Insights Informed Generative AI for Design: Incorporating Real-world Data for Text-to-Image Output

• Current interior-design text-to-image models provide visual representations but lack actionable metrics and real-world data such as sustainability, cost, and material properties.Source: §1 · p. 2
• The proposed approach shifts generative AI from visual output generation toward a decision-support system enriched with material-related CO₂e insights.Source: §1 · p. 2
• Because early design decisions strongly affect environmental outcomes, the study tests whether CO₂e feedback can support more informed prompt iteration from the earliest workflow stages.Source: §1 · p. 2
• The broader aim is to merge materials, sustainability, and generative models into a streamlined workflow that supports more responsible, informed AI-assisted design.Source: §1 · p. 3

• Text-to-image models such as DALL-E enable designers to rapidly transform natural-language concepts into speculative interior visualizations.Source: §2.1 · p. 3
• Interior design remains iterative because designers repeatedly critique, evaluate, and adjust evolving artifacts during open-ended problem solving.Source: §2.2 · p. 3
• Early sustainability decisions can strongly affect long-term environmental impact, while treating sustainability as an afterthought makes later optimization less effective.Source: §2.3 · pp. 3–4
• The paper argues that generative outputs should increasingly be accompanied by measurable environmental insights that guide responsible architectural solutions.Source: §2.3 · p. 4

System Implementation

• DALL·E 3 generates interior images, after which vision-language processing identifies up to ten architectural materials for database matching and CO₂e retrieval.Source: §3.1 · pp. 5–7
• The pipeline combines a generative image model, vision-language material descriptor, material mapping agent, database connector, and insights interface.Source: §3.1 · p. 5, Figure 2 · p. 5
• The interface presents generated images with identified materials and CO₂e scores so users can refine prompts using visual and sustainability feedback.Source: §3 · p. 4, §3.1 · p. 7

Experiments and User Studies

• Each participant completes five prompt iterations per test, and the study includes nine design professionals and students who complete all three conditions.Source: §3.1 · pp. 7–8
• The user study uses three tests: no sustainability consideration, a sustainability goal without metrics, and a sustainability goal with the top ten materials and associated CO₂e values.Source: §3.1 · p. 7
• The study uses separate text-to-image and text-to-image-plus-insights pipelines, with the latter providing material CO₂e information alongside Test 3 outputs.Source: Figure 5 · p. 8, Figure 4 · p. 8

Cross-test comparison

• The three tests produced distinct responses, showing that sustainability prompts and embedded material metrics affected participants’ prompting behavior.Table 2 summarizes participant responses across the experimental conditions.Source: §4 · pp. 10–11, Table 2 · p. 11

Test 1

• Without sustainability goals or metrics, approximately 72% of Test 1 participants were satisfied, while most omitted sustainability from their prompts.Participants primarily evaluated outputs through visual and thematic alignment, establishing a passive baseline.Source: §4 · p. 10

Test 2

• Test 2 increased sustainability-oriented behavior to ~83% and satisfaction to ~77%, but participants relied mainly on symbolic strategies such as greenery and natural light.These responses reflected an “eco-aesthetic” rather than specific material trade-offs.Source: §4 · p. 10

Test 3

• In Test 3, sustainability informed 100% of design decisions, but satisfaction fell to ~50% as users faced higher cognitive load and decision fatigue.Participants incorporated specific materials such as “rammed earth,” “mass timber,” and “lowcarbon substitutes.”Source: §4 · p. 10
• Participants found the material insights helpful in some cases, but also reported insufficient metric context and difficulty interpreting nuanced material intent.Comments described the system as “a good way to summarize impact” while criticizing the lack of comparative baselines and its tendency to force decisions into carbon logic.Source: §4 · pp. 10–11

Implications

• Actionable metrics can encourage deliberate prompt iterations and materially grounded decisions, but complex trade-offs increase cognitive burden and reduce design satisfaction.The system must balance accuracy and interpretability with creative agency and user-centered choices.Source: §5 · pp. 11–12

5.1 Challenges

• Inconsistent material units make CO₂e values difficult to compare, and emissions may not scale linearly because of fixed operational or manufacturing costs.The dictionary combines cubic meters, square meters, and kilograms, so apparently high or low values may reflect different measurement bases.Source: §5.1 · p. 12
• Dictionary granularity can misclassify functionally different materials, while DALL-E 3 does not provide precise material descriptions for fully automated matching.Cross-referencing the dictionary, prompt, and generated image could mitigate these inaccuracies.Source: §5.1 · p. 12

5.2 Future Works

• Consistent material units would improve interpretability, but robust design support should also incorporate cost, durability, procurement, and other practical criteria.The discussion frames these additions as a progression from a speculative aid toward a more robust decision-support tool.Source: §5.2 · p. 12
• Linking generated designs to specific materials and carbon footprints, and extending insights from 2D images to 3D software, could improve practical applicability.These extensions are identified as future directions for translating visual concepts into real-world projects.Source: §5.2 · p. 13

Responsible AI Parallels

• The proposed post-generative framework can support transparency, user agency, and adaptability by letting designers refine outputs using interpretable data.This approach differs from end-to-end generation that may overlook regional availability, preferences, or cultural nuances.Source: §5.2 · p. 13
• Embedding CO₂e metrics and material references can anchor idealized AI images in actionable data and help designers and clients discuss environmental trade-offs.The framework is presented as a way to reduce false expectations and support more informed design collaborations.Source: §5.2 · p. 13

• The pipeline provides CO₂e metrics for up to ten materials in generated interior designs, and this information significantly shapes user behavior.Users reported more iterative prompt refinements and deeper consideration of environmental impact.Source: §6 · p. 13
• Including actionable sustainability data promotes a more informed, user-centric approach to generative AI design workflows.Source: §6 · p. 13
• Additional metrics such as cost, resource availability, and life-cycle impact could further improve the system’s relevance to design stakeholders.These extensions are framed as a future direction for balancing creativity with practical constraints.Source: §6 · pp. 13–14