SPAD : Spatially Aware Multiview Diffusers

Yash Kant, Ziyi Wu, Michael Vasilkovsky, Gordon Qian, Jian Ren, Riza Alp Guler, Bernard Ghanem, Sergey Tulyakov*, Igor Gilitschenski*, Aliaksandr Siarohin*
Published at CVPR, 2024

---

Paper: https://arxiv.org/abs/2402.05235

Project Page: https://yashkant.github.io/spad/

Model pipeline. (a) We fine-tune a pre-trained text-to-image diffusion model on multi-view rendering of 3D objects. (b) Our model jointly denoises noisy multi-view images conditioned on text and relative camera poses. To enable cross-view interaction, we apply 3D self-attention by concatenating all views, and enforce epipolar constraints on the attention map. (c) We further add Plücker Embedding to the attention layers as positional encodings, to enhance camera control.

Citation

@misc{kant2024spad,
      title={SPAD : Spatially Aware Multiview Diffusers}, 
      author={Yash Kant and Ziyi Wu and Michael Vasilkovsky and Guocheng Qian and Jian Ren and Riza Alp Guler and Bernard Ghanem and Sergey Tulyakov and Igor Gilitschenski and Aliaksandr Siarohin},
      year={2024},
      eprint={2402.05235},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Filtered High-Quality Objaverse

If you are looking for the objaverse assets we used to train SPAD models, you can find that list here: filtered_objaverse.txt.

To see how this list was generated / tweak its parameters, you can try this colab notebook here: filter_objaverse.ipynb

See this issue for more discussion on data.

Repository Setup

Create a fresh conda environment, and install all dependencies.

conda create -n spad python=3.8 -y
conda activate spad

Clone the repository. Then, install pytorch (tested with CUDA 11.8), dependencies, pytorch3d, taming-transformers, and ldm:

git clone https://github.com/yashkant/spad
cd spad

# get pytorch (select correct CUDA version)
pip install --ignore-installed torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

# get dependencies
pip install -r requirements.txt

# get and install pytorch3d (from source)
pip install "git+https://github.com/facebookresearch/pytorch3d.git"

# get and install taming-transformers (from source)
git clone [email protected]:CompVis/taming-transformers.git
cd taming-transformers && pip install -e . && cd ..

# install ldm (from spad)
pip install -e .

Download Files

Download checkpoints and cache using following command:

python scripts/download.py

You can also download from browser / manually from huggingface hub: https://huggingface.co/yashkant/spad/tree/main/data

Pretrained Model

We provide two intermediate (slightly undertrained) checkpoints, with following specifications:

• spad_two_views: Trained jointly to denoise two views with learning rate 1e-4, relative cameras (between views) and no intrinsics,random viewpoints.
• spad_four_views: Trained jointly to denoise four views with learning rate 2e-5, absolute cameras (between views) with intrinsics, random + orthogonal viewpoints

You can test these models out using:

python scripts/inference.py --model spad_two_views --caption "Yellow Toyota Celica sports car."

Remove the caption argument to test on few evaluation captions provided at data/captions_eval.npy.

You can adjust the following hyperparameters for best results:

--cfg_scale: 3.0 to 9.0 (default 7.5)
--blob_sigma: 0.2 to 0.7 (default 0.5)
--ddim_steps: 50 to 1000 (default 100)

To understand the camera convention take a look at this issue.

To train on custom dataset take a look at this comment.

Visualizing and Creating Epipolar Masks

We provide code that allows you to visualize epipolar lines corresponding to each pixel between any two given views (see example below).

LHS above shows the source view (see highlighted pixel) from where the camera ray emerges, and RHS shows the target view and the epipolar line corresponding to the source pixel.

To visualize the epipolar masks and plucker embeddings (or use them as separate module), read and run the following script:

python scripts/visualize_epipolar_mask.py 

Important note: Some of the cameras of the uploaded sample visuals have zero translation, and are thus incorrect. Please ignore their outputs, I will attempt to clean these cameras from codebase very soon.