Text2Room

Text2Room generates textured 3D meshes from a given text prompt using 2D text-to-image models.

This is the official repository that contains source code for the ICCV 2023 paper Text2Room.

[arXiv] [Project Page] [Video]

If you find Text2Room useful for your work please cite:

@preprint{hoellein2023text2room,
  title={Text2Room: Extracting Textured 3D Meshes from 2D Text-to-Image Models},
  author={H{\"o}llein, Lukas and Cao, Ang and Owens, Andrew and Johnson, Justin and Nie{\ss}ner, Matthias},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2023}
}

Prepare Environment

Create a conda environment:

conda create -n text2room python=3.9
conda activate text2room
pip install -r requirements.txt

Then install Pytorch3D by following the official instructions. For example, to install Pytorch3D on Linux (tested with PyTorch 1.13.1, CUDA 11.7, Pytorch3D 0.7.2):

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

Download the pretrained model weights for the fixed depth inpainting model, that we use:

• refer to the official IronDepth implemention to download the files normal_scannet.pt and irondepth_scannet.pt.
• place the files under text2room/checkpoints

(Optional) Download the pretrained model weights for the text-to-image model:

• git clone https://huggingface.co/stabilityai/stable-diffusion-2-inpainting
• git clone https://huggingface.co/stabilityai/stable-diffusion-2-1
• ln -s <path/to/stable-diffusion-2-inpainting> checkpoints
• ln -s <path/to/stable-diffusion-2-1> checkpoints

Generate a Scene

As default, we generate a living room scene:

python generate_scene.py

Outputs are stored in text2room/output.

Generated outputs

We generate the following outputs per generated scene:

Mesh Files:
    <output_root>/fused_mesh/after_generation.ply: generated mesh after the first stage of our method
    <output_root>/fused_mesh/fused_final.ply: generated mesh after the second stage of our method
    <output_root>/fused_mesh/x_poisson_meshlab_depth_y.ply: result of applying poisson surface reconstruction on mesh x with depth y
    <output_root>/fused_mesh/x_poisson_meshlab_depth_y_quadric_z.ply: result of applying poisson surface reconstruction on mesh x with depth y and then decimating the mesh to have at least z faces
    
Renderings:
    <output_root>/output_rendering/rendering_t.png: image from pose t, that was rendered from the final mesh
    <output_root>/output_rendering/rendering_noise_t.png: image from a slightly different/noised pose t, that was rendered from the final mesh
    <output_root>/output_depth/depth_t.png: depth from pose t, that was rendered from the final mesh
    <output_root>/output_depth/depth_noise_t.png: depth from a slightly different/noised pose t, that was rendered from the final mesh

Metadata:
    <output_root>/settings.json: all arguments used to generate the scene
    <output_root>/seen_poses.json: list of all poses in Pytorch3D convention used to render output_rendering (no noise)
    <output_root>/seen_poses_noise.json: list of all poses in Pytorch3D convention used to render output_rendering (with noise)
    <output_root>/transforms.json: a file in the standard NeRF convention (e.g. see NeRFStudio) that can be used to optimize a NeRF for the generated scene. It refers to the rendered images in output_rendering (no noise).

We also generate the following intermediate outputs during generation of the scene:

    <output_root>/fused_mesh/fused_until_frame_t.ply: generated mesh using the content until pose t
    <output_root>/rendered/rendered_t.png: image from pose t, that was rendered from mesh_t
    <output_root>/mask/mask_t.png: mask from pose t, that signals unobserved regions
    <output_root>/mask/mask_eroded_dilated_t.png: mask from pose t, after applying erosion/dilation
    <output_root>/rgb/rgb_t.png: image from pose t, that was inpainted with the text-to-image model
    <output_root>/depth/rendered_depth_t.png: depth from pose t, that was rendered from mesh_t
    <output_root>/depth/depth_t.png: depth from pose t, that was predicted/aligned from rgb_t and rendered_depth_t
    <output_root>/rgbd/rgbd_t.png: combination of rgb_t and depth_t placed next to each other

Create a scene from a fixed start-image

Already have an in-the-wild image, from which you want to start the generation? Specify it as --input_image_path and the generated scene kicks-off from there.

python generate_scene.py --input_image_path sample_data/0.png

Create a scene from another room type

Generate indoor-scenes of arbitrary rooms by specifying another --trajectory_file as input:

python generate_scene.py --trajectory_file model/trajectories/examples/bedroom.json

We provide a bunch of example rooms.

Customize Generation

We provide a highly configurable method. See opt.py for a complete list of the configuration options.

Get creative!

You can specify your own prompts and camera trajectories by simply creating your own trajectory.json file.

Trajectory Format

Each trajectory.json file should satisfy the following format:

[
  {
    "prompt": (str, optional) the prompt to use for this trajectory,
    "negative_prompt": (str, optional) the negative prompt to use for this trajectory,
    "n_images": (int, optional) how many images to render between start and end pose of this trajectory,
    "surface_normal_threshold": (float, optional) the surface_normal_threshold to use for this trajectory
    "fn_name": (str, required) the name of a trajectory_function as specified in model/trajectories/trajectory_util.py
    "fn_args": (dict, optional) {
      "a": value for an argument with name 'a' of fn_name,
      "b": value for an argument with name 'b' of fn_name,
    },
    "adaptive": (list, optional) [
      {
        "arg": (str, required) name of an argument of fn_name that represents a float value,
        "delta": (float, required) delta value to add to the argument during adaptive pose search,
        "min": (float, optional) minimum value during search,
        "max": (float, optional) maximum value during search
      }
    ]
  },
  
  {... next trajectory with similar structure as above ...}
]

Adding new trajectory functions

We provide a bunch of predefined trajectory functions in trajectory_util.py. Each trajectory.json file is a combination of the provided trajectory functions. You can create custom trajectories by creating new combinations of existing functions. You can also add custom trajectory functions in trajectory_util.py. For automatic integration with our codebase, custom trajectory functions should have the following pattern:

def custom_trajectory_fn(current_step, n_steps, **args):
    # n_steps: how many poses including start and end pose in this trajectory
    # current_step: pose in the current trajectory
    
    # your custom trajectory function here...

def custom_trajectory(**args):
    return _config_fn(custom_trajectory_fn, **args)

This lets you reference custom_trajectory as fn_name in a trajectory.json file.

Render an existing scene

We provide a script that renders images from a mesh at different poses:

python render_cameras.py -m <path/to/mesh.ply> -c <path/to/cameras.json>

where you can provide any cameras in the Pytorch3D convention via -c. For example, to re-render all poses used during generation and completion:

python render_cameras.py \
-m <output_root>/fused_mesh/fused_final_poisson_meshlab_depth_12.ply \
-c <output_root>/seen_poses.json

Optimize a NeRF

We provide an easy way to train a NeRF from our generated scene. We save a transforms.json file in the standard NeRF convention, that can be used to optimize a NeRF for the generated scene. It refers to the rendered images in <output_root>/output_rendering. It can be used with standard NeRF frameworks like Instant-NGP or NeRFStudio.

Acknowledgements

Our work builds on top of amazing open-source networks and codebases. We thank the authors for providing them.

• IronDepth [1]: a method for monocular depth prediction, that can be used for depth inpainting.
• StableDiffusion [2]: a state-of-the-art text-to-image inpainting model with publicly released network weights.

[1] IronDepth: Iterative Refinement of Single-View Depth using Surface Normal and its Uncertainty, BMVC 2022, Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla

[2] High-Resolution Image Synthesis with Latent Diffusion Models, CVPR 2022, Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer