Supervisor weights: AdaFace IR101 backbone

AdaFace IR101 (WebFace12M) backbone weights, mirrored from the AdaFace project so that H1/H2 training can fetch them from a stable location instead of a Google Drive link. Only the backbone (model.*) is kept: 261 MB instead of 1.53 GB, every tensor byte-identical to the original. Non-commercial research use only — see ADAFACE-NOTICE.md.

Installer

Download DeepFaceLab-installer.zip, extract it into an empty folder and run install.bat (Windows) or install.sh (Linux).

0.0.2

facelib weights for the two additional extraction engines: RetinaFace-R50 (detector) and PIPNet-68 (landmarker), plus the attribution notice for the upstream projects they were converted from. Defaults are unchanged: extraction still uses s3fd and fan-2d unless the new engines are chosen explicitly. Release 0.0.1 is left untouched, so existing installations keep verifying against the manifest they were installed with.

Installer assets 0.0.1

Assets downloaded by install.sh / install.bat. You do not need to download anything here by hand — the installer fetches, verifies (SHA-256) and extracts them for you.

asset	contents	required
facelib.tar.gz	network weights for detection, landmarks and segmentation	yes
ffmpeg.tar.gz	ffmpeg build used for video in/out	yes
model_generic_xseg.tar.gz	the generic XSeg model	yes
pretrain_Quick96.tar.gz	Quick96 pretraining data	yes
pretrain_faces.tar.gz.000/.001/.002	face pretraining set (1.8 GB, split into blocks)	no
EbSynth.tar.gz	EbSynth, Windows only	no

pretrain_faces is split across three blocks because it exceeds the per-part limit; the installer reassembles it and verifies the SHA-256 of the whole archive before extracting.