YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
X-Dub — MLX (lip-sync / dub)
MLX-format weights for the X-Dub lip-sync / dubbing model, for Apple-silicon (MLX) inference. Pick one DiT tier (your Mac's RAM decides) + the shared decoder/detectors below.
DiT tiers (download one)
| Tier | File | Size | When |
|---|---|---|---|
| Quality | X-Dub_model_int8_blockonly.safetensors |
6.9 GB | best fidelity — 15-step, full VAE |
| Fast / Balanced (default) | X-Dub_model_turbo_s100_int8.safetensors |
6.9 GB | 4-step turbo (speed-LoRA merged in), full VAE |
| Low-memory | X-Dub_model_turbo_s100_int6.safetensors |
5.4 GB | ≤16 GB Macs — turbo int6 + tiny decoder + tiled VAE |
Shared (all tiers)
Wan2.2_VAE.safetensors(1.4 GB) — VAE decoderlighttaew2_2.safetensors(45 MB) — tiny TAEW2.2 continuity decoder (low-mem path)dwpose/yolox_l.onnx+dwpose/dw-ll_ucoco_384.onnx(350 MB) — face/pose detection (ONNX, CPU)
Attribution & license
- X-Dub — KlingAIResearch/X-Dub, Apache-2.0 (public release built on Wan-5B). VAE: Wan2.2 (Alibaba Wan). DWPose: the DWPose project. TAEW: TAESD-style tiny autoencoder.
- Derivative work: MLX conversion + quantization. Distributed under Apache-2.0.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support