ferdinandl007/VibeKit-FLARE-Fast

VibeKit FLARE Fast

An Apple Silicon inference export of FLARE Large for English visual speech recognition. VibeKit compiles the FP16 Core ML encoder on the Mac and runs the FP16 decoder through Swift MLX with a KV cache. The export uses only the mouth input. The weights come from the original model without additional training.

The source is the Large fine-tuned checkpoint from the authors of Learning Visual Speech Representations via Cross-Modal Distillation and Joint Face-Lip Modeling (Jing-Xuan Zhang et al.). See the original implementation. We converted the published checkpoint to an FP16 Core ML encoder and FP16 safetensors decoder for VibeKit.

The model is available under Creative Commons Attribution-NonCommercial 4.0 International. Preserve the upstream attribution and the description of these changes when sharing it.