
fictionalization in visual research. We expect to contribute to well-position qualitative research in the evolving AI
landscape.
Acknowledgements
All presented AI-replicas (Figures 1, 3, 4, 5, 7, 8, 9 and 10) have been generated using the SD-XL 1.0-
base model (CreativeML Open RAIL++-M License, Stability-AI, 2023b).
Notes
1. DALL•E and Midjourney allow model configuration but restrict the choice to proprietary, in-house trained
models (Midjourney, 2024; Niles and OpenAI, 2024).
2. The Stable Diffusion community provides a vast set of comprehensive guides for the installation and use of
Stable Diffusion and its modules:
• Stable Diffusion installation guide: https://stable-diffusion-art.com/install-windows (Andrew, 2023b)
• ControlNet installation and application guide: https://stable-diffusion-art.com/controlnet (Andrew,
2023a)
3. We intentionally ignore the possibility of using a consistent seed variable throughout multiple generations.
4. Seed values are used to control the randomness of image generation. Using the same seed value with stable
parameters will allow you to replicate the same image generation. The seed values are randomized by default.
5. Suitable image-to-video models are already in active development (Stability-AI, 2023a, 2023d).
References
Allen L (2015) Losing Face? Photo-Anonymisation and Visual Research Integrity. Visual Studies 30(3): 295–
308.
Andrew (2023a) ControlNet v1.1: A complete guide. Available at: https://stable-diffusion-art.com/controlnet/
(Accessed: 11 January 2024).
Andrew (2023b) How to install Stable Diffusion on Windows (AUTOMATIC1111). Available at: https://stable-
diffusion-art.com/install-windows/ (Accessed: 11 January 2024).
AUTOMATIC1111 (2023) Stable Diffusion web UI. Available at: https://github.com/AUTOMATIC1111/stable-
diffusion-webui (Accessed: 16 December 2023).
Banks M (2001) Visual Methods in Social Research. London: Sage.
17
Beaumont R (2022) Laion-5B: A New Era of Open Large-Scale Multi-Modal Datasets. Available at:
https://laion.ai/blog/laion-5b/ (Accessed: 14 January 2024).
Berger PL and Luckmann T (1967) The Social Construction of Reality: A Treatise in the Sociology of Knowledge.
New York, USA: Anchor Books.
Berger R (2015) Now I see it, now I don’t: researcher’s position and reflexivity in qualitative research. Qualitative
Research 15(2): 219–234.
Betker J, Brooks T, Wang J et al. (2023) DALLE3: Improving Image Generation with Better Captions. arXiv.
Epub ahead of print 2023.
Bie F, Yang Y, Zhou Z et al. (2023) RenAIssance: A Survey into AI Text-to-Image Generation in the Era of Large
Model. pp. 1–25.
Bohnsack R (2008) The Interpretation of Pictures and the Documentary Method. Forum: Qualitative Social
Research 9(3): 1–17.
Burgos-Thorsen S and Munk A (2023). Opening alternative data imaginaries in urban studies: Unfolding COVID
place attachments through Instagram photos and computational visual methods. Cities 141.
Common Crawl (2023) Common Crawl. Available at: https://commoncrawl.org/ (Accessed: 14 January 2024).
CompVis (2022) Stable Diffusion v1 Model Card. Available at: https://github.com/CompVis/stable-
diffusion/blob/main/Stable_Diffusion_v1_Model_Card.md#training (Accessed: 14 January 2024).
Douglas K (2022) Anonymizing People in Images Using Generative Adversarial Networks. MSc Thesis,
University of Amsterdam.
Elo S and Kyngäs H (2008) The qualitative content analysis process. Journal of Advanced Nursing 62(1): 107–
115.
Emerson R, Fretz R and Shaw L (1995) Writing Ethnographic Fieldnotes. Chicago, USA: The University of
Chicago Press.
Gan Y, Greiffenhagen C, and Licoppe, C (2020) Orchestrated openings in video calls: Getting young left-behind
children to greet their migrant parents. Journal of Pragmatics 170: 364-380.
Geertz C (1973) Deep Play: Notes on the Balinese Cockfight. Daedalus 101(1): 1–37.
