Co-Creation and Computation: Navigating Creative Agency in AI-Assisted Graphic Design

Heniel Fourie

In Publications Posted September 30, 2025

DOI: 10.5281/zenodo.17235570

Abstract

The rapid proliferation of generative AI image-making platforms has introduced a set of provocative questions for graphic design practice and education — questions that concern not merely workflow efficiency but the ontological conditions under which creative agency is exercised. This paper presents a post-facto account of a hands-on workshop conducted at the Digital Humanities Spring School (September 2025), in which participants were introduced to three generative AI image-making environments: Adobe Firefly, Google Whisk, and open-source latent diffusion models. Drawing on posthumanist philosophy of technology — particularly the work of Barad, Braidotti, Haraway, and Stiegler — I argue that these tools do not simply augment the designer’s established practice but reconfigure the very terms on which creative agency is exercised, distributing authorship across a network of human intentions, algorithmic processes, and training data. I propose the concept of the negotiating co-creator to describe the practitioner’s reoriented role within such systems, and I reflect on what this reorientation demands from design education and from digital humanists who engage with computational creative tools. The paper concludes by proposing that a critically informed engagement with AI image-making — predicated on reflective practice rather than unreflective adoption — is both urgent and consonant with the interrogative values that underpin digital humanities as a discipline.

Keywords: generative AI, graphic design, co-creation, diffusion models, Adobe Firefly, Google Whisk, algorithmic authorship, design education, posthumanism

1. Introduction

The arrival of generative AI image-making platforms in professional and educational design contexts constitutes what one might reasonably describe as a paradigm disruption. Within a remarkably compressed timeframe — roughly from 2022 to the present — platforms predicated on latent diffusion models have moved from experimental curiosity to mainstream adoption, appearing in commercial software suites, professional design workflows, and educational curricula at a pace that has consistently outrun critical scrutiny. Adobe Firefly, launched in March 2023 and subsequently integrated across the Creative Cloud ecosystem, made generative image synthesis available to millions of professional designers without requiring any knowledge of the underlying computational processes. Google Whisk, released in December 2024, displaced the text prompt as the primary mode of input and invited users to guide image generation through the submission of reference images — a development that represents a significant shift in the phenomenology of AI-assisted image-making. The open-source ecosystem of latent diffusion models, anchored by Stable Diffusion and its many community-developed extensions, provides an alternative terrain in which the constraints imposed by corporate platforms are largely absent and technical transparency coexists with considerably greater ethical complexity.

In September 2025, I presented a hands-on workshop at the Digital Humanities Spring School in which participants were introduced to these three environments through a series of structured exercises designed to elicit not merely proficiency but critical reflection. The workshop was not conceived as technical training in any narrow sense; its underlying pedagogical ambition was to surface the conceptual tensions that attend the use of generative AI in creative practice — tensions pertaining to authorship, agency, originality, and the redistribution of creative labour between human and machine. This paper is a post-facto account of that workshop and the theoretical concerns it was designed to explore.

I argue, centrally, that generative AI image-making tools do not simply extend the designer’s existing repertoire — augmenting it as the camera once augmented the painter — but rather reconfigure the ontological conditions under which creative agency is exercised. The designer who engages with these systems does not operate upon a passive material but enters into a relationship with a system that is itself generative, probabilistic, and shaped by the vast, largely opaque archive of human visual production on which it was trained. This redistribution of creative capacity — from a single, clearly locatable human author to a diffuse network of intentions, algorithms, and data — demands a new conceptual vocabulary. For present purposes, I propose the figure of the negotiating co-creator: a practitioner who exercises agency not through unilateral authorial control but through a continuous process of prompt articulation, iterative evaluation, and critical selection within a system whose generative logic remains, in significant measure, beyond direct human command.

The paper proceeds as follows. I first address the theoretical framework that informs my analysis, drawing on posthumanist philosophy of technology and scholarship in the philosophy of computational creativity. I then provide an account of the three platforms demonstrated in the workshop, situating them within the broader technical and institutional landscape of generative image synthesis. I turn next to the workshop itself — its design, its pedagogical approach, and the observations it yielded. I continue with a discussion of the tensions and provocations that emerged, before concluding with reflections on the implications of these findings for design education and for digital humanities practice more broadly.

