Rodrigo Santos, a former student of Ciências ULisboa, develops a new approach to image editing using Artificial Intelligence

Artificial Intelligence is transforming the way we interact with images and information. It is at this intersection between language and image that the work of Rodrigo Santos is centred, a former bachelor’s, master’s and doctoral student at the Faculty of Sciences of the University of Lisbon and a researcher with the NLX group.

His doctoral research proposes a text‑guided image‑editing method that does not require task‑specific training, relying instead on existing Artificial Intelligence models. In this interview, he discusses the challenges and results of the work developed, the contribution of conducting research in Portuguese, and his path within NLX, as well as the transition from academic research to creating a company dedicated to transforming scientific knowledge into concrete applications.

Your research focuses on text‑guided image editing using artificial intelligence. How would you explain this work to an audience without training in the field?

My research focuses on teaching computers to edit images based on written instructions in natural language, for example, “put a hat on the cat”, without opening any editing software. Most systems that do this need to be trained specifically for the task, using thousands of manually annotated examples. The method I proposed in my doctoral work does not require any training at all. It takes advantage of existing artificial intelligence models that were trained for other tasks.

What motivated you to choose this research topic, and what were the main challenges throughout your PhD?

I completed all my academic training, bachelor’s, master’s and PhD, at the Faculty of Sciences of the University of Lisbon, always within NLX. My master’s degree focused on Portuguese–Chinese machine translation. Moving into multimodal AI, combining language and image, felt like a natural progression, also driven by the rapid evolution of this field. The biggest technical challenge was proving that it was possible to compete with much heavier supervised methods without training anything, relying solely on pre‑trained models used through inference. Another challenge was methodological. Benchmark datasets only contain one “correct” image for each instruction, when in reality there are several visually valid responses. This led us to rethink the evaluation metrics.

The CLARIN interview highlights your contribution to the development of language models and resources for Portuguese. What do you consider to be the impact of this work on science and on society?

Most advances in AI are made first in English. By building open models that are trained and evaluated specifically for Portuguese, we give other researchers and companies a solid foundation to work from, without having to replicate everything from scratch. These are the models that support, on a daily basis, applications such as chatbots, machine‑translation engines or educational tools in Portuguese. Without them, we remain dependent on technology designed for other languages.

How did your training and the scientific environment at Ciências contribute to the development of your research?

As I mentioned, my entire academic path took place at the Faculty of Sciences of the University of Lisbon. This gave me an unusual sense of continuity. I joined NLX during my master’s degree and grew scientifically within the same group, instead of starting from scratch at each stage. That continuity allowed me to consolidate technical skills early on and begin my PhD already with research experience, publications and a clear scientific framework. NLX is a group with a long‑standing tradition in natural language processing, coordinated by Professor António Branco. It is also the Portuguese national centre of the European research infrastructure PORTULAN CLARIN. This gave me access not only to close scientific guidance, but also to high‑quality linguistic corpora, models and computational infrastructure that I would hardly have had on my own.

What are the practical applications of the work you developed? In which areas might it have the greatest impact?

The most direct application is in image‑editing tools accessible to anyone, without technical training, with potential in areas such as design, marketing and accessibility. Because it does not rely on training or annotated data, the method adapts easily to different languages and domains, and tends to improve automatically as the underlying models evolve. Meanwhile, the Portuguese language models I worked on already have more immediate applications in areas such as machine translation, chatbots, customer support, content moderation and educational tools.

After completing your PhD, what are your next scientific challenges and goals?

The next step is to turn the knowledge gained over these years of research into something with concrete application. I am beginning the process of creating a company with that purpose, taking some of the ideas developed during the PhD beyond the academic context and transforming them into products that can be used by people and businesses. This transition from research to product development represents a shift in perspective. During the PhD, the main concern was understanding whether an idea was scientifically valid. Now, I want to understand whether an idea can be turned into something that has real‑world usefulness.

What advice would you give to students who want to pursue a research career in artificial intelligence?

Three things. First, get involved in a research group early on, ideally during your bachelor’s or master’s degree. Second, don’t always follow the dominant trend in the field, which is based on the idea of “more data, more parameters, more training”. There is scientific value in questioning that approach and looking for more efficient paths. That was, to some extent, the logic behind my own work. Third, stay curious and don’t become attached to the tools of the moment. What is state‑of‑the‑art today may be outdated in a few months. What endures is the ability to identify good problems and design rigorous experiments to test them.