LipForcing 14B — MLX prepared packs
Prepared BF16, fused INT8 (affine, group 64) and optional Cider W8A8 weights for Natively Engine. This is the 14B Wan2.1 causal student, not the 1.3B variant. Select only the precision you need through the matching engine catalog. release-manifest.json records exact checksums, pinned source identity and the owned BF16 → INT8 → Cider preparation chain. These are existing accepted artifacts; this upload introduces no new numerical conversion or quality claim.
The pack includes the fixed a person talking embedding, face mask and converted TAEW2.1 tiny decoder. The engine catalog obtains Wan2.1 VAE and Wav2Vec2 separately at pinned revisions. Tiny/full decoder selection changes quality and speed. Cider additionally needs a local native helper build and has narrower hardware acceptance than the portable INT8 pack.
Face detector weights are not included. Current LipForcing preprocessing requires InsightFace buffalo_l (det_10g.onnx and 2d106det.onnx). Upstream pretrained face models are restricted to noncommercial research unless separately licensed. Acquire them yourself under applicable terms and pass --insightface-root; the engine does not silently download them. The Apache license of the video model does not remove this separate pipeline restriction.
Original LipForcing checkpoint: Jinhyuk Jang and LipForcing contributors; its model card licenses the weights Apache-2.0 and describes Wan2.1, OmniAvatar-LS and Self-Forcing lineage. Preserve their attribution and LICENSES/. The converted TAEW2.1 decoder is MIT (Ollin Boer Bohan); the LatentSync mask is Apache-2.0. MLX conversion/packaging: Shraey, Natively. Natively Engine code is separately GPL-3.0-or-later.
For install/runtime/terms details follow the engine's docs/lipforcing.md. Installation and integrity checks do not guarantee output quality for every video or support on every memory tier. No portraits, voice recordings, model outputs or platform-specific binaries are included.
Quantized