Florida International University, doctoral research
Research question
How does a designer’s relationship to visual work in progress shift when inferential discovery can recur across a generative workflow?
In a generative workflow, model inference produces or transforms visual material, and that material can become grounds for the designer’s inference: the designer can take something visible in it to support a conclusion about the work in progress.
Where a workflow invokes model inference recurrently rather than episodically, the work can change through model inference at more points in the process, and the occasions on which the designer finds something consequential can recur. The possibility of recurrent inferential discovery makes this research an inquiry into deciding under an abundance of options.
The research examines how designers exercise agency in responding to inferential discovery across workflows that combine automated production with direct editing, compositing and physical making. It asks how the conditions of their work can enable or constrain what they can do.
Human inference is drawing a conclusion from available information and prior understanding. The word also names the conclusion drawn.
When information is limited, prior knowledge can shape which explanations a person considers and how plausible they find them (Tenenbaum and colleagues).
In his account of human reasoning with beliefs, Boghossian explains that a person draws a conclusion because they take other beliefs to support it. In design work, a designer can take something visible in the work as grounds for a conclusion about it.
In machine learning, inference applies a trained or fitted model to supplied input to produce a prediction or generated result.
In neural networks, training adjusts numerical weights and other parameters using data; inference processes the supplied input through the network using those learned values (Google machine learning glossary).
A text-to-image system, for example, uses a trained model to generate an image guided by a prompt (Rombach and colleagues). That image becomes visual material from which the designer can draw an inference.
In natural language processing, a trained model can perform a task called natural language inference by predicting whether a premise entails, contradicts, or is neutral toward a hypothesis (Bowman et al., 2015). Using inference for both people and models can encourage anthropomorphism when the shared term is taken to imply human reasoning or understanding. Following McDermott (1976) and Shanahan (2024), this account describes model inference through the mechanisms that produce its results.
In this research, an inference a designer finds becomes an inferential discovery as the designer determines what it means for adaptation and continuation of their work.
Design adaptation names the changes a discovery can prompt in the material being developed and the aims guiding it. It develops through situated judgments, and a judgment can shape the work the designer encounters in subsequent decisions, as when the designer prepares what the next generation receives (Suchman, 1987; Hutchins, 1995; Kirsh, 1995).
Here, abundance includes the information available in generated material, the possibility of further discovery, and the range of responses available across blended practice. A blended practice combines generative computation with other ways of making and editing, within one application or across several.
Simon (1971) observed that “a wealth of information creates a poverty of attention.” Abundance in this inquiry carries that cost: each generated result, each route through blended practice and each possible further discovery asks for judgment.
A designer can pursue something found in a generated result through another kind of operation. The response may develop the existing material, change the conditions of further production, or change the intention guiding the work. Further discovery can remain possible whichever response the designer chooses. That possibility can affect present decisions about continuation, as the designer considers what further work might reveal.
Further discovery need not depend on generating another output. Relationships can also become apparent through working with material already available. In his account of designing as reflective conversation with materials, Schön describes how designers’ moves can reveal unintended consequences and how their intentions can develop through the work. A discovery can change the designer’s understanding without being incorporated into a finished artifact.
More available material does not establish what matters to the designer. An undesirable result can occasion a consequential discovery.
Further discovery can occur after work has come to depend on an earlier decision. Reconsidering that decision can then affect the work that depends on it. What pursuing an interpretation requires depends partly on the possibilities for further work that remain after an operation. Those requirements can affect whether and how the designer continues.
Green and Blackwell call this resistance to change viscosity: changing one part of a working structure can require further changes to connected parts.
Continuation concerns what becomes of the work as the designer responds to it, including the possibility of deciding not to proceed. The work can be sufficient for its purpose while further discovery remains possible. A decision to stop can therefore rest on what the work accomplishes. Looking only at a finished output would leave much of the designer’s response and its consequences unexplained.
How the designer continues the work depends partly on their reasons and agency—the exercise of their capacity to act. The amount of automated production alone does not establish that capacity or explain how it is exercised.
The research has a technical component and a qualitative component: the Inference Resolution Map assesses software operations, and a visual workbook examines the designer’s response through evolving workflow examples.
