---
title: 'WeDesign: AI-Driven Urban Consultation'
url: https://www.emergentmind.com/topics/wedesign
type: topic
---

# WeDesign: AI-Driven Urban Consultation

Searching arXiv for the specified paper and related metadata.
WeDesign is a custom platform that integrates generative text-to-image methods, specifically Stable Diffusion XL, into community consultations for urban public space design, with the stated aim of supporting more equitable participatory processes in urban planning [2508.19256]. It was examined through a half-day workshop in Montreal involving five mixed focus groups and through additional semi-structured interviews with urban planning professionals. The reported results indicate that immediate visual outputs can facilitate creativity and dialogue, while also revealing persistent limitations in representing marginalized groups’ needs, local architectural specificity, and bilingual interaction. The system and its evaluation therefore occupy a specific position within participatory urban design research: generative AI is treated not as an autonomous design agent, but as an interactive consultation aid whose value depends on workflow, facilitation, and institutional safeguards [2508.19256].

## 1. System architecture and model integration

WeDesign combines a React-based front-end, a Python/Flask back-end, local hosting of Stable Diffusion XL on a GPU server, and PostgreSQL-based persistence for session and interaction data [2508.19256]. The front-end supports scenario management, prompt entry, image previews, and user reactions such as hearts and votes. The back-end mediates between the user interface and the text-to-image engine, handles session states, archives prompts and metadata, and exposes REST endpoints for generating and retrieving images. Stable Diffusion XL runs locally on a GPU server, exemplified by an NVIDIA A100, and is accessed through an internal HTTP API. Workshop sessions, scenario definitions, prompts, generated image URIs, user reactions, and timestamps are recorded in PostgreSQL.

The integration workflow is explicitly specified. The UI sends a POST request to `/api/generate` with JSON containing `"prompt"`, `"negative_prompt"`, and a `"seed_list"` of four seeds. The Flask application then invokes the Stable Diffusion XL Python client through the diffusers library, using a call of the form `images = pipe(prompt, negative_prompt=neg, guidance_scale=7.5, seeds=seed_list)`. Generated images are stored, their URLs are returned to the front-end, and metadata including prompt text, user ID, group ID, and timestamp are persisted in the database [2508.19256].

Images are generated in batches of four per prompt, with adjustable random seeds. This four-image batch structure is significant because it couples model stochasticity to collective deliberation: participants compare multiple variants rather than treating a single output as authoritative. A plausible implication is that the system’s architecture was designed not merely for inference throughput but also for facilitating comparative judgment during group consultation.

The paper did not report formal latency or accuracy benchmarks. It instead proposes candidate quantitative measures for follow-up evaluation, including average generation latency,
$$
L \;=\;\frac{1}{N}\sum_{i=1}^{N}t_i
$$
where $t_i$ is the round-trip time for the $i$th prompt over $N$ requests, and a representational-accuracy-style measure for inclusivity features [2508.19256]. These formulas are presented as prospective evaluation instruments rather than validated experimental results.

## 2. User interface and consultation workflow

The user interface is organized around a set of components identified in Figures 2–3 of the paper: a Scenario Panel, Prompt Editor, Image Canvas, Reaction Buttons, and Session Log [2508.19256]. The Scenario Panel allows creation, viewing, and switching between scenarios labeled A–E. The Prompt Editor contains text fields for “prompt” and “negative prompt,” with on-the-fly synonym and translation suggestions. The Image Canvas displays four generated thumbnails per prompt, with click-to-enlarge behavior. Reaction Buttons include a heart icon for marking preferred images, with voting buttons described as a potential future feature. The Session Log maintains a live list of past prompts and images for the group, supplemented by a facilitator notes sidebar.

The workshop workflow is also explicitly sequenced. In pre-session setup, facilitators define five main scenarios in WeDesign, with examples including a suburban park and a pedestrian promenade, and participants are divided into five mixed groups composed of one AI specialist, one urban professional, and two to three citizens. Brainstorming then proceeds for approximately 30 minutes, during which each group reviews the scenario and lists desired features such as accessibility, seating, and lighting. This is followed by a prompt iteration phase lasting three hours total, with a lunch break. Citizen leads type an initial prompt, AI specialists advise on keyword ordering and negative prompts, the system generates four images, the group discusses the outputs, citizens mark preferred images with hearts, and the urban professional annotates feasibility notes. Prompts are then refined—for example by adding “realistic” or removing “overhead view”—and resubmitted. This loop repeats until the group is satisfied or time elapses. In closing, each group selects its top two to three images, presents them to the full workshop, and facilitators record observations and written feedback [2508.19256].