Graneheim UH and Lundman B (2004) Qualitative content analysis in nursing research: Concepts, procedures
and measures to achieve trustworthiness. Nurse Education Today 24(2): 105–112.
Heikkilä M (2023) These new tools let you see for yourself how biased AI image models are. Available at:
https://www.technologyreview.com/2023/03/22/1070167/these-news-tool-let-you-see-for-yourself-how-
biased-ai-image-models-are/ (Accessed: 13 January 2024).
Hukkelås H and Lindseth F (2022) DeepPrivacy2: Towards Realistic Full-Body Anonymization. Available at:
https://arxiv.org/pdf/2211.09454v1.pdf (Accessed: 14 January 2024).
Imdahl M (1996) Giotto, Arenafresken: Ikonographie – Ikonologie – Ikonik. Munich, Germany: Wilhelm Fink
Verlag.
Jenka (2023) AI and the American Smile: How AI misrepresents culture through a facial expression. Available
at: https://medium.com/@socialcreature/ai-and-the-american-smile-76d23a0fbfaf (Accessed: 7 July
2023).
Kozinets RV (2020) Netnography: The Essential Guide to Qualitative Social Media Research (3rd ed.). SAGE
Publications Ltd.
Laurier E (2014) The graphic transcript: Poaching comic book grammar for inscribing the visual, spatial and
temporal aspects of action. Geography Compass 8(4): 235-248.
Luhmann N (1995) Social System. Stanford: Stanford University Press.
MacEntee K., Kendrick C and Flicker S (2022) Quilted cellphilm method: A participatory visual health research
method for working with marginalised and stigmatised communities. Global Public Health 17(7): 1420-
1432.
MacLaury RE (1991) Prototypes Revisited. Annual Review of Anthropology 20(1): 55–74.
Mao J, Wang X and Aizawa K (2023) Guided Image Synthesis via Initial Image Editing in Diffusion Model. MM
2023 - Proceedings of the 31st ACM International Conference on Multimedia: 5321–5329.
Markham A (2012) Fabrication as ethical practice: Qualitative inquiry in ambiguous Internet contexts.
Information Communication and Society 15(3): 334–353.
Maximov M, Elezi I and Leal-Taixe L (2020) CIAGAN: Conditional Identity Anonymization Generative
Adversarial Networks. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition
(CVPR), June 2020, pp. 5446–5455. IEEE. Available at: https://ieeexplore.ieee.org/document/9157102/.
18
Meng C, He Y, Song Y et al. (2022) Sdedit: Guided Image Synthesis and Editing With Stochastic Differential
Equations. ICLR 2022 - 10th International Conference on Learning Representations. Epub ahead of print
2022.
Midjourney I (2024) Version. Available at: docs.midjourney.com (Accessed: 13 January 2024).
Mishkin P, Ahmad L, Brundage M et al. (2022) DALL·E 2 Preview - Risks and Limitations.
Mulhall A (2003) In the field: Notes on observation in qualitative research. Journal of Advanced Nursing 41(3):
306–313.
Nicoletti L and Bass D (2023) HUMANS ARE BIASED. GENERATIVE AI IS EVEN WORSE. Available at:
https://www.bloomberg.com/graphics/2023-generative-ai-bias/ (Accessed: 30 January 2024).
Niles R and OpenAI (2024) DALL-E 3 API. Available at: https://help.openai.com/en/articles/8555480-dall-e-3-
api (Accessed: 13 January 2024).
Nutbrown C (2011) Naked by the pool? Blurring the image? Ethical issues in the portrayal of young children in
arts-based educational research. Qualitative Inquiry 17(1): 3–14.
OpenAI (2020) CLIP. Available at: https://github.com/OpenAI/CLIP.
Parmar N, Vaswani A, Uszkoreit J et al. (2018) Image transformer. 35th International Conference on Machine
Learning, ICML 2018 9: 6453–6462.
Papademas D and IVSA (2009) IVSA Code of Research Ethics and Guidelines. Visual Studies 24(3): 250–57.
Phillippi J and Lauderdale J (2018) A Guide to Field Notes for Qualitative Research: Context and Conversation.
Qualitative Health Research 28(3): 381–388.
Podell D, English Z, Lacey K et al. (2023) SDXL: Improving Latent Diffusion Models for High-Resolution Image
Synthesis. Epub ahead of print 2023.
