What is flowlet?

FlowLet is the official code release for a research paper on generating synthetic 3D brain MRI scans, accepted at the journal Medical Image Analysis. It is built with PyTorch and is aimed at researchers working in medical imaging rather than general developers. The core idea is that FlowLet generates realistic 3D brain scans that can be conditioned on a person's age, meaning the generated scan reflects age related changes in brain structure. It does this using a technique called flow matching, applied inside a wavelet transformed version of the data rather than on the raw image. Working in this transformed space lets the model generate images at multiple levels of detail without needing a separate compression step, and produces samples through a deterministic, step by step process rather than a random one. Age information is fed into the model in two ways, through a global conditioning mechanism and through an attention mechanism that lets the model focus on different spatial regions. The repository includes the full training and generation code, along with pretrained model checkpoints hosted on Hugging Face, though several of the largest checkpoint variants are still listed as in training at the time of this writing. Because of patient privacy rules, no actual MRI scan data is included in the repository. Instead, the README points to three publicly available research datasets that qualified researchers can request access to, along with a detailed preprocessing pipeline covering bias correction, spatial alignment, skull removal, and intensity normalization. Setup requires Python 3.11 or newer and a CUDA capable GPU, with training documented to run within 24 gigabytes of GPU memory. This is a specialized research tool intended for people already working in medical image analysis, not a general purpose image generator.