A new framework developed for Florentine leather craftsmanship explores how artificial intelligence can accelerate design and prototyping while preserving the distinctive knowledge, techniques and visual identity of artisan workshops.
Generative AI is increasingly entering creative industries, offering designers the possibility to turn written descriptions, sketches and reference images into visual concepts within seconds. For traditional craftsmanship, however, this opportunity comes with a fundamental concern: if generative models are trained primarily on large collections of generic online images, will their output gradually make handcrafted products look more alike?
Researchers from the University of Florence and the Foundation for Research and Innovation have developed a Heritage-Preserving Generative AI framework designed to address precisely this problem. Tested in the Florentine leathercraft sector, the approach combines generative AI with the historical archives and specialist knowledge of individual artisan workshops.
The aim is not to automate craftsmanship. Instead, AI becomes a collaborative design instrument that helps artisans experiment faster while keeping them in control of the creative process.
Turning artisan heritage into a digital resource
Traditional craftsmanship is more than the physical object being produced. It also encompasses knowledge accumulated and transmitted across generations: construction methods, material choices, decorative patterns, proportions, finishing techniques and workshop-specific stylistic conventions.
The proposed system begins by converting this heritage into a structured digital archive.
For each craft object, artisans provide photographs from different perspectives alongside information such as dimensions, materials, accessories, construction techniques and historical references. A multimodal AI model then analyses the photographs to extract additional visible characteristics, including shape, colour, texture and decorative elements.
Crucially, the automated description does not replace artisan knowledge. Instead, it complements information that cannot be inferred from photographs alone, particularly specialised manufacturing processes and techniques.
The result is a digital representation of the workshop’s heritage combining visual information with expert knowledge.
Retrieving the right heritage for each new design
Once digitised, the archive becomes part of a Retrieval-Augmented Generation (RAG) system.
When an artisan requests a new design—for example, a blue leather duffle bag—the system does not immediately send that instruction to an image generator. It first searches the workshop archive for the heritage information most relevant to the request.
It may retrieve details about traditional construction techniques, typical dimensions, materials, decorative features or characteristic stylistic elements.
This information is then incorporated into an expanded prompt for the generative model.
Instead of simply asking AI to create a generic blue leather bag, the system can therefore specify characteristics derived directly from the artisan’s own archive: the type of leather, construction method, proportions, hardware and finishing details.
The researchers argue that this intermediate retrieval step helps reduce one of the major risks associated with generative AI in cultural and creative industries: stylistic homogenisation.
From text prompts and sketches to digital prototypes
The framework supports two complementary design workflows.
The first is text-to-image generation. An artisan describes a product idea in natural language, the RAG system enriches the request with relevant heritage information, and a diffusion model generates a visual prototype.
The second is image-to-image generation. Here, an existing product, an AI-generated concept or even a hand-drawn artisan sketch becomes the visual starting point. The artisan can then request modifications while asking the system to preserve specific structural or stylistic characteristics.
In one experiment, for example, a hand-drawn bag sketch was transformed into a realistic brown leather prototype with three-dimensional floral details, while retaining the proportions and silhouette of the original drawing.
This iterative approach is particularly important because it keeps the artisan at the centre of the workflow. AI proposes visual alternatives, but the craftsperson continues to modify the instructions, evaluate the results and determine the final design direction.
Encouraging results—and important limitations
The framework was evaluated using an archive containing 30 representative craft products, documented with artisan annotations and photographs.
Artisan representatives assessed generated designs according to their innovation, structural coherence and consistency with the reference collection or workshop style. Four company representatives participated in one evaluation, producing an average score of approximately 8.3 out of 10.
The participants considered the system particularly useful for early-stage ideation and for rapidly exploring variations before investing time and materials in physical prototypes.
The experiments also highlighted limitations. Even when the prompts contained information retrieved from the workshop archive, current generative models did not always reproduce highly specific construction details or distinctive heritage elements accurately.
The researchers therefore position the technology as a design-support system rather than an autonomous designer.
Faster prototyping with less material waste
The potential impact extends beyond creativity.
Traditional prototyping can require repeated manual work and physical materials for every design variation. Digital generation makes it possible to explore multiple possibilities before producing a physical sample, potentially reducing development time, material waste and production costs.
For small artisan workshops competing with industrial manufacturing, this could provide a way to benefit from advanced digital technology without sacrificing the very characteristics that make their products distinctive.
The system also incorporates access controls and data-governance mechanisms through the Snap4City platform, an important consideration because workshop archives can contain proprietary designs, technical knowledge and intellectual property.
AI as a tool for continuity, not replacement
Perhaps the most significant idea behind the project is that technological innovation and cultural preservation do not necessarily have to pull in opposite directions.
Rather than asking artisans to adapt their creative identity to generic AI models, the framework reverses the relationship: AI is supplied with the artisan’s own heritage before it generates anything.
The resulting model of human-AI collaboration offers an alternative vision for generative technology in cultural industries. The value of AI lies not in replacing tacit expertise accumulated over generations, but in making that expertise usable within new digital workflows.
Future research will explore the approach across multiple workshops and other craft sectors, including jewellery and fashion. A major challenge remains the digitisation of heritage archives: the richer and more carefully documented these collections become, the more effectively AI systems will be able to draw on them.
For traditional craftsmanship, the question may therefore be shifting from whether to adopt generative AI to how to make AI understand what must be preserved when innovation begins.
work performed in the context of project ELLIE: ELLIE: On the UsE of Internet of Senses for the CuLturaL HerItagE