This workflow positions prompting as a collaborative and iterative design practice rather than an individual interaction with a generative model. The assignment of roles—citizen lead, AI specialist, urban professional, facilitator—indicates that WeDesign was evaluated in a socio-technical configuration where the interface alone was not expected to guarantee inclusion. This suggests that the platform’s operation is inseparable from moderation and expert scaffolding.

## 3. Empirical study design in Montreal

The Montreal workshop included 14 citizens, five computer scientists, five urban architects or planners, and five note takers, distributed across five parallel focus groups during a single half-day session of approximately four hours [2508.19256]. The citizen cohort is described as balanced across age, 2SLGBTQ+, wheelchair users, racial or religious minorities, French-dominant, English-dominant, and bilingual participants. In addition, six one-hour semi-structured interviews were conducted with municipal facilitators and urban planning experts.

Data collection combined several sources. All group discussions were audio recorded and documented through field notes. The platform automatically logged 440 textual prompts and associated images. The six expert interviews were transcribed, and group-level metadata were retained, including which group produced which prompt and which image received hearts [2508.19256]. This mixture of interaction logs, observational records, and interviews reflects a multi-modal qualitative dataset, with some structured digital traces suitable for later quantitative or computational analysis.

The analytical approach combined open coding, Latent Dirichlet Allocation, and word-cloud analysis. Open coding, attributed to Creswell and Creswell (2022), was used for manual identification of themes including inclusivity, language barriers, AI limitations, platform features, and facilitation. LDA topic modeling, attributed to Blei et al. (2003), was applied separately to interview transcripts, workshop notes, and the 440-prompt corpus, using standard settings with approximately $K \approx 6$ topics. Topics were labeled with phrases such as “model biases,” “bilingual friction,” and “community ownership.” The paper gives the underlying generative equation as
$$
p(w\,|\,d) \;=\; \sum_{z=1}^{K}p(w\,|\,z)\,p(z\,|\,d)\,.
$$
Word-cloud analysis was then used to count prompt-token frequencies, including terms such as “mobility,” “greenery,” and “accessibility,” and to cross-reference those frequencies with coded themes [2508.19256].

Methodologically, the study is structured as an exploratory evaluation rather than a controlled benchmark. Because the paper does not report formal model-comparison experiments, latency tests, or quantitative outcome metrics, its evidentiary strength lies primarily in documented interaction patterns, participant testimony, and thematic synthesis.

## 4. Reported benefits in participatory design

Participants reported that immediate visual outputs facilitated both creativity and dialogue [2508.19256]. One quoted response characterizes the process as collective brainstorming: “It feels like we’re all brainstorming together, seeing what works or doesn’t, instead of just listening to a lecture.” The LDA analysis reportedly identified co-occurrences including “fun,” “try again,” “prompt tweak,” and “discuss,” which align with this interpretation of the platform as a mechanism for iterative engagement rather than static consultation.

The paper further states that immediate visuals functioned as boundary objects, enabling non-experts to articulate preferences [2508.19256]. In this framing, generated images mediate among heterogeneous participants—residents, AI specialists, architects, planners—by providing shared referents for discussion. An expert quotation captures both the affordance and its constraint: “An infinity of images can create expectations… but early visuals help non-specialists own the concept.” The first clause warns of overproduction and expectation inflation; the second identifies the consultative value of early-stage visualization.

More broadly, the discussion section argues that generative AI lowers the skill barrier to visual design and transforms passive town-hall formats into collaborative co-creation spaces [2508.19256]. These claims are consistent with the workshop workflow, in which residents actively steer prompts, react to variants, and participate in iterative refinement. A plausible implication is that WeDesign’s main contribution is procedural rather than purely representational: it reconfigures participation by making image production sufficiently rapid and interactive to be embedded in live group deliberation.

## 5. Limitations, failure modes, and contested representations

The reported limitations are specific and substantive. One major issue concerns the visualization of mobility needs. Participants stated that when they entered “wheelchair-friendly benches,” the generated image produced “random wheelchairs everywhere,” while pavement slopes and ramp clearances were never accurate [2508.19256]. This indicates a failure to render accessibility not merely as visible tokens but as spatially and functionally correct environmental design. The distinction is important: token insertion can simulate inclusivity iconography without representing the physical requirements that make spaces accessible.

