Developer weighs licences behind ML models, including training-data terms for a photo culler

A developer building krites, a photo-culling tool that runs locally, argues that choosing a machine-learning model depends first on licensing rather than accuracy. Each model can carry three separate licences: one for the code, one for the weights, and one for the training data, which can conflict. The author highlights two inpainting models, LaMa and MI-GAN, both permissively licensed but trained on Places2 data under non-commercial terms, leaving unresolved whether those data terms bind the weights. To manage this, the project gates object removal behind a configuration flag requiring explicit acceptance of the risk. The author also notes that a licence-less Hugging Face re-host of MI-GAN and a GPL-licensed mirror pose pinning hazards, and plans to build and host a properly licensed export.