Open-Source Watermark Segmentation Model from Diffusion Dynamics: Powering clear.photo

Diffusion Dynamics has open-sourced the core technology behind its watermark removal product, clear.photo: a watermark segmentation model. This deep learning model generates masks highlighting watermark regions, excelling at segmenting logo-based watermarks. The project provides a complete workflow for training and inference, including dataset generation, model training, and post-processing, and supports fine-tuning on Apple M-series chips. A key feature is its data augmentation strategy which randomizes watermark parameters, leading to robust performance. This aims to provide a clear, easily modifiable baseline for building more complex tools.