| license: apache-2.0 | |
| library_name: mlx | |
| tags: | |
| - image-to-video | |
| - audio-driven-video | |
| - ai2apps | |
| # FlashHead Pro MLX | |
| Self-contained checkpoint for AI2Apps FlashHead native MLX Model Worker. Runtime and model Python code are distributed separately. | |
| Includes Pro DiT, its VAE and Wav2Vec2 base. Original tensor layouts are converted to MLX on load; Pro Wan VAE was converted offline from a pure tensor state dictionary to safetensors. No Torch is required for inference. | |
| Fixed upstream revisions, exact runtime file list, sizes and SHA-256 values are in `ai2apps-checkpoint.json`. | |
| Sources: [SoulX-FlashHead](https://github.com/Soul-AILab/SoulX-FlashHead), [official weights](https://huggingface.co/Soul-AILab/SoulX-FlashHead-1_3B), [Wav2Vec2](https://huggingface.co/facebook/wav2vec2-base-960h). Upstream model cards declare Apache-2.0. Preserve LICENSE and NOTICE.md. | |
| 512x512, 25 FPS, four denoising steps. This is a local offline generation path; no real-time streaming claim. | |