Beyond Renders: Insight-Informed Generative AI for Post-Generative Grounding via Vision–Language Models and Sequential Reasoning
Name
Gupta-richag-smarchs-arch-2026-thesis.pdf
Size
4.87 MB
Format
Adobe PDF
Checksum (MD5)
277ac38b41323a0d9ddd999303cee293
Author(s)
Gupta, Richa
Advisor(s)
Knight, Terry W.
Nagakura, Takehiko
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
Generative diffusion models excel at high-fidelity image synthesis but lack inherent mechanisms for grounding their outputs in factual, domain-specific data. This creates a critical "expectation gap" in domains like architectural design, where compelling but unverifiable AI visuals become detached from the material, cost, and environmental realities of construction. Existing approaches focus on steering AI models during training or inference, while post-generative grounding for downstream decision-making remains underexplored.
This thesis presents IIGenAI (Insight-Informed Generative AI), a post-generative grounding framework that augments AI-generated interior design images with real-world material data through vision-language models (VLMs) and sequential reasoning. The system employs structured prompting: few-shot material identification, semantic alignment to a curated materials database, and API-based retrieval of embodied carbon (CO₂e) metrics. This enables human-AI collaboration through feedback-driven prompt refinement based on grounded data and material aware AI generation. The primary aim of this human-in-the-loop is to shift design AI from purely visual output to a decision-support system enriched with actionable insights, transforming early-stage design workflows from speculation to specification.
A user study with designers compared three conditions: (1) baseline generation, (2) sustainability goals without metrics, and (3) IIGenAI with material CO₂e insights. IIGenAI prompted nearly all participants to explicitly address environmental impact (vs. 83% and 5%), with greater prompt divergence (cosine similarity 0.83 vs. 0.88-0.92) and increased lexical diversity (0.42), reflecting material-specific language ("rammed earth," "mass timber"). However, satisfaction decreased from 72-77% to 50%, indicating cognitive load and tension between aesthetic and environmental priorities and interactions. The approach transforms design dialogue, stakeholder engagement, and accountability by grounding early-stage visuals in verifiable data, creating a shared evidence space for design tools and designer-client-AI collaboration maintaining the professional rigor.
The work addresses responsible AI through post-generation validation with applications beyond architectural design. Future work includes multi-objective optimization, BIM integration, multi-modal interactions (sketch/voice/3D) and agentic pipelines for deploying generative AI in high-stake creative domains.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Massachusetts Institute of Technology. Department of Architecture
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
