Remove AI Watermarks
Inspect AI provenance and clean content you generated or edited. Use the image, video, and text operations below. Images support PNG / JPEG / WebP / HEIC / AVIF.
Use the Video, Text, and Image tools tabs for the complete operation catalog. Invisible image regeneration needs NVIDIA CUDA; video reconstruction also supports CPU. Optional models download on first use.
Source, docs and issues: github.com/wiltodelta/remove-ai-watermarks (Apache-2.0).
Read every locally-readable provenance signal (C2PA issuer, EXIF/XMP generator, IPTC 'Made with AI', China TC260 AIGC label, xAI signature, SynthID metadata proxy, visible marks) into one verdict.
Estimate the likely AI generator family (OpenAI, Google, ByteDance, FLUX, Stable Diffusion, Meta, Midjourney, Ideogram, Adobe Firefly, xAI, or unknown) and model version from pixel statistics alone (wiltodelta/raiw-source-classify). Runs on CPU in pure NumPy without metadata and abstains as unknown on camera photos or unrecognized pipelines. Separate from identify and not SynthID watermark detection.
Detect and remove every registered visible AI-provenance mark in one pass by localize -> fill. Only the mark footprint is touched; the rest of the image is left alone. Sensitivity sets how hard a borderline mark is trusted: auto relaxes only on same-vendor evidence, while strict never relaxes. The fill here is the classical cv2 inpaint; MI-GAN / LaMa run locally.
Inspect or losslessly strip AI / provenance metadata (C2PA, EXIF, XMP, IPTC, PNG chunks, the China TC260 AIGC label). Stripping never touches the pixels: PNG/WebP stay pixel-identical and JPEG keeps its DCT scan bit-for-bit.
Work on content you generated or edited. Image regeneration requires NVIDIA CUDA; video profiles can use CPU, MPS, or CUDA. First use may download model weights. An unknown result does not prove that a file is human-made or free of watermarks.
Work on content you generated or edited. Image regeneration requires NVIDIA CUDA; video profiles can use CPU, MPS, or CUDA. First use may download model weights. An unknown result does not prove that a file is human-made or free of watermarks.
Work on content you generated or edited. Image regeneration requires NVIDIA CUDA; video profiles can use CPU, MPS, or CUDA. First use may download model weights. An unknown result does not prove that a file is human-made or free of watermarks.
Watermarking is a weak trust signal: a mark that is almost always present yet removable can make a cleaned forgery look more trustworthy. Half of this tool is identify, which reads provenance rather than removing it. The point is user autonomy over your own AI generations, not stripping a third party's paid asset. See the repo for the legal limits on removing AI labels.