Decoding nature: Rebuilding coral reefs with Wildflow

Wildflow’s 3D reconstruction pipeline is designed to reliably process terabytes of data without sacrificing visual fidelity or spatial accuracy. While chunking and tiling steps for 3D training are GPU-intensive, other stages depend on CPU processing, like camera positioning estimations.

Purpose-built to power varying workloads, Nebius Serverless AI lets Wildflow run each job with its own container image without worrying about cluster setup or overprovisioning costs. With Serverless Jobs, Wildflow matches GPU capacity to their AI lifecycle, seamlessly scaling from a handful of jobs to over 20 running in parallel, with agents helping automate the workflow. “We would have saved a lot of time and resources had we used Nebius from day one”, Nozdrenkov said.

Backed by Nebius' high-performance infrastructure, Wildflow optimized every stage of the reconstruction pipeline to accelerate 3D generation and prepare for AI model development.

Delivering photorealistic visuals for web-based exploration from any angle, on any device requires real-time rendering. With a custom 3D Gaussian Splatting (3DGS) architecture, Wildflow recreates complex environments at much lower computational costs than alternative computer vision methods.

By distributing millions of small, programmable ellipses as dense point clouds in a virtual space, this cutting-edge technique encodes color, defines opacity and reconstructs surfaces much faster for a browser-friendly experience.

With a 3DGS pipeline, Wildflow relies on Nebius AI Cloud to piece together terabytes of environmental data and raw footage from GoPro cameras, mobile phones and even robots — like autonomous or remotely operated vehicles — into topographically aware, precisely georeferenced digital twins.

To replicate underwater scenes all the way to the polyp, Wildflow must move TB-scale datasets swiftly through preprocessing, alignment, reconstruction and rendering stages. Any lags can result in IO bottlenecks, inconsistent rendering and quality loss.

Designed to scale AI workloads, Nebius’ shared filesystems let Wildflow retain full-resolution imagery across the pipeline, delivering the fine ecological details needed for reliable environmental monitoring. High-speed storage also prevents costly delays and lets the startup fully leverage high-throughput infrastructure.

Reef landscapes must be split into smaller scenes to fit GPU memory, but what’s the optimal number of chunks to maximize returns on infrastructure? With too many patches, a lot of compute is spent processing little data, but oversizing tiles risks compromising visual quality.

Wildflow solved it with an algorithm that dynamically partitions each site at the lowest number of patches that yields the best possible information density — ensuring efficient rendering and training tasks on Nebius AI Cloud.

Wildflow is on track to digitally recreate models from dozens of coral reef sites in partnership with over 20 research institutions around the world. In five years, hundreds of survey plots have been collected by academic partners including University College London (UCL), Derby University, the Lancaster University’s LEC Reefs research group, IPB University in Indonesia and the Australian Institute of Marine Science (AIMS).

Around 625 square meters in size, each survey plot is roughly half the size of an Olympic swimming pool and can be fully documented with 8,000pictures from two GoPro cameras. To design the model, Wildflow estimates the reef’s 3D layout from 2D images captured from various angles with Structure from Motion (SfM) techniques — recreating the basis of underwater scenes in under a day with Nebius’ high-throughput infrastructure. To render the details while preserving spatial accuracy, scale and color fidelity with sub-centimeter resolution, the 3D Gaussian Splatting step takes just a few hours running on Nebius AI Cloud.

“Nebius lets us quickly process large 3D reef datasets without needing a full cluster setup, which makes high-quality monitoring accessible even with limited resources — helping scientists and restoration teams scale their work faster and more affordably”, Nozdrenkov emphasized.

In an experience akin to Google Street View, you can explore a restoration site in Indonesia’s Spermonde Archipelago by navigating Wildflow’s web-based prototype, including close-up views for the highest level of detail. This other demo lets scientists and enthusiast dive into a nook of Australia’s colorful reefs — you can even count the polyp branchlets and lobes by zooming in.

Building on established field surveys, the digital twins unlock multimodal AI analysis to enhance long-term restoration planning and reef health monitoring. Wildflow’s 3D pipeline turns diver-collected footage into spatially consistent, full-resolution records at a fraction of the traditional processing time.

Combining deep expertise in ecology, machine learning and remote sensing, Wildflow’s 1-year vision is to train the first multimodal foundation models on diverse reef ecosystems globally. Their 3D reconstruction pipeline generates structured outputs, including camera poses, segmentation overlays and splats, all ready for foundation model training.

To unlock a more comprehensive understanding of the ecosystem, the training pipeline will be complemented with genomics, acoustics, satellite imagery and even hyperspectral imaging, alongside environmentally relevant indicators like temperature, salinity and oxygen levels.

The foundation model in development will help predict how localized interventions can strengthen reef communities by combining innovative AI capabilities:

• Delivering insights from sensor data: By automating species identification, area coverage assessments and disease detection with AI, restoration projects can define more precise risk mitigation strategies.

• Modeling ecosystem dynamics: An in-depth understanding of how the colorful polyps, booming biodiversity and dynamic lifecycles make up a thriving reef community is essential to accelerate conservation initiatives. Wildflow aims to model natural processes like population growth and predator-prey interactions for scientists to plan more effective, targeted environmental interventions.

• Measuring what nature gives us: Valuating specific ecosystem services is essential to help restoration teams demonstrate the economic returns only made possible by healthy reefs — a delicate ecosystem which harbors 25% of marine life supports the livelihoods of hundreds of millions of people, according to the Global Coral Reef Monitoring Network’s latest estimates.

• Tracking what we give back to nature: A foundation model will make it easier to quantify our impact on coral reefs — both positive and negative. Wildflow plans to model the effects of conservation initiatives like marine protected areas or restoration projects, as well as the consequences of pollution, agriculture runoff and uncontrolled coastal development.

Starting from coral reefs, Wildflow’s vision is to recreate other oceanic habitats within four years and model all Earth’s biomes with AI. Their solution is biome-agnostic by design, as its multimodal data capture, Structure-from-Motion reconstruction, and 3D Gaussian Splatting rendering pipeline can be applied to other endangered biomes. Wildflow’s eight-year goal is to develop a “digital nervous system for our planet”, able to deliver informed insights at scale to help design more efficient conservation initiatives and spread successful restoration methods across all ecosystems.

Beyond Wildflow’s own plans, the company actively contributes to AI model development worldwide. They shared over 350 GB of georeferenced GoPro footage and color-corrected imaging under an open-source license on Hugging Face to encourage people to build their own AI applications for coral reefs.

In the coming decades, Wildflow sees little limit to what humanity can achieve with compounding advances in AI-enhanced ecosystem modeling. The ability to accurately predict how targeted interventions will holistically affect complex biodiversity dynamics could unlock restoration goals beyond what we thought was possible. “Our goal is to be the first generation to leave nature better than we found it. By establishing a symbiotic relationship with our biosphere today, maybe one day we’ll even use that deep understanding to create a jungle out of a desert or terraform Mars, ” concludes Wildflow’s 16-year AI roadmap.