生成360度单批次图像：这是一种评估方式，简单说，就是让模型生成出一组图像，这组图像是从不同角度（全方位360度）观看同一3D物体的结果。单批次意味着这组图像是一次性，即在同一时间生成的。

创建平滑的对象渲染：这个是让模型生成一系列图像，这些图像展示的是当你沿着一个平滑轨迹（例如，环绕物体一周）移动摄像机时，物体会如何变化。这种方法的目的是检查模型生成的3D物体是否在所有视角下都很真实（也就是3D一致性）。

在单个输入图像的条件下生成图像：这是指给模型一个物体的单个图像（例如，一个苹果的侧面图），然后让模型生成该物体在其他视角下的图像。这个能力对于许多应用来说都很有用，像是虚拟现实、视频游戏或电影特效等，因为你可能只有物体的一两个视角的图像，但需要生成其他视角的图像。

NeRF（神经重建场）：是一种3D重建技术，用于从一组2D图像（拍摄自不同角度）重建出3D场景。NeRF被训练去学习一个函数，这个函数可以将3D空间中的任一点和相应的视角映射到颜色和不透明度。在生成的图像中训练NeRF，实际上是在从这些图像重建一个3D模型。

ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models

ViewDiff generates high-quality, multi-view consistent images of a real-world 3D object in authentic surroundings.

This is the official repository that contains source code for the CVPR 2024 paper ViewDiff.

[arXiv] [Project Page] [Video]

If you find ViewDiff useful for your work please cite:

@inproceedings{hoellein2024viewdiff,
    title={ViewDiff: 3D-Consistent Image Generation with Text-To-Image Models},
    author={H{\"o}llein, Lukas and Bo\v{z}i\v{c}, Alja\v{z} and M{\"u}ller, Norman and Novotny, David and Tseng, Hung-Yu and Richardt, Christian and Zollh{\"o}fer, Michael and Nie{\ss}ner, Matthias},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    year={2024}
}

Installation

Create a conda environment with all required dependencies:

conda create -n viewdiff python=3.10
conda activate viewdiff
pip install -r requirements.txt

Then install Pytorch3D by following the official instructions. For example, to install Pytorch3D on Linux (tested with Pytorch3D 0.7.4):

conda install -c fvcore -c iopath -c conda-forge fvcore iopath
pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"

Then manually update triton to the required version:

pip install --upgrade triton==2.1.0

Data Preparation

• Download CO3D categories that you would like to train on. Follow the official instructions here: https://github.com/facebookresearch/co3d. You should end up with a directory structure like this:

<co3d_root>
<co3d_root>/teddybear
<co3d_root>/hydrant
<co3d_root>/donut
<co3d_root>/apple
...

• Generate BLIP2 text captions from the images for each category:

export CO3DV2_DATASET_ROOT=<path/to/co3d>
python -m viewdiff.data.co3d.generate_blip2_captions --dataset-config.co3d_root <path/to/co3d> --output_file <path/to/co3d>/co3d_blip2_captions.json

• Generate the prior preservation (aka Dreambooth) dataset for each category:

export CO3DV2_DATASET_ROOT=<path/to/co3d>
python -m viewdiff.data.co3d.generate_co3d_dreambooth_data --prompt_file <path/to/co3d>/co3d_blip2_captions.json --output_path <path/to/co3d>/dreambooth_prior_preservation_dataset

• Recenter the poses of each object, such that the object lies within the unit cube:

export CO3DV2_DATASET_ROOT=<path/to/co3d>
python -m viewdiff.data.co3d.save_recentered_sequences --dataset-config.co3d_root <path/to/co3d>

Training

Execute the following script (requires 2x A100 80GB GPUs):

./viewdiff/scripts/train.sh <path/to/co3d> "stabilityai/stable-diffusion-2-1-base" outputs/train <category=teddybear>

If you only have a smaller GPU available and want to sanity check that everything is working, you can execute this script (e.g. on a RTX 3090 GPU):

./viewdiff/scripts/train_small.sh <path/to/co3d> "stabilityai/stable-diffusion-2-1-base" outputs/train <category=teddybear>

In our experiments, we train the model for 60K iterations.

Evaluation

First, export a trained model to a runnable checkpoint:

python -m viewdiff.convert_checkpoint_to_model --checkpoint-path <path/to/checkpoint-XXXXX>

Create 360 Degree Images In A Single Batch

Execute the following script:

./viewdiff/scripts/test/test_spherical_360_256x256.sh <path/to/co3d> <path/to/saved_model_from_checkpoint-XXXXX> outputs/single-batch-uncond-generation <num_images=10> <category=teddybear> <num_steps=50>

This creates num_images images of a single object in a single forward pass of the model (first row in the teaser image). In total num_steps objects will be created.

Create Smooth Rendering Around An Object (Elevation=60 Degrees)

Execute the following script:

./viewdiff/scripts/test/test_sliding_window_smooth_alternating_theta_60_360_256x256.sh <path/to/co3d> <path/to/saved_model_from_checkpoint-XXXXX> outputs/smooth-autoregressive-theta-60 <category=teddybear>

This creates a video rendering of an object in a spherical trajectory at 60 degrees elevation.

Create Smooth Rendering Around An Object (Elevation=30 Degrees)

Execute the following script:

./viewdiff/scripts/test/test_sliding_window_smooth_alternating_theta_30_360_256x256.sh <path/to/co3d> <path/to/saved_model_from_checkpoint-XXXXX> outputs/smooth-autoregressive-theta-30 <category=teddybear>

This creates a video rendering of an object in a spherical trajectory at 30 degrees elevation.

Condition Generation On Single Input Image

Execute the following script:

./viewdiff/scripts/test/eval_single_image_input.sh <path/to/co3d> <path/to/saved_model_from_checkpoint-XXXXX> outputs/single-image-eval <category=teddybear>

This renders novel views for an object of the test set given a single image input. We also save the quantitative metrics PSNR, SSIM, LPIPS in the output directory.

Optimize a NeRF

We provide an easy way to train a NeRF from our generated images. When creating a smooth rendering, we save a transforms.json file in the standard NeRF convention, that can be used to optimize a NeRF for the generated object. It can be used with standard NeRF frameworks like Instant-NGP or NeRFStudio.

LICENSE

The majority of this repository is licensed under CC-BY-NC, however portions of the project are available under separate license terms:

• diffusers is licensed under the Apache 2.0 license. We use the repository to extend the default U-Net architecture, by adapting the model definition found in the original library.