Likeness uses converted versions of open-source research models. The converted models are bundled in the app and run on-device; they are never downloaded or replaced at runtime.
Face restoration is based on RestoreFormer++ by Zhouxia Wang, Jiawei Zhang, Tianshui Chen, Wenping Wang, and Ping Luo. The upstream repository and released checkpoint are provided under the Apache License 2.0.
For Core ML deployment, Zebel replaced the upstream one-hot codebook lookup with the mathematically equivalent embedding gather and specialized attention shapes for a single fixed 512×512 input. Each converted build is compared pixel-for-pixel with the untouched upstream PyTorch graph before release.
Likeness does not present the raw generative output directly. It transfers fine detail into structure retained from the source photograph and applies an on-device identity-deviation check.
Background and full-image upscaling are based on Real-ESRGAN x4plus by Xintao Wang and collaborators, provided under the BSD 3-Clause License.
The model is converted to a fixed 256×256 Core ML tile. Larger photographs are processed as non-overlapping output interiors with padded model inputs, keeping memory bounded without writing any output pixel twice.
The app includes the applicable Apache 2.0 and BSD 3-Clause licence and attribution text. Source links above identify the exact upstream projects. Likeness does not include CodeFormer, GFPGAN, or another model whose applicable terms do not clearly permit our commercial distribution.
Questions about model provenance: likeness [at] zebel.ai