Polat Z (2022) Facilitating a ‘virtual space’ for social change during the COVID-19 pandemic: working with high-
risk population using an arts-informed method. Visual Studies 37(1-2): 22-32.
Popper K (1959) The Logic of Scientific Discovery. London: Hutchinson & Co.
Prosser J (2000) The Moral Maze of Image Ethics. In: Simons H and Usher R (Eds.), Situated ethics in educational
research (pp. 116–32). London: Routledge.
Radford A, Kim JW, Hallacy C et al. (2021) Learning Transferable Visual Models From Natural Language
Supervision. Proceedings of Machine Learning Research 139: 8748–8763.
Regulation (EU) 2016/679 (2016) REGULATION (EU) 2016/679 OF THE EUROPEAN PARLIAMENT AND
OF THE COUNCIL. Available at: https://eur-lex.europa.eu/legal-
content/EN/TXT/HTML/?uri=CELEX:32016R0679&qid=1706168522662.
Rombach R, Blattmann A, Lorenz D et al. (2022) High-Resolution Image Synthesis with Latent Diffusion Models.
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
2022-June: 10674–10685.
Schütz A (1972) The Phenomenology of the Social World. London: Heinemann Educational.
Smith M and Miller S (2022) The ethical application of biometric facial recognition technology. AI and Society
37(1): 167–175.
Smith M, Mann M and Urbas G (2018) Facial recognition. In: Biometrics, Crime and Security. Routledge.
Spencer D (2020) The Face in Visual Representations of Children. Qualitative Research 22(2): 236–50.
Stability-AI (2023a) SDXL-Turbo Model Card. Available at: https://huggingface.co/stabilityai/sdxl-turbo
(Accessed: 13 January 2024).
Stability-AI (2023b) stable-diffusion-xl-base-1.0. Available at: https://huggingface.co/stabilityai/stable-diffusion-
xl-base-1.0 (Accessed: 11 January 2024).
Stability-AI (2023c) Stable Diffusion Version 2. Available at: https://github.com/Stability-AI/stablediffusion
(Accessed: 16 December 2023).
Stability-AI (2023d) Stable Video Diffusion Image-to-Video Model Card. Available at:
https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt (Accessed: 13 January 2024).
Tiidenberg K and Baym NK (2017) Learn It, Buy It, Work It: Intensive Pregnancy on Instagram. Social Media
and Society 3(1).
Wahle JP, Ruas T, Mohammad SM et al. (2023) AI Usage Cards: Responsibly Reporting AI-Generated Content.
In: Proceedings of the ACM/IEEE Joint Conference on Digital Libraries, 2023. Association for Computing
Machinery.
Wang C and Burris MA (1997) Photovoice: Concept, Methodology, and Use for Participatory Needs Assessment.
Health Education & Behavior 24(3): 369–387.
19
Wang P, Li Y, Singh KK et al. (2021) IMAGINE: Image Synthesis by Image-Guided Model Inversion.
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition:
3680–3689.
Ward J (2017) What are you doing on Tinder? Impression management on a matchmaking mobile app.
Information Communication and Society 20(11): 1644–1659.
Watt D (2015) On Becoming a Qualitative Researcher: The Value of Reflexivity. The Qualitative Report 12(1):
82–101.
Weber M (1949) The Methodology of the Social Sciences. New York: Free Press.
Weng L (2021) What are Diffusion Models? Available at: https://lilianweng.github.io/posts/2021-07-11-
diffusion-models/ (Accessed: 30 January 2024).
Yang L, Zhang Z, Song Y et al. (2024) Diffusion Models: A Comprehensive Survey of Methods and Applications.
ACM Computing Surveys 56(4): 1–39.
Yang Y, Lyu J, Wang Ruixin et al. (2022) A digital mask to safeguard patient privacy. Nature Medicine 28(9):
1883–1892.
Zhang L (2023) Released Checkpoints. Available at: https://huggingface.co/lllyasviel/sd-controlnet-
openpose#released-checkpoints (Accessed: 27 January 2024).
Zhang L, Rao A and Agrawala M (2023) Adding Conditional Control to Text-to-Image Diffusion Models. Epub
ahead of print 2023.