2. Theoretical Framework: Agency, Entanglement, and the Computational Turn

The question of creative agency in AI-assisted image-making is not reducible to a question of skill, control, or workflow; it is, at a deeper level, an ontological question — a question about the nature and location of the capacity to bring new things into being. In this section, I explore the theoretical resources I find most salient to this question, drawing primarily on posthumanist philosophy of technology and on recent scholarship in the philosophy of computational creativity.

I begin with Barad’s (2007) concept of intra-action, developed within her broader framework of agential realism. Barad argues, against the classical presumption of pre-formed, independently constituted entities that subsequently interact, that what she terms ‘agencies’ do not precede their relational entanglements but emerge through them. Applied to the context of AI image-making, this framework suggests that the creative agency exercised in a generative image-making encounter is not the property of the human designer alone, nor of the algorithmic system alone, but emerges in and through the intra-action between them. The designer who submits a prompt to Adobe Firefly or manipulates reference images in Google Whisk is not deploying pre-formed creative intentions through a neutral instrument; she is entering into a generative entanglement in which the terms of her own agency are simultaneously reshaped by the affordances, constraints, and outputs of the system. This corroborates my argument that the tool reconfigures the ontological conditions of creative practice rather than merely extending them.

Braidotti (2013) provides a complementary theoretical resource in her account of posthuman subjectivity. Braidotti argues for a relational, nomadic conception of the subject — one constituted through ongoing processes of becoming-with rather than through the stable, autonomous, self-identical subjecthood of the Cartesian tradition. The posthuman designer, on this account, is not a bounded creative agent who brings pre-existing ideas into material form but a relational entity whose creative capacities are continuously shaped by the non-human assemblages — tools, platforms, archives, algorithms — within which she operates. This reframing is not, I wish to emphasise, a capitulation to technological determinism; it is, rather, an invitation to think more precisely about the distributed and heterogeneous character of creative practice, in a manner that becomes especially pressing when the tool is itself generative.

Haraway (2016) further enriches this theoretical landscape through her concept of making kin with non-human others — a disposition of attentiveness and response-ability towards the entities, processes, and histories with which one is entangled. In the context of AI image-making, this concept invites practitioners not to treat generative systems as mere instruments of human will but to attend to what these systems do, what histories they encode, and what kinds of visual culture they tend to reproduce or exclude. This is an ethical demand as much as an epistemological one — a demand that I sought to foreground in the workshop through activities requiring participants to reflect on the training data provenance, aesthetic biases, and representational tendencies of each platform.

From the philosophy of technology, Stiegler’s (1998) account of technical objects as exteriorisations of human memory provides a further analytical resource. Stiegler proposes that human beings are constitutively technical creatures — that humanity and technicity co-evolve in a relationship he designates as the pharmakon: the tool is simultaneously remedy and poison, augmenting human capacities while creating new forms of dependency and transforming the very nature of the capacities themselves. Generative AI image-making platforms are, in this sense, a particularly potent instantiation of the pharmakon: they dramatically expand the range of visual forms that a practitioner can produce while simultaneously displacing the technical skills and deliberative processes through which those forms were previously achieved. This ambivalence — that the tool both enables and transforms — pervades the contemporary discourse on AI in creative practice and constitutes one of the central tensions I sought to activate in the workshop.

Hertzmann (2018) provides a philosophically measured intervention in the question of computational creativity that is directly relevant to the concerns of the workshop. Hertzmann argues, contra both the enthusiasm of those who credit AI systems with genuine artistic agency and the scepticism of those who deny the creative character of algorithmically mediated output, that creativity is a fundamentally social and relational concept: what counts as creative is constituted by communities of interpretation, not by any property intrinsic to the process of production. On this view, the question of whether AI ‘creates’ is not answerable at the level of the algorithm but at the level of the social practices through which creative recognition is extended or withheld. This is a conclusion that bears directly on design education: it suggests that the salient task is not to adjudicate the metaphysics of machine creativity but to cultivate the critical capacities through which practitioners can navigate the social, ethical, and aesthetic dimensions of AI-assisted practice.

3. The Generative Toolkit: Adobe Firefly, Google Whisk, and Latent Diffusion Models

The three platforms demonstrated in the workshop represent distinct positions within the current landscape of generative AI image-making — positions differentiated by their underlying architectures, their institutional contexts, their modes of human input, and their implicit theories of the practitioner-tool relationship. In this section, I describe each platform and draw out the conceptual implications most salient to the workshop’s concerns.

