This is a quickly produced unofficial reproduction of Fantasia3D for text-to-3D generation, featuring the ability to generate high-quality meshes with PBR textures directly.

Thanks for the author's kind help in elaborating the details! This repo is also heavily based on Nvdiffrec (with extremely untidy modifications...).

Important Notice: This repo is like a preview of this impressive method, and may not be actively maintained since the authors will release the official implementation. The performance seems not as good as the original paper, and there are still some details may be different, but it's still a start point if you are interested.

pip install -r requirements.txt

First you need to create a config file under configs, you can copy and edit one of the provided examples like this:

{
    "text": "a ripe strawberry",
    "random_textures": true,
    "iter": 5000,
    "save_interval": 10,
    "texture_res": [ 1024, 1024 ],
    "train_res": [512, 512],
    "batch": 8,
    "learning_rate": 0.001,
    "ks_min" : [0, 0.25, 0],
    "out_dir": "strawberry",
    "base_mesh": "./data/sphere.obj"
}

We provide two base meshes under ./data (sphere.obj and ellipsoid.obj), but you can use any mesh to initialize the geometry.

The default settings are tested under a 32GB V100. Lower batch and increase iter if your GPU memory is limited.

Then you can run training by:

# single GPU
python train.py --config configs/strawberry.json

# multi GPU (experimental)
torchrun --nproc_per_node 4 train.py --config configs/strawberry.json

For single GPU (V100), it takes about 4 hours to train a single model (5000 iters at batch size of 8).

The validation/checkpoints/final mesh will be stored to ./out/<out_dir>

• the major logic of SDS is hard coded in geometry/dmtet.py.
• It seems important to apply antialiasing on the normal and visibility mask (render/render.py) so the gradient can be propagated to DMTet.

• The awesome original paper:

@misc{chen2023fantasia3d,
      title={Fantasia3D: Disentangling Geometry and Appearance for High-quality Text-to-3D Content Creation}, 
      author={Rui Chen and Yongwei Chen and Ningxin Jiao and Kui Jia},
      year={2023},
      eprint={2303.13873},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

• Nvdiffrec codebase.