The qualitative component is a visual workbook in Miro of evolving workflow examples. The examples connect supplied material and relevant settings with working versions and, where a discovery is documented, the designer’s account of it and what follows from the response. Through the examples, the research examines how the designer’s understanding, aims and judgments develop within the work, including what makes a result “coherent enough”: sufficient for its purpose, even when further discovery remains possible. Comparing the examples can test and refine the vocabulary for describing discoveries, what the available evidence supports about them, and the forms continuation takes. Where the records support a distinction the existing terms do not describe, the vocabulary is revised or extended.
A workflow example connects the two components when it links the designer’s response and a working version to relevant operation assessments, including provisional X coordinates where supported. X summarizes the assessed operation’s predictive task and implementation; the accompanying account describes how the result is produced, as documented in the sources for that operation. The workbook describes what the designer finds significant and what responding requires, including what needs to be rebuilt when an earlier decision changes. For example, a generated clip may remain adjustable in timing and placement within an editable composition while its depicted movement stays fixed in the generated frames. If that composition is nested in another sequence, changing the clip’s timing changes when its movement occurs in that sequence. The designer may then find a different relationship between the movement and the next shot, prompting them to reconsider the earlier edit or what they want the sequence to achieve.
Shneiderman’s counterpoint also shapes the research’s direction. Greater automation need not be treated in advance as either a loss or a gain in the designer’s capacity to continue. A particular arrangement of model determination and retained controls may make one response possible while making another difficult. What that means is examined in relation to the work, the designer’s developing intentions and the response under consideration.
Three revision capabilities describe which responses an operation keeps possible. Shneiderman explicitly discusses interfaces that support revised intent and reversible actions. That supports a connection: these capabilities may describe part of the human control that Shneiderman keeps separate from automation. They make particular questions about further work available for examination:
| Capability | What it helps investigate |
|---|---|
| Recovery and reuse | What earlier material remains accessible and usable when the designer considers another response. |
| Scope of available revision | Which aspects of the work the retained controls can change. |
| Preservation of dependent later work | What remains applied, and what needs to be rebuilt, when an earlier decision changes. |
These descriptions can help explain control over revision across continuation without a second coordinate. They also give the research a particular concern: how earlier decisions and subsequent work remain connected when something is reconsidered. Considering these capabilities does not make preserving earlier decisions a goal in itself; their significance depends on what the designer is trying to pursue at that point.
The Inference Resolution Map assigns provisional X coordinates on a proposed scale of how trained or fitted prediction affects an operation’s returned result. Because what prediction determines depends on the task and its implementation, the map assesses named software tasks, provisionally defined here as operations. For each operation, the assessment identifies the software version, model where applicable, and settings relevant to its predictive task and computation.
It then asks what the designer specifies, what specified computation determines, and what depends on prediction from a trained or fitted model in the returned result. The predictive task receives graded ratings describing which parts of the result prediction determines (scope), how much the supplied material and task leave the model to resolve (latitude), and which relationships it resolves together (coordination). X combines these ratings with implementation counts of the model’s parameters and computation. X is 1 when trained or fitted prediction makes no contribution to the returned result; where prediction contributes, X moves toward 0 as the ratings and transformed counts increase:
Ratings describe the task the prediction performs. Each is scored 0–5 against written anchors, and the support for a rating depends on the documentation, code, model files or papers available for the assessed implementation.
No placed operations match these filters.
Values of X are rounded to three decimal places.
The contribution I am developing is a shared vocabulary for designers and a method for comparing operations, and for comparing designers’ responses to what those operations produce, so that claims about agency and automation can be examined against both technical evidence and documented visual practice. When I explain to clients, and to colleagues in the AI innovation department where I work, how I produced a visual result, they can come to understand the result differently, and with that understanding they can say more precisely what they want. Where the workbook examples document what the designer supplied, what they found in the developing work, what they determined it meant for continuation and what followed from their response, others can examine those judgments.