3.1 The Technical Ground: Latent Diffusion Models

All three platforms I examine share a common technical substrate — or, in the case of open-source diffusion models, embody it directly. The foundational mechanism is the latent diffusion model, first described at scale by Rombach et al. (2022) and predicated on the earlier formulation of denoising diffusion probabilistic models by Ho et al. (2020). The generative process in these systems proceeds through an iterative denoising procedure: beginning with random noise in a compressed latent representation of image space, the model progressively removes noise according to a learned distribution, eventually arriving at a coherent image. The conditioning of this process on a text prompt — or, in image-to-image variants, on a reference image — is achieved through cross-attention mechanisms that align the denoising trajectory with the semantic content of the conditioning input (Rombach et al., 2022). Crucially, these models do not retrieve images from a database or recombine pre-existing visual elements in any straightforward sense; they synthesise novel images from a learned statistical representation of the visual and semantic patterns present in their training corpora. Dhariwal and Nichol (2021) further demonstrated that appropriately conditioned diffusion models could exceed the generative quality of previously dominant Generative Adversarial Networks across a range of tasks — a finding that significantly accelerated their adoption in applied and creative contexts.

For the purposes of the workshop, I found it important to convey this technical grounding without reducing the session to a lecture on machine learning — not because the technical details are unimportant, but because the conceptual implications of the generative mechanism are most productive when encountered in the context of practice rather than in the abstract. The understanding that generative AI images are not retrieved but synthesised — that they emerge from probabilistic distributions rather than from deliberate compositional acts — is, I argue, prerequisite to any serious critical engagement with these tools.

3.2 Adobe Firefly

Adobe Firefly occupies a distinctive position in the generative AI landscape by virtue of its institutional context and its explicit claims regarding training data ethics. Adobe has positioned Firefly as trained exclusively on licensed Adobe Stock content, openly licensed material, and public domain works — a claim designed to address the copyright concerns that have attended other generative AI platforms and to render Firefly appropriate for professional commercial use. This institutional framing is itself an object of critical interest: Firefly’s ethical positioning is as much a market strategy as a principled commitment, and practitioners engaging with it are well-served by attending to the gap between Adobe’s claims and the fuller complexity of training data governance in the field.

Within the workshop context, Firefly’s deep integration into familiar Creative Cloud environments — most notably Photoshop’s Generative Fill and Adobe Express — made it the most immediately accessible platform for participants already operating within Adobe’s ecosystem. This accessibility is pedagogically double-edged: the frictionless integration of generative AI within established workflows reduces the threshold of engagement but simultaneously diminishes the conceptual distance from which critical reflection is most readily achieved. I return to this tension below.

3.3 Google Whisk

Google Whisk, released in December 2024, represents a significant departure from the text-prompt paradigm that has dominated generative image-making since the emergence of CLIP-conditioned diffusion models. Rather than requiring users to construct verbal descriptions of desired images, Whisk invites them to submit reference images — for subject, scene, and style separately — which the system then synthesises into a novel visual output. This inversion of the primary input mode carries substantial implications for design practice. On one hand, it lowers the linguistic burden on the practitioner and enables a more directly visual mode of ideation; on the other, it introduces a new set of questions about the provenance and ethical implications of the reference images submitted, and it displaces a new form of agency — the capacity to articulate visual intention through language — with a different but equally demanding form: the capacity to select, curate, and combine reference images in ways that generate purposeful and generative outputs.

For participants in the workshop, Whisk consistently elicited a distinctive phenomenological response — a sense of productive surprise at the gap between the input images and the synthesised output, which the system renders as a novel synthesis rather than a composite. This unpredictability, I argued during the session, is not a deficiency to be overcome but a characteristic of the intra-active generative encounter that invites the practitioner to negotiate, rather than command, the image-making process.

