MVHOI Weights Bundle
This folder contains the model files needed by the public MVHOI release.
This repository is intended as a private archival full-weight package. It contains MVHOI checkpoints/adapters together with third-party base models needed for local reproduction. Check the licenses and access terms of MANO, OmniHands, InstantMesh, Zero123Plus, and Stable Diffusion before redistributing it publicly.
Size
total: ~16G
omnihands: 5.2G
instantmesh: 8.3G
sd_inpaint: 2.0G
sd_lora_dual: 49M
sd_lora_hand: 49M
mano_assets: 3.7M
hamer_data: 3.7M
Layout
MVHOI_weights/
├── omnihands/
│ ├── Demo_Multiview.pth
│ ├── config_multi.yaml
│ └── epoch58-mpvpe1.63.ckpt
├── mano_assets/
│ └── models/MANO_RIGHT.pkl
├── hamer_data/
│ ├── mano/MANO_RIGHT.pkl
│ └── mano_mean_params.npz
├── sd_inpaint/
├── sd_lora_dual/lora/
│ ├── lora_epoch_009/
│ │ ├── adapter_config.json
│ │ ├── adapter_model.safetensors
│ │ └── conv_in.pt
│ └── lora_epoch_019/
│ ├── adapter_config.json
│ ├── adapter_model.safetensors
│ └── conv_in.pt
├── sd_lora_hand/lora/lora_epoch_019/
│ ├── adapter_config.json
│ ├── adapter_model.safetensors
│ └── conv_in.pt
└── instantmesh/
├── instant_mesh_large.ckpt
├── diffusion_pytorch_model.bin
└── zero123plus-v1.2/
Environment
Set these before running the release code:
export MVHOI_WEIGHTS_ROOT=/path/to/MVHOI_weights
export MVHOI_OMNI_BACKBONE=$MVHOI_WEIGHTS_ROOT/omnihands/Demo_Multiview.pth
export MVHOI_OMNI_CONFIG=$MVHOI_WEIGHTS_ROOT/omnihands/config_multi.yaml
export MVHOI_MANO_ROOT=$MVHOI_WEIGHTS_ROOT/mano_assets
export MVHOI_HAMER_DATA_ROOT=$MVHOI_WEIGHTS_ROOT/hamer_data
export MVHOI_SD_INPAINT_DIR=$MVHOI_WEIGHTS_ROOT/sd_inpaint
export MVHOI_SD_LORA_ROOT=$MVHOI_WEIGHTS_ROOT/sd_lora_dual/lora
export MVHOI_INSTANTMESH_CKPT=$MVHOI_WEIGHTS_ROOT/instantmesh/instant_mesh_large.ckpt
export MVHOI_ZERO123PLUS_UNET=$MVHOI_WEIGHTS_ROOT/instantmesh/diffusion_pytorch_model.bin
export MVHOI_ZERO123PLUS_MODEL=$MVHOI_WEIGHTS_ROOT/instantmesh/zero123plus-v1.2
The main optimizer consumes the OmniHands, MANO, and hamer_data files. The SD and InstantMesh files are needed when running the upstream image-completion and object-reconstruction stages.
The default dual-task SD completion command uses sd_lora_dual/lora/lora_epoch_009
with --epoch 9. The hand-only sd_lora_hand/lora/lora_epoch_019 adapter is an
optional backup and is not loaded by the default dual-task inference command.