threestudio is a unified framework for 3D content creation from text prompts, single images, and few-shot images, by lifting 2D text-to-image generation models.

👆 Results obtained from methods implemented by threestudio 👆
| DreamFusion | Magic3D | SJC | Latent-NeRF | Fantasia3D |

Installation

The following steps have been tested on Ubuntu20.04.

• You must have a NVIDIA graphics card with at least 6GB VRAM and have CUDA installed.
• Install Python >= 3.8.
• (Optional, Recommended) Create a virtual environment:

python3 -m virtualenv venv
. venv/bin/activate

• Install PyTorch >= 1.12. We have tested on torch1.12.1+cu113 and torch2.0.0+cu118, but other versions should also work fine.

# torch1.12.1+cu113
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 --extra-index-url https://download.pytorch.org/whl/cu113
# or torch2.0.0+cu118
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

• (Optional, Recommended) Install ninja to speed up the compilation of CUDA extensions:

pip install ninja

• Install dependencies:

pip install -r requirements.txt

• (Optional, Recommended) The best-performing models in threestudio uses the newly-released T2I model DeepFloyd IF which currently requires signing a license agreement. If you would like use these models, you need to accept the license on the model card of DeepFloyd IF, and login in the huggingface hub in terminal by huggingface-cli login.

• For contributors, see here.

Quickstart

Here we show some basic usage of threestudio. First let's train a DreamFusion model to create a classic pancake bunny.

IMPORTANT NOTE: Multi-GPU training is not fully tested and can be erroneous at the moment.

If you are experiencing unstable connections with HuggingFace, we suggest you either (1) setting environment variable TRANSFORMERS_OFFLINE=1 DIFFUSERS_OFFLINE=1 before your running command after all needed files have been fetched on the first run, to prevent from connecting to HuggingFace each time you run, or (2) downloading the guidance model you used to a local folder following here and here, and set pretrained_model_name_or_path of the guidance and the prompt processor to the local path.

# if you have agreed the license of DeepFloyd IF and have >20GB VRAM
# please try this configuration for higher quality
python launch.py --config configs/dreamfusion-if.yaml --train --gpu 0 system.prompt_processor.prompt="a zoomed out DSLR photo of a baby bunny sitting on top of a stack of pancakes"
# otherwise you could try with the Stable Diffusion model, which fits in 6GB VRAM
python launch.py --config configs/dreamfusion-sd.yaml --train --gpu 0 system.prompt_processor.prompt="a zoomed out DSLR photo of a baby bunny sitting on top of a stack of pancakes"

threestudio uses OmegaConf for flexible configurations. You can easily change any configuration in the YAML file by specifying arguments without --, for example the specified prompt in the above cases. For all supported configurations, please see our documentation.

The training lasts for 10,000 iterations. You can find visualizations of the current status in the trial directory which defaults to [exp_root_dir]/[name]/[tag]@[timestamp], where exp_root_dir (outputs/ by default), name and tag can be set in the configuration file. A 360-degree video will be generated after the training is completed. In training, press ctrl+c one time will stop training and head directly to the test stage which generates the video. Press ctrl+c the second time to fully quit the program. If you want to resume from a checkpoint, do:

# resume training from the last checkpoint, you may replace last.ckpt with any other checkpoints
python launch.py --config path/to/trial/dir/configs/parsed.yaml --train --gpu 0 resume=path/to/trial/configs/last.ckpt
# if the training has completed, you can still continue training for a longer time by setting trainer.max_steps
python launch.py --config path/to/trial/dir/configs/parsed.yaml --train --gpu 0 resume=path/to/trial/configs/last.ckpt trainer.max_steps=20000
# you can also perform testing using resumed checkpoints
python launch.py --config path/to/trial/dir/configs/parsed.yaml --test --gpu 0 resume=path/to/trial/configs/last.ckpt
# note that the above commands use parsed configuration files from previous trials
# which will continue using the same trial directory
# if you want to save to a new trial directory, replace parsed.yaml with raw.yaml in the command

See here for example running commands of all our supported models. Please refer to here for tips on getting higher-quality results, and here for reducing VRAM usage.

For feature requests, bug reports, or discussions about technical problems, please file an issue. In case you want to discuss the generation quality or showcase your generation results, please feel free to participate in the discussion panel.

Supported Models

DreamFusion

Results obtained by threestudio (DeepFloyd IF, batch size 8)

dreamfusion-if.mp4

Notable differences from the paper

• We use open-source T2I models (StableDiffusion, DeepFloyd IF), while the paper uses Imagen.
• We use a guiandance scale of 20 for DeepFloyd IF, while the paper uses 100 for Imagen.
• We do not use sigmoid to normalize the albedo color but simply scale the color from [-1,1] to [0,1], as we find this help convergence.
• We use HashGrid encoding and uniformly sample points along rays, while the paper uses Integrated Positional Encoding and sampling strategy from MipNeRF360.
• We adopt camera settings and density initialization strategy from Magic3D, which is slightly different from the DreamFusion paper.
• Some hyperparameters are different, such as the weighting of loss terms.