3.4 Open-Source Latent Diffusion Models

The third environment examined in the workshop was the broader ecosystem of openly available latent diffusion models — principally Stable Diffusion and its community-developed derivatives — which provide access to generative image synthesis without the corporate mediation and proprietary constraints of the first two platforms. This openness affords significant creative freedom, enabling practitioners to fine-tune models on custom datasets, to access a wider range of stylistic and subject-matter capabilities, and to engage more directly with the technical dimensions of the generative process. It also entails, however, substantially greater ethical complexity: the training data of open-source models has been the subject of sustained critical attention, particularly in relation to the inclusion of images scraped from the web without the consent of their creators and the aesthetic tendencies encoded in those datasets. Attending to this complexity was, for present purposes, an important component of the workshop’s critical agenda — and one that generated the most searching discussion of the session.

4. Workshop Design and Pedagogical Approach

The workshop was designed as a structured hands-on session within the Digital Humanities Spring School, lasting approximately three hours and accommodating participants drawn from a range of disciplinary backgrounds — digital humanists, designers, educators, and researchers — with correspondingly diverse prior experience of AI image-making tools. This heterogeneity of participant background was not incidental but constitutive of the workshop’s pedagogical ambition: I sought to create conditions in which disciplinary perspectives could be brought into productive dialogue around shared practical encounters with the tools.

I structured the session in three phases. The first was a brief conceptual introduction in which I outlined the technical grounding described in Section 3.1, situating the tools within the broader history of generative image synthesis and foregrounding the theoretical framework — agency, authorship, entanglement — that would inform the subsequent reflective discussion. This introductory phase was kept deliberately concise; the workshop’s primary mode was practice rather than lecture, and I was conscious of Bearman et al.’s (2023) caution against pedagogical approaches that treat AI tools as stable objects of neutral knowledge transmission rather than as contested sites of value and power. The second phase comprised three successive hands-on exercises, one with each platform, in which participants were invited to pursue a shared generative brief — an open visual prompt related to the Spring School’s broader themes of computation and culture — using each tool in turn, documenting their outputs and their reflections on the generative process at each stage. The third phase was a facilitated discussion in which participants shared their observations, compared their experiences across the three environments, and began to formulate the conceptual vocabulary adequate to describing what they had encountered.

This structure is consonant with emerging scholarship on generative AI in design education. Hwang (2025) has argued, in a study of graphic design education using text-to-image generation, that the productive integration of these tools requires not merely technical familiarisation but a transformation of educational goals — from the cultivation of execution skills to the development of critical, curatorial, and conceptual capacities. In a similar vein, Bartlett and Camba (2024) have noted that generative AI in design education raises irreducible questions of originality and ethics that cannot be addressed through technical instruction alone but require deliberate pedagogical scaffolding. For present purposes, I understand the workshop not as a training event but as what one might describe, borrowing loosely from the discourse of critical making, as a site of productive entanglement: a space in which participants are invited to encounter the tools in sufficient depth to feel their generative power and their ethical friction simultaneously, and to hold both responses in view.

5. Observations and Emerging Tensions

In this section, I reflect on the observations that emerged from the workshop — in participants’ expressed responses, in the images they produced, and in the conversations that accompanied and followed the hands-on exercises. I do not claim the status of systematic empirical findings for what follows; the workshop was not designed as a research study with controlled conditions and measurable outcomes. I offer these observations, rather, as theoretically informed reflections on practice — provisional insights that I hold to be of sufficient salience to warrant documentation and discussion.

5.1 The Seduction of Ease

The most immediately striking observation was what I term the seduction of ease: participants across all three platforms tended, in the initial phase of engagement, to experience the rapid generation of visually coherent and aesthetically plausible images as intrinsically satisfying — a satisfaction that, at least initially, precluded critical distance. This is not a trivial observation. The aesthetic plausibility of generative AI outputs creates a powerful implicit endorsement of the first result: when a system produces a visually compelling image within seconds of a prompt being submitted, the practitioner is inclined to accept it, to continue from it, or at most to iterate minor variations — rather than to interrogate the assumptions encoded in its composition, the biases of its training data, or the aesthetic regime it silently normalises. This dynamic is, I argue, one of the most significant pedagogical challenges attending the use of generative AI in design education. The tool’s productivity is simultaneously its most attractive feature and its most significant epistemological risk; the speed with which it forecloses alternatives is directly proportional to the speed with which it generates outputs.

