Design cannot be disconnected from the last—lines, proportions, and viability depend on it.
Project develops AI models that generate shoe sketches constrained by actual last geometry, moving beyond visual inspiration to technically feasible designs. Custom dataset of 37,000 shoe sketches trained the AI, addressing lack of specialized public databases for industrial footwear design applications.
- 37,000 shoe sketches in custom dataset trained the AI models
- Project funded by Valencian Institute for Business Competitiveness (IVACE+i)
- Developed by ITI (technology research) and Inescop (footwear technology center)
- Participating companies: UNISA, Pedro García, Pikolinos, Dian
AIGEN4FASHION project integrates generative AI into shoe design by creating sketches adapted to real shoe lasts, bridging creative vision with manufacturing constraints through specialized training datasets.
In the design studios of Valencia's shoe companies, a familiar tension has long defined the work: a designer sketches something beautiful, technically precise, and then the sketch must be translated—often clumsily—into the actual geometry of a shoe last, the wooden or plastic form around which leather and canvas are shaped. That translation costs time. It requires back-and-forth between creative and technical teams. It produces waste in the form of abandoned ideas that look good on paper but won't work in three dimensions.
AIGEN4FASHION, a project funded by the Valencian Institute for Business Competitiveness and backed by two research centers—ITI, which specializes in information technology, and Inescop, the Footwear Technology Center—is designed to collapse that distance. Rather than treating generative AI as a tool for producing visually appealing reference images, the project integrates the technology directly into the technical design process itself. The goal is to generate shoe sketches that are not merely attractive but geometrically adapted to a real last from the moment the designer begins working.
To accomplish this, the team built generative AI models trained to work within the constraints of actual last geometry. This required solving a fundamental problem: no public database of shoe sketches existed with enough specialization and scale to train an industrial-grade system. The researchers constructed their own, assembling roughly 37,000 examples organized by shoe type and design characteristics. The dataset was built through a semi-automated process that combined master models, coherent design lines from an industrial perspective, and various reference lasts. This specialized foundation allows the AI to operate at a level of precision that general-purpose tools cannot match.
François Signol, the project's principal investigator at ITI, described the shift in approach: generative AI can deliver real value to shoe design when trained on sector-specific data and integrated into actual company workflows. The technology is not about generating attractive images. It is about moving toward sketches adapted to technical constraints—the geometry of the last—so that creativity and manufacturability are connected from the earliest phases of design.
The companies involved in the project—UNISA, Pedro García, Pikolinos, and Dian among them—have articulated what this means in practice. Pablo Jaspers, a designer at UNISA, noted that many current tools produce visually appealing images that are difficult to manufacture or misaligned with real shoe proportions. AIGEN4FASHION's focus on the footwear sector and its commitment to adapting sketches to actual lasts and shoe types addresses that gap directly. Bernardo García, a designer at Pedro García, emphasized the value of being able to sketch directly onto a last: entering a text prompt and immediately seeing a model sketched onto the company's own last represents a significant advance in translating creative vision into technical language.
Oscar Franco, a last specialist at Pikolinos, sees potential for greater agility in design and better alignment with technical requirements. One of the most valuable improvements, he suggested, would be the ability to visualize sketches quickly on the last and to create curves on it without manual digitization. Juan José Hinojo, a CAD systems specialist and principal investigator at Inescop, framed the core benefit: in footwear, design cannot be disconnected from the last. Lines, proportions, and model viability all depend on it. Generating proposals adapted to real lasts can reduce iterations and improve communication between creative and technical designers.
The system also has the potential to make CAD tools more accessible to designers who find them complex or overly technical. The AI acts as a bridge, placing design lines on the last coherently while serving as both a technical tool for creative designers and a source of inspiration for technical ones. From a business perspective, the value extends beyond generating designs: it frees time for higher-value work. Francisco Bernabé, from Dian's graphic design department, noted that having a large database and predefined structures can accelerate the placement of baseline elements, allowing teams to invest more time directly in design itself.
The project has validated two generative AI technologies based on diffusion models and advanced toward an integrated design system embedded in a CAD environment where users can request sketches from the AI and refine them afterward. Looking forward, the teams envision 3D volume generation for virtual prototyping, direct CAD integration, material and comfort simulation, design adaptation based on manufacturing systems, and customer personalization tools. Bernardo García sees the system evolving toward functions increasingly connected to technical development: improved sketch placement on the last, functional line generation, part differentiation, texture application, and combination creation with minimal user interaction. This vision reinforces the idea that generative AI can become a cross-functional tool within the design workflow, not merely a source of visual inspiration. The real measure of success will be whether the technology makes the shoe industry faster, more precise, and more competitive—and whether it remains accessible to the small and medium-sized companies that define Valencia's footwear sector.
Citas Notables
Generative AI can deliver real value to shoe design when trained on sector-specific data and integrated into actual company workflows—not just generating attractive images, but moving toward sketches adapted to technical constraints.— François Signol, principal investigator at ITI
In footwear, design cannot be disconnected from the last. Lines, proportions, and model viability all depend on it. Generating proposals adapted to real lasts can reduce iterations and improve communication between creative and technical designers.— Juan José Hinojo, CAD systems specialist and principal investigator at Inescop
La Conversación del Hearth Otra perspectiva de la historia
Why does it matter that the AI understands the geometry of the last? Couldn't a designer just use any AI tool and then adapt the sketch manually?
Because that manual adaptation is where time and money disappear. A designer sketches something, then a technical person has to redraw it to fit the actual last. Back and forth, revisions, wasted effort. If the AI understands the last from the start, you skip that entire loop.
So this is really about speed and efficiency, not creativity?
It's both. The designers I read about weren't saying the AI replaces their creativity—they were saying it removes friction. You can try more ideas faster. You can see immediately whether something will actually work on the last. That's liberating, not limiting.
The dataset of 37,000 sketches—why was that so hard to build?
Because no one had done it before. Public databases either had too few examples, too many generic categories, or weren't organized in ways that made sense for actual shoe manufacturing. The team had to build it themselves, which meant understanding the industry deeply enough to know what to collect and how to organize it.
What happens next? Is this tool ready for factories to use?
Not yet. They've proven the concept works, but the companies involved are already talking about what comes next—3D shapes, material simulation, brand-specific design styles. The real work is making it so integrated into existing workflows that it feels natural, not like an add-on.
Does this threaten shoe designers' jobs?
The designers quoted in the reporting don't see it that way. They see it as a tool that lets them spend less time on technical grunt work and more time on actual design thinking. Whether that's how it plays out in practice—that's the question the industry will have to answer.