nvlabs/cbottle

cBottle is an AI model from NVIDIA that generates realistic snapshots of the Earth's atmosphere at kilometer-scale resolution. Think of it as a system that can create detailed weather and climate data, things like temperature, pressure, and moisture levels, across the globe, or take a low-resolution weather picture and fill in the fine details to make it sharp and realistic. At its core, the project uses a "diffusion" approach, which is the same family of AI techniques behind image generators. Instead of creating photos, it creates atmospheric states, grids of weather data that look like they came from real observations or physics-based simulations. The project actually includes a few different models: a coarse global model, a video version that can roll weather forward through time in multi-step sequences, and a super-resolution model that takes blurry or low-detail climate data and sharpens it. NVIDIA provides pre-trained versions of these models through their Earth2Studio toolkit, so you can start generating data without training anything from scratch. The primary audience is climate and weather researchers, or teams building applications that need high-quality atmospheric data. For example, a research lab studying extreme weather might use this to generate detailed storm simulations. A company building weather-risk tools for agriculture or insurance could use it to produce fine-grained historical or scenario data. It could also help fill gaps where real observation data is sparse, say, downscaling a coarse global climate model to regional, kilometer-level detail. A few things stand out. The project is explicitly labeled research-and-development only, not a production tool. It's built to run on NVIDIA GPUs (the inference code references CUDA directly). Training these models is resource-intensive, the training commands suggest large batch sizes and thousands of steps. The project also bundles third-party software, so anyone using it should be mindful of those licenses. The README is sparse on conceptual explanation, but the linked documentation and example notebooks fill in practical guidance for getting started.