Instructions to use xocialize/hairdeck-mlx with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use xocialize/hairdeck-mlx with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] hf download xocialize/hairdeck-mlx --local-dir hairdeck-mlx
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
HairDeck weights for mlx-hairdeck-swift
The released checkpoints of HairDeck (hair cards as tokens: anime-style hair with
spring-bone chains from a point cloud or a picture), laid out for
mlx-hairdeck-swift, a Swift-MLX port. The tensors are the
upstream ones, unchanged (float32); only the file names differ, and the statistics the models need are gathered in
config.json.
| File | What | Upstream |
|---|---|---|
pointcloud_dit.safetensors |
point-cloud set DiT, 181.3M parameters | ckpt/pointcloud_model/model.safetensors, byte for byte |
image_dit.safetensors |
image set DiT, 182.3M | ckpt/image_model/model.safetensors, byte for byte |
color_dit.safetensors |
per-card colour model, 183.5M | ckpt/color_model/model.safetensors, byte for byte |
chain_dit.safetensors |
spring-bone chain DiT, 30.5M | ckpt/chain_model/model.safetensors, byte for byte |
card_vae.safetensors |
card VAE, 8.1M | ckpt/card_vae/model.safetensors, byte for byte |
dora_encoder.safetensors |
the encoder tensors of Dora-VAE 1.1 (the point-cloud condition) | Seed3D/Dora-VAE-1.1 @ 7187b01, dora_vae_1_1.ckpt, the shape_model.encoder.* tensors |
config.json |
run arguments of every model, token / latent / card / chain statistics, the canonical head | HairDeck's data/ and ckpt/ json files |
SHA256SUMS |
checksums of the files above |
The image path also needs DINOv3 ViT-L/16. It is not in this repository: the port reads dino.safetensors from
xocialize/trellis2-mlx, the same file byte for byte as
facebook/dinov3-vitl16-pretrain-lvd1689m's model.safetensors (sha256 dcb2e451โฆ8179), under the DINOv3 License
("Built with DINOv3").
Licences
HairDeck's code and checkpoints are Apache-2.0 (LICENSE). Dora-VAE 1.1 is Apache-2.0. DINOv3 is under Meta's DINOv3
License and is distributed separately as described above.
The checkpoints were trained by HairDeck's authors on VRoid Hub characters; see the upstream repository for the data and its terms. This repository adds no training.
- Downloads last month
- 12
Quantized
Model tree for xocialize/hairdeck-mlx
Base model
hyz317/HairDeck