Example running commands

# uses DeepFloyd IF, requires ~15GB VRAM to extract text embeddings and ~10GB VRAM in training
# here we adopt random background augmentation to improve geometry quality
python launch.py --config configs/dreamfusion-if.yaml --train --gpu 0 system.prompt_processor.prompt="a delicious hamburger" system.background.random_aug=true
# uses StableDiffusion, requires ~6GB VRAM in training
python launch.py --config configs/dreamfusion-sd.yaml --train --gpu 0 system.prompt_processor.prompt="a delicious hamburger"

Tips

• DeepFloyd IF performs way better than StableDiffusion.
• Validation shows albedo color before system.material.ambient_only_steps and shaded color after that.
• Try increasing/decreasing system.loss.lambda_sparsity if your scene is stuffed with floaters/becoming empty.
• Try increasing/decreasing system.loss.lambda_orient if you object is foggy/over-smoothed.
• Try replacing the background to random colors with a probability 0.5 by setting system.background.random_aug=true if you find the model incorrectly treats the background as part of the object.
• DeepFloyd IF uses T5-XXL as its text encoder, which consumes ~15GB VRAM even when using 8-bit quantization. This is currently the bottleneck for training with less VRAM. If anyone knows how to run the text encoder with less VRAM, please file an issue. We're also trying to push the text encoder to Replicate to enable extracting text embeddings via API, but are having some network connection issues. Please contact bennyguo if you would like to help out.

Magic3D

Results obtained by threestudio (DeepFloyd IF, batch size 8; first row: coarse, second row: refine)

magic3d-if.mp4

Notable differences from the paper

• We use open-source T2I models (StableDiffusion, DeepFloyd IF) for the coarse stage, while the paper uses eDiff-I.
• In the coarse stage, we use a guiandance scale of 20 for DeepFloyd IF, while the paper uses 100 for eDiff-I.
• There are many things that are ommited from the paper such as the weighting of loss terms and the DMTet grid resolution, which could be different.

Example running commands

First train the coarse stage NeRF:

# uses DeepFloyd IF, requires ~15GB VRAM to extract text embeddings and ~10GB VRAM in training
python launch.py --config configs/magic3d-coarse-if.yaml --train --gpu 0 system.prompt_processor.prompt="a delicious hamburger"
# uses StableDiffusion, requires ~6GB VRAM in training
python launch.py --config configs/magic3d-coarse-sd.yaml --train --gpu 0 system.prompt_processor.prompt="a delicious hamburger"

Then convert the NeRF from the coarse stage to DMTet and train with differentiable rasterization:

# the refinement stage uses StableDiffusion, requires ~5GB VRAM in training
python launch.py --config configs/magic3d-refine-sd.yaml --train --gpu 0 system.prompt_processor.prompt="a delicious hamburger" system.weights=path/to/coarse/stage/trial/ckpts/last.ckpt

Tips

• For the caorse stage, DeepFloyd IF performs way better than StableDiffusion.
• Magic3D uses a neural network to predict the surface normal, which may not resemble the true geometric normal, so it's common to see that your object becomes extremely dark after system.material.ambient_only_steps.
• Try increasing/decreasing system.loss.lambda_sparsity if your scene is stuffed with floaters/becoming empty.
• Try replacing the background to random colors with a probability 0.5 by setting system.background.random_aug=true if you find the model incorrectly treats the background as part of the object.

Score Jacobian Chaining

Results obtained by threestudio (Stable Diffusion)

sjc.mp4

Notable differences from the paper: N/A.

Example running commands

# train with sjc guidance in latent space
python launch.py --config configs/sjc.yaml --train --gpu 0 system.prompt_processor.prompt="A high quality photo of a delicious burger"
# train with sjc guidance in latent space, trump figure
python launch.py --config configs/sjc.yaml --train --gpu 0 system.prompt_processor.prompt="Trump figure" trainer.max_steps=30000 system.loss.lambda_emptiness="[15000,10000.0,200000.0,15001]" system.optimizer.params.background.lr=0.05 seed=42

Tips

• SJC uses subpixel rendering which decodes a 128x128 latent feature map for better visualization quality. You can turn off this feature by system.subpixel_rendering=false to save VRAM in validation/testing.

Latent-NeRF

Results obtained by threestudio (Stable Diffusion)

latent-nerf.mp4

Notable differences from the paper: N/A.

We currently only implement Latent-NeRF for text-guided and Sketch-Shape for (text,shape)-guided 3D generation. Latent-Paint is not implemented yet.

Example running commands

# train Latent-NeRF in Stable Diffusion latent space
python launch.py --config configs/latentnerf.yaml --train --gpu 0 system.prompt_processor.prompt="a delicious hamburger"
# refine Latent-NeRF in RGB space
python launch.py --config configs/latentnerf-refine.yaml --train --gpu 0 system.prompt_processor.prompt="a delicious hamburger" system.weights=path/to/latent/stage/trial/ckpts/last.ckpt

# train Sketch-Shape in Stable Diffusion latent space
python launch.py --config configs/sketchshape.yaml --train --gpu 0 system.guide_shape=load/shapes/teddy.obj system.prompt_processor.prompt="a teddy bear in a tuxedo"
# refine Sketch-Shape in RGB space
python launch.py --config configs/sketchshape-refine.yaml --train --gpu 0 system.shape_guide=load/shapes/teddy.obj system.prompt_processor.prompt="a teddy bear in a tuxedo" system.weights=path/to/latent/stage/trial/ckpts/last.ckpt

Fantasia3D

Results obtained by threestudio (Stable Diffusion)

fantasia-3d.mp4

Notable differences from the paper: N/A.

We currently only implement the geometry stage of Fantasia3D.

Example running commands

python launch.py --config configs/fantasia3d.yaml --train --gpu 0 system.prompt_processor.prompt="a DSLR photo of an ice cream sundae"
# Fantasia3D highly relies on the initialized SDF shape
# the default shape is a sphere with radius 0.5
# change the shape initialization to match your input prompt
python launch.py --config configs/fantasia3d.yaml --train --gpu 