A second limitation concerns local architectural accuracy. Attempts to reference “Old Montreal” often yielded generic European façades rather than recognizable local built elements. One participant summarized this mismatch: “I wanted Notre-Dame or cobblestones; instead I got a vague old-town look” [2508.19256]. This suggests a locality gap between prompt intent and model priors, where geographically grounded cultural references are absorbed into broader stylistic stereotypes.

A third limitation is bilingual friction. French prompts reportedly yielded lower image quality, whereas English prompts risked losing local nomenclature. The LDA analysis surfaced a topic containing “language,” “translation,” “confusion,” and “culture” [2508.19256]. This is not merely an interface inconvenience; it has implications for deliberative equity because uneven language support may privilege participants more comfortable with the language producing higher-fidelity outputs.

These limitations are discussed alongside ethical concerns about bias in training data and power asymmetries in consultations [2508.19256]. The paper states that training-data bias can perpetuate tokenism or exclusion of marginalized groups, and that uneven language support may privilege English-speaking participants over French-dominant ones. In this respect, the platform does not eliminate the longstanding problems of participatory planning; rather, it can reproduce them in model-mediated form.

## 6. Proposed extensions and evaluation framework

The paper records several platform features requested by participants and interviewees. These proposals are summarized below.

| Feature | Description |
|---|---|
| Region-specific in-painting | Brush tool to edit parts of an image without full regeneration |
| Multilingual support | Automatic prompt translation (French↔English), synonym suggestions |
| Group voting | Collective ranking/discarding of variants via buttons |
| Preference sliders | Weight criteria (e.g., accessibility vs. biodiversity) to steer generation |
| Open-source codebase | Public inspection of model fine-tuning, data provenance |

These requested extensions indicate that participants wanted more granular control, better linguistic mediation, stronger collective-decision affordances, and greater transparency [2508.19256]. Region-specific in-painting addresses the observed problem of global regeneration producing off-target local details. Multilingual support responds directly to bilingual friction. Group voting and preference sliders would formalize collective preference aggregation. An open-source codebase is framed as enabling public inspection of model fine-tuning and data provenance.

The paper also proposes an evaluation framework for future studies. The Representational Accuracy Score (RAS) is defined as
$$
\text{RAS}
\;=\;
\frac{1}{M}\sum_{j=1}^{M}
\frac{\#\{\text{features}_j\text{ correctly rendered}\}}
{\#\{\text{features}_j\text{ requested}\}} \,,
$$
where $M$ is the number of design scenarios. The Engagement Index (EI) is defined as
$$
EI
\;=\;
\frac{H + Q + R}{T}
\quad\text{where}\quad
\begin{cases}
H = \text{\# hearts (preference flags)},\\
Q = \text{\# participant questions/comments},\\
R = \text{\# prompt refinements},\\
T = \text{session time (minutes)}.
\end{cases}
$$
[2508.19256]

These measures are explicitly prospective. They do not constitute reported experimental findings from the Montreal study. Their significance lies in providing candidate operationalizations for two central concerns in participatory generative design: whether the system renders requested features correctly, and whether it sustains meaningful interaction over time.

## 7. Implications for participatory urban design

The paper concludes that generative AI can broaden participation and enable iterative interactions, but only if accompanied by structured facilitation approaches [2508.19256]. Facilitators are described as needing to manage expectations and communicate that AI outputs should not be treated as final designs. They must also contextualize generated imagery in relation to practical constraints such as budget and zoning. This emphasis on facilitation is central: the platform’s outputs are conceptual artifacts within a planning process, not substitutes for planning judgment.

The discussion also foregrounds transparency and trust. The authors state that transparency in model provenance and open-source code can help address community trust [2508.19256]. This is consistent with the participant demand for an open-source codebase and with the broader concern that opaque generative systems may embed inaccessible assumptions about data, fine-tuning, and representational defaults.

Within participatory urban design, WeDesign thus exemplifies a bounded use of generative AI: rapid visual synthesis is deployed to support dialogue, but the study documents persistent failures in accessibility representation, local specificity, and bilingual equity. The overall evidentiary pattern does not support the view that text-to-image systems straightforwardly democratize planning. Instead, the findings support a narrower conclusion: under carefully structured workshop conditions, they may function as consultation aids that broaden participation while simultaneously requiring explicit safeguards against bias, misinterpretation, and linguistic exclusion [2508.19256].

Source: https://www.emergentmind.com/topics/wedesign