jjihwan / Novel-View-Refinement

SV3D finetuning for NVR

Novel View Refinement

Setting up

PyTorch 2.0

conda activate sv3d python==3.10.14
pip3 install -r requirements.txt

Install deepspeed for training

pip3 install deepspeed

Get checkpoints

Store them as following structure:

cd Novel-View-Refinement
    .
    └── checkpoints
        └── sv3d_p.safetensors
        └── sv3d_u.safetensors # might not be used.

Training (WIP)

sh scripts/sv3d_finetune.sh

Inference

Store the input images in assets

sh scripts/inference.sh

Datasets

Use Objaverse 1.0 dataset with preprocessing pipeline.

See objaverse dataloader for detail.