In the northeast of the United States, the Gulf of Maine represents one of the most diversified marine ecosystems on the planet-which houses whales, sharks, jellyfish, herring, plankton and hundreds of other species. But even if this ecosystem supports rich biodiversity, it undergoes a rapid environmental change. The Gulf of Maine warms more quickly than 99% of the world's oceans, with consequences that still take place.
A new research initiative developing at MIT Sea Grant, called Lobstger – Abbreviation of learning oceanic bioecological systems through generative representations – brings together artificial intelligence and underwater photography to document the oceanic life that has left vulnerable to these changes and share them with the public of new visual ways. Co-directed by an underwater photographer and invited artist at MIT Sea Grant Keith Ellenbogen and the MIT mechanical engineering of MIT Andreas Mentzelopoulos, the project explores how the generator can extend the scientific narration by relying on photographic data on the ground.
Just as the 19th century camera has transformed our ability to document and reveal the natural world – capture life with unprecedented details and put in sight of distant or hidden environments – the generative AI marks a new border in visual narration. Like early photography, AI opens a creative and conceptual space, contesting the way we define authenticity and how we communicate scientific and artistic perspectives.
In the Lobstger project, the generative models are formed exclusively on an organized library of original underwater photographs of Ellenbogen – each image manufactured with an artistic intention, technical precision, precise identification of species and a clear geographic context. By building a high -quality data set anchored in real world observations, the project guarantees that resulting imagery maintains both visual integrity and ecological relevance. In addition, Lobstger models are built using the personalized code developed by Mentzellopoulos to protect the process and outputs from all potential biases from external data or models. The generative AI of Lobstger relies on real photography, expanding the visual vocabulary of researchers to deepen the public link with the natural world.
This Ocean Sunfish image (Mola Mola) was generated by unconditional Lobstger models.
Image generated by AI: Keith Ellenbogen, Andreas Mentzelopoulos and Lobstger.
In its heart, Lobstger operates at the intersection of art, science and technology. The project draws from the visual language of photography, the rigor of observation of marine science and the computing power of the generative AI. By uniting these disciplines, the team does not only develop new ways of visualizing ocean life – they also reinvent how environmental stories can be told. This integrative approach makes Lobstger both a research tool and a creative experience – which reflects the long -standing tradition of interdisciplinary MIT innovation.
The underwater photography in the coastal waters of New England is notoriously difficult. Limited visibility, swirling sediments, bubbles and the unpredictable movement of marine life all pose constant challenges. In recent years, Ellenbogen has led to these challenges and built a complete recording of the region's biodiversity through the project, Space to Sea: Visualization of the New England Wilderness. This large set of underwater image data provides the basics of the formation of Lobstger generative AI models. The images cover various angles, lighting conditions and animal behavior, resulting in a visual archive which is both striking artistic and biologically precise.
Image synthesis via reverse broadcast: this short video shows the latent gaussian noise shift trajectory at the photorealistic output using the unconditional Lobstger models. The iterative lag requires that 1,000 passes forward in the neural network formed.
Video: Keith Ellenbogen and Andreas Mentzellopoulos / Mit Sea Grant
Lobstger's personalized broadcasting models are formed to reproduce not only the Ellenbogen biodiversity documents, but also the artistic style it uses to capture it. By learning thousands of real underwater images, models internalize fine grain details such as natural lighting gradients, specific coloring for species and even atmospheric texture created by suspended particles and refracted sunlight. The result is the imagery which not only seems visually precise, but also feels immersive and moving.
The models can both generate new, synthetic, but scientifically precise, unconventional images (that is to say, requiring any user input / guidance), and improve the real conditionally (i.e. image generation in the image). By integrating the AI into the photographic workflow, Ellenbogen will be able to use these tools to recover the details in troubled water, adjust the lighting to underline the key subjects, or even simulate scenes which would be almost impossible to capture on the ground. The team also considers that this approach can benefit other underwater photographers and image editors facing similar challenges. This hybrid method is designed to speed up the conservation process and allow storytellers to build a more complete and coherent visual story of life below the surface.
Left: improved image of an American lobster using image models with Lobstger image. Right: original image.
Left: genotated image of Ai by Keith Ellenbogen, Andreas Mentzelopoulos and Lobstger. Right: Keith Ellenbogen
In a series of keys, Ellenbogen has captured high resolution images of the lion's mane, blue sharks, American lobsters and oceanic sun (Mola mola) By diving free in coastal waters. “It is not easy to get a high quality data set,” says Ellenbogen. “This requires several dives, missed opportunities and unpredictable conditions. But these challenges are part of what makes underwater documentation that are both difficult and enriching. ”
Mentzellopoulos has developed original code to form a family of latent diffusion models for Lobstger based on Ellenbogen images. The development of these models requires a high level of technical expertise, and training models from zero are a complex process requiring hundreds of calculation hours and meticulous hyperparameter adjustment.
The project reflects a parallel process: documentation in the field through photography and development of models thanks to iterative training. Ellenbogen works in the field, capturing rare and ephemeral meetings with marine animals; Mentzellopoulos works in the laboratory, translating these moments in automatic learning contexts which can extend and reinterpret the visual language of the ocean.
“The objective is not to replace photography,” says Mentzellopoulos. “It is to rely on and complete it – make the invisible visible, and help people see environmental complexity in a way that resonates both emotionally and intellectually. Our models aim to capture not only biological realism, but the emotional charge which can lead to the commitment and action of the real world. ”
Lobstger points to a hybrid future which merges direct observation with a technological interpretation. The long -term objective of the team is to develop a complete model which can visualize a wide range of species found in the Gulf of Maine and, possibly, apply methods similar to marine ecosystems around the world.
The researchers suggest that the photography and the generators form a continuum rather than a conflict. Photography captures what is – texture, light and animal behavior during real encounters – while AI extends this vision beyond what is seen, towards what could be understood, deduced or imagined on the basis of scientific data and artistic vision. Together, they offer a powerful framework to communicate science by creating images.
In a region where ecosystems change quickly, the act of visualization becomes more than just documentation. It becomes a tool for awareness, commitment and, ultimately, conservation. Lobstger is still in its infancy, and the team is looking forward to sharing more discoveries, images and information as the project evolves.
Left image response: The left image has been generated using unconditional Lobstger models and the right image is real.
For more information, contact Keith Ellenbogen And Andreas Mentzelopoulos.
