Downloads · 30 days
0
hunyuan3d-mlx-paint-large
zimengxiong/hunyuan3d-mlx-paint-large
hunyuan3d-mlx-paint-large is a image-to-3d model from zimengxiong. Use it for the image-to-3d task on the model card, and read the license before you ship it in a product. It is set up for diffusers. The card lists the license as other.
Texture-paint weights for the large (PBR) paint slot of the Hunyuan3D-Swift / Modelr MLX pipeline.
Access
Public
Updated Jul 10, 2026
Parameters
2B
8.7 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors8.7 GB · 100%
At a glance
- Task
- Image-to-3D
- Library
- diffusers
- License
- other
- Access
- Public
- Created
- Jul 10, 2026
- Updated
- Jul 10, 2026
- SHA
- b56e8b86
Model card
From the Hugging Face model README
hunyuan3d-mlx-paint-large
Texture-paint weights for the large (PBR) paint slot of the Hunyuan3D-Swift / Modelr MLX pipeline.
This bundle is the official hunyuan3d-paintpbr-v2-1 PBR texture UNet (multi-attention + reference-attention + material-dual-attention + DINO conditioning + pose RoPE; outputs albedo and metallic-roughness) from Tencent's Hunyuan3D-2.1, together with the supporting components it needs at inference:
- the DINOv2-giant image encoder (facebook/dinov2-giant) used for image conditioning,
- the Hunyuan3D-2.0 VAE (the PBR pipeline reuses the 2.0 VAE; the upstream
paintpbr-v2-1VAE directory ships a config only), - the Real-ESRGAN ×4 upscaler for multiview super-resolution,
- the diffusion scheduler configuration.
The tensor data is repackaged unmodified; only the on-disk layout is fixed so the byte layout is a stable contract for the MLX loader. The bundle is self-contained: the VAE, DINOv2 encoder, and upscaler are included directly rather than cross-referenced from another repo. The upstream duplicate diffusion_pytorch_model.bin is intentionally omitted (the .safetensors is authoritative).
Contents
| Path | Size | Description |
|---|---|---|
unet/config.json | 911 B | PBR multiview UNet configuration |
unet/diffusion_pytorch_model.safetensors | 3.92 GB | PBR multiview UNet weights |
vae/config.json | 553 B | VAE configuration (Hunyuan3D-2.0 VAE) |
vae/diffusion_pytorch_model.safetensors | 167.3 MB | VAE weights (Hunyuan3D-2.0 VAE) |
dinov2/config.json | 548 B | DINOv2-giant configuration |
dinov2/model.safetensors | 4.55 GB | DINOv2-giant image-encoder weights |
dinov2/preprocessor_config.json | 436 B | DINOv2-giant image preprocessor configuration |
realesrgan/rrdbnet_mlx.safetensors | 66.9 MB | Real-ESRGAN RRDBNet ×4 upscaler (MLX-format safetensors) |
scheduler/scheduler_config.json | 387 B | Diffusion scheduler configuration |
Total: 8.70 GB (8,704,938,622 bytes).
Usage
Consumed by the hy3d CLI from the Hunyuan3D-Swift package. Paint operates on an existing shape mesh plus the reference image (point --weights at the downloaded repo directory):
# 1. generate a shape (see hunyuan3d-mlx-shape-* bundles)
hy3d shape penguin.png -o mesh.glb --weights ./hunyuan3d-mlx-shape-large
# 2. paint it with PBR materials
hy3d paint mesh.glb penguin.png -o painted.glb --weights ./hunyuan3d-mlx-paint-large
Provenance
- PBR multiview UNet + scheduler:
hunyuan3d-paintpbr-v2-1— Tencent Hunyuan3D-2.1. - VAE:
hunyuan3d-paint-v2-0VAE — Tencent Hunyuan3D-2 (reused by the PBR pipeline). - Image encoder: DINOv2-giant — facebook/dinov2-giant.
- Upscaler: Real-ESRGAN (RRDBNet ×4).
License
- Paint UNet, VAE, and scheduler: Tencent Hunyuan3D-2 Community License — see https://huggingface.co/tencent/Hunyuan3D-2 and the license text.
- DINOv2-giant (
dinov2/): Apache-2.0 — facebook/dinov2-giant. - Real-ESRGAN (
realesrgan/rrdbnet_mlx.safetensors): BSD 3-Clause License (upstream xinntao/Real-ESRGAN).
Your use of each component is subject to its respective license.
- Task
- Image-to-3D
- Library
- diffusers
- License
- other
- Created
- Jul 10, 2026
- Updated
- Jul 10, 2026