Green and Blackwell (1998) describe cognitive dimensions as discussion tools that name concepts people may have half-formulated, so that discussants can use the name rather than explain the idea again each time. The workbook’s terms, such as “coherent enough,” name distinctions of this kind, and following Gaver and Bowers (2012), I develop them through the examples they describe. Schön and Wiggins (1992), drawing on Vickers, describe intentions that develop through appreciation of an intermediate result, and judgments that designers can make tacitly. The workbook accounts can record how intentions develop and state the criteria behind such judgments, as the copper-figure account below does for the do-rag and for the cultural fidelity I weighed against patina texture. Yuan et al. (2026) present concepts, grounded in a single artist’s practice, that others can test, adapt, contest or extend, and propose documentation as both creative support and a record for collective sharing. This research develops concepts and documentation of this kind for inferential discovery across visual generative workflows, so that someone working on very different material can take up the distinctions and questions rather than the workflow. The artist in Yuan et al. (2026, p. 4190) works from “working conceptions of AI operation”; the workbook examples can also link the designer’s account to assessments of the operations involved.
As a further possible application, an expert assessment of human inference and automation in the development of a work could examine, for a relationship such as the copper figure’s head, do-rag and crown, what the designer constructed or specified, what prediction resolved from that input, and how the designer judged and responded to the result, drawing on working versions, the designer’s account and technical evidence. A buyer or client could use such an assessment to choose work according to what they value; it would need its own definition, separate from X, which describes operations.
The same records could inform a benchmark comparing production arrangements. ReDesign (Yun et al., 2026) compares the same vision-language model, tools and memory organized as a linear sequence of tool calls with validation at the end and a branching reconstruction process that verifies each proposed layer decomposition, showing that the arrangement of a workflow can be evaluated alongside the model.
In this work, I depicted Jaylen Brown as the Statue of Liberty. While developing the figure, I found that the head needed a bridge to the crown and considered a do-rag as a way to connect them. The process included generation, Photoshop Generative Fill, the Pen tool, layer masks and color grading. When generation failed to produce the required relationship, I constructed and composited the do-rag in Photoshop, then used the composite as the reference image for further generation.
Constructing the composite made the head–do-rag–crown relationship I had worked out part of the material supplied for further generation. This gives the technical assessment a specific question: how did the generative operation use that supplied relationship, and what did it determine in the returned material? Assessing that operation would require its software and model versions, relevant settings and technical evidence about how it uses reference material.
I valued the do-rag as a culturally recognizable part of the portrayal of Jaylen Brown, as well as a way to connect the head and crown. In discussing this work, I described wanting details that would resonate with African-American viewers, within a broader aim of making work that moves people or helps them identify with themselves. I accepted some loss of patina texture to retain the cultural fidelity I wanted in the do-rag, and judged that compromise sufficient for the work’s purpose.
Earlier work examined how different input modalities supply partial spatial information, how generative systems resolve that information, and how people interpret the resulting structures.
While teaching architecture in Miami, I explored how people could learn to use tools through experiences tailored to them. My early research considered visualization and gamification as ways of making learning interactive, with climate education as an initial context. The emphasis subsequently moved toward personalized education and how learning could respond to individual needs.
As generative practice became a stronger focus, my attention shifted toward what designers encounter while working with these tools. The current dissertation focuses on the designer’s response to inferential discovery.
Literature review and research-paper drafts examined 3D visualization, gamification, and AI-supported personalization in climate education.
A Digital Futures paper considered prospective changes to architectural education, curricula, and relationships between educators and students.
Personalized Learning in Higher Education: AI-Driven Analytics and Competency-Based Models as Catalysts for Reform placed adaptive learning and individual needs at the center of the inquiry.
Listening is an important part of how I learn and develop my work. My reading and listening span technology, creativity, perception, culture, and lived experience.
Linked titles and covers open their audiobook editions on Audible.
78 titles
Algorithms to Live ByBrian Christian and Tom Griffiths
The Thinking MachineStephen Witt
Blue MindWallace J. Nichols
GoogooshGoogoosh and Tara Dehlavi
Tupac ShakurStaci Robinson
CasteIsabel WilkersonSources cited or held for the current research. Updated 3 November 2026.
137 sources
Technical documentation and research supporting the operation records on this page.
32 sources