You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Removed --restore-faces and --restore-faces-method CLI flags. Every face-restore approach we evaluated (GFPGAN-on-cleaned, PhotoMaker-V2, InstantID txt2img, InstantID img2img-on-cleaned at three parameter sweeps, full record in docs/synthid-robust-identity-research-2026-06-08.md) regenerated the face via SDXL diffusion — the output face inherited SDXL's "clean skin" aesthetic and lost original identity precision, looking more AI-generated than the cleaned image. The cleaned controlnet 0.20 image is the least-AI face state achievable without re-introducing SynthID, so that's now the prod output.
Dropped restore, photomaker, instantid extras (and the [tool.hatch.metadata] allow-direct-references block that was only there for the photomaker git URL).
Dropped restore_faces / restore_faces_method parameters from InvisibleEngine.remove_watermark and AutoConfig.
License
MIT → Apache 2.0. Apache adds an explicit patent grant and the trademark-use clause, and matches the licenses of the upstream projects this library mirrors / depends on (diffusers, transformers, controlnet-aux, xinsir's canny ControlNet weights).
Notes
--pipeline controlnet and the canny ControlNet path are unchanged. Canny preserves face structure, not identity — the regenerated face drifts in likeness. That drift is now the documented prod behavior.
Vendored MIT attribution for mertizci/noai-watermark is preserved in noai/__init__.py as required by the original MIT license; Apache 2.0 and MIT are compatible.