A command-line tool and Python library on GitHub removes visible sparkle overlays from Google Gemini outputs and invisible watermarks such as SynthID v1/v2, StableSignature and TreeRing from images produced by multiple generators. The repository at https://github.com/wiltodelta/remove-ai-watermarks documents reverse alpha blending for Gemini logos, diffusion-based regeneration at SDXL 1024 px resolution for frequency-domain marks, and direct stripping of EXIF, PNG chunks and C2PA manifests. Three-stage NCC detection locates watermarks after resize or crop operations. The library supports batch processing and includes optional face protection via YOLO extraction before diffusion passes. Related technical references include Google’s 2023 SynthID paper and the C2PA technical specification v2.1 released in 2024. The diffusion regeneration step resizes to native SDXL dimensions, adds controlled noise over approximately 50 steps at strength 0.05, then decodes and upsamples, a pipeline verified on Gemini 3 Pro outputs after May 2026. Earlier SD-1.5 768 px paths were dropped once testing showed insufficient removal of SynthID v2 patterns.