5.2 Prompt Literacy and Its Limits

The second observation pertains to what participants frequently described as the challenge of translating visual intentions into effective inputs. With Adobe Firefly and the open-source diffusion models, this challenge centred on text prompt construction — the capacity to articulate in language the visual qualities, compositional arrangements, mood, and stylistic register of a desired image. Participants with stronger verbal-visual translation capacities consistently achieved more directed outputs; those with more developed visual intuitions but less facility in verbal articulation found the text-prompt interface to be a significant constraint on the specificity of their generative engagement. With Google Whisk, this challenge was reconfigured rather than resolved: the ability to select reference images that would interact productively — to understand, in practice, how the system would triangulate between submitted exemplars — required a different but equally sophisticated form of visual literacy. The concept of ‘prompt literacy’ that has circulated in popular discourse on generative AI therefore requires substantial refinement when applied to design practice; it encompasses not only verbal fluency but a complex of visual, conceptual, and systemic capacities that resist straightforward pedagogical definition.

5.3 Authorship Anxieties

The third and most theoretically generative set of observations concerned questions of authorship. Without exception, participants raised — at some point in the facilitated discussion — questions about the legitimacy of the claim to creative ownership entailed in the use of these tools. These questions took diverse forms: some participants were concerned about the relationship between their outputs and the uncredited training data on which the models were built; others were preoccupied with the degree to which the work could be said to be ‘theirs’; still others raised concerns about the displacement of human creative labour in professional contexts. What these diverse anxieties share is a cluster of assumptions — about creativity as origination, about authorship as solitary and unmediated, about the creative work as the expression of a singular human consciousness — that are, as my theoretical framework suggests, already contestable before the question of AI is raised. Hertzmann’s (2018) observation that the attribution of creativity is a social rather than an intrinsic act was, in this context, among the most productively disorienting theoretical resources the workshop could offer: it reframed the question of AI authorship not as a metaphysical problem to be resolved but as a social practice to be examined and, where necessary, renegotiated.

5.4 Productive Friction

Notwithstanding the tensions described above — or, more precisely, through them — the most consistent observation across the workshop was that the moments of greatest creative and conceptual engagement arose not when the tools performed smoothly but when they failed, surprised, or resisted. The unexpected output — the image that diverged from the submitted prompt in ways that were dissonant, ambiguous, or generative of possibilities the participant had not initially envisaged — consistently provoked more animated and substantive reflection than outputs that closely matched prior intentions. This observation suggests that the pedagogical framing of these tools — the manner in which they are introduced and the expectations they are invited to generate — has significant consequences for the quality of critical engagement they elicit. A framework that positions the tool as a means of efficiently executing pre-formed ideas will produce a different practitioner, and a different critical consciousness, than one that positions the tool as a generative partner whose unexpected contributions are occasions for reflection.

6. Discussion: Towards a Critical Computational Practice in Design

The observations reported in the preceding section converge on a set of implications that I wish to develop here — for design practice, for design education, and for the digital humanities community this workshop was designed, in part, to address.

I return, first, to the concept of the negotiating co-creator. This concept is intended to capture a practitioner’s reoriented relationship with the tools of image-making in the generative AI context. The negotiating co-creator does not exercise unilateral authorial control over passive material but enters into a dynamic and iterative process of articulation, evaluation, and response with a system that is itself generative. This is not a loss of creative agency; it is a transformation of its character — a shift from the model of the artist who brings intention to inert matter, to the model of the designer who navigates, selects, and composes within a generative system whose outputs are neither fully predictable nor fully under direct command. The analogy with other creative practices shaped by the introduction of new technical capacities — the photographer who composes within the constraints and affordances of the optical apparatus; the graphic designer who works within and against the possibilities of digital layout software — suggests that this transformation is neither unprecedented nor necessarily to be resisted. What is required, however, is that it be understood rather than merely undergone.

This demand for understanding — for a critically informed engagement with the conditions under which one’s creative practice is being transformed — is the primary implication of the workshop experience for design education. The integration of generative AI tools into design curricula is already well underway, but the critical dimension of this integration lags behind its technical dimension. Hwang (2025) is right to argue that graphic design education in the generative AI era requires new educational goals; I would add that it requires, specifically, a sustained engagement with the theoretical and ethical questions these tools raise — questions of authorship, data provenance, aesthetic bias, and the redistribution of creative labour — as constitutive elements of the curriculum rather than supplementary reflections upon it. The practitioner who can evaluate a generative AI output with the same rigour she brings to any other design decision — asking what assumptions it encodes, what communities it represents or excludes, what creative labour it displaces — is a more genuinely capable practitioner than one who can merely prompt effectively.

The digital humanities dimension of this argument is equally important. The spring school context of the workshop foregrounded a community of practitioners for whom computational tools are simultaneously instruments of research and objects of scholarly inquiry. This dual orientation — as both user and critic of computational systems — seems to me to be precisely what is called for in relation to generative AI image-making. The risk, which the humanities are well placed to identify, is that the adoption of these tools proceeds on the basis of a fundamentally uncritical technocentrism: a disposition in which the impressive capabilities of the technology occlude the critical examination of its assumptions, affordances, and effects. The tools I have described in this paper are, by the standards of creative technology, remarkable. They are also opaque, politically situated, aesthetically partial, and shaped by commercial interests that do not always align with the values of the communities that use them. To adopt them is, for many practitioners, both necessary and productive. To do so without critical scrutiny is, I argue, a failure of the humanistic commitment to reflective practice that both design and the digital humanities, at their best, have always upheld.

7. Conclusion

This paper has offered a post-facto account of a workshop in AI-assisted graphic design conducted at the Digital Humanities Spring School in September 2025, and has sought to develop a theoretical framework adequate to the conceptual challenges the workshop was designed to surface. Drawing on posthumanist philosophy of technology — particularly the work of Barad, Braidotti, Haraway, and Stiegler — I have argued that generative AI image-making tools such as Adobe Firefly, Google Whisk, and open-source latent diffusion models do not simply extend design practice but reconfigure the ontological conditions under which creative agency is exercised. I have proposed the concept of the negotiating co-creator to describe this reoriented practitioner role, and I have argued that design education adequate to this reorientation requires sustained critical engagement with the theoretical, ethical, and aesthetic dimensions of AI-assisted practice, rather than technical familiarisation alone.

The observations drawn from the workshop — the seduction of ease, the complexities of prompt literacy, the anxieties surrounding authorship, and the creative productivity of friction — together suggest that the most significant pedagogical challenges attending generative AI in design are not technical but conceptual. They demand practitioners who can think with and against these tools simultaneously; who can experience the generative encounter as both creative opportunity and critical provocation; and who can situate their own practice within the broader institutional, ethical, and cultural contexts these technologies bring into view.

In a certain respect, this conclusion extends a longstanding concern of the digital humanities: the insistence that the adoption of new computational tools must be accompanied by critical interrogation of the assumptions those tools embody and the practices they transform. The advent of generative AI image-making does not require the digital humanities to develop a fundamentally new critical posture; it requires the application of the one the field has always championed to a set of tools that are, by any measure, among the most consequential to enter creative practice in recent memory.

References

Barad, K. (2007). Meeting the universe halfway: Quantum physics and the entanglement of matter and meaning. Duke University Press.

Bartlett, K. A., & Camba, J. D. (2024). Generative artificial intelligence in product design education: Navigating concerns of originality and ethics. International Journal of Interactive Multimedia and Artificial Intelligence, 8(5), 55–64. https://doi.org/10.9781/ijimai.2024.02.006

Bearman, M., Ryan, J., & Ajjawi, R. (2023). Discourses of artificial intelligence in higher education: A critical literature review. Higher Education, 86, 369–385. https://doi.org/10.1007/s10734-022-00937-2

Braidotti, R. (2013). The posthuman. Polity Press.

Dhariwal, P., & Nichol, A. (2021). Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems, 34, 8780–8794.

Haraway, D. (2016). Staying with the trouble: Making kin in the Chthulucene. Duke University Press.

Hertzmann, A. (2018). Can computers create art? Arts, 7(2), 18. https://doi.org/10.3390/arts7020018

Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33, 6840–6851.

Hwang, Y. (2025). Graphic design education in the era of text-to-image generation: Transitioning to contents creator. International Journal of Art & Design Education. https://doi.org/10.1111/jade.12558

Latour, B. (1993). We have never been modern. Harvard University Press.

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022). High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 10684–10695). https://doi.org/10.1109/CVPR52688.2022.01042

Stiegler, B. (1998). Technics and time, 1: The fault of Epimetheus. Stanford University Press.