Choosing the Models: Why Open Source, Why Europe, Why It Matters

The choice of which AI models to use is not a neutral technical decision. It is political, ethical, and strategic.

This project deliberately works with European, open-source foundation models: BLOOM (developed by the BigScience initiative, coordinated by Hugging Face with French and European research institutions), Mistral (Paris-based, Apache 2.0 license), Stable Diffusion (CompVis, Stability AI, and LMU Munich), and Public Diffusion (Spawning, by Holly Herndon and Mat Dryhurst — a consent-driven generative image model that emphasizes ethical use of training data).

There are several reasons for these choices. First, transparency: open-source models allow full inspection of architecture, weights, and training data — essential for the bias audits that are central to this research. Second, EU AI Act compliance: the European regulatory framework increasingly requires transparency and bias documentation for AI systems. Working with European models positions this research within that regulatory conversation. Third, ethics: Public Diffusion's consent-based approach to training data aligns directly with the feminist data ethics framework of this project.

The technical pipeline involves fine-tuning these models using DreamBooth and LoRA (Low-Rank Adaptation) — a method that allows efficient model customization with as few as 5–10 reference images and minimal computational resources. This is crucial for a practice-based PhD: it means the feminist architectural dataset curated in this project can directly shape how the AI generates spatial visions.

The distinction between language models (BLOOM, Mistral) and image generation models (Stable Diffusion, Public Diffusion) is important. The language models support the analytical and textual components — prompt engineering, text analysis, conceptual exploration. The diffusion models generate the visual outputs that form the core of the artistic practice. Both are necessary; neither is sufficient alone.

References: BLOOM. https://huggingface.co/bigscience/bloom Mistral AI. https://mistral.ai Stable Diffusion. https://github.com/Stability-AI/stablediffusion Public Diffusion (Spawning). https://spawning.ai/public-diffusion Ruiz, N. et al. (2022). DreamBooth: Fine Tuning Text-to-Image Diffusion Models. https://arxiv.org/abs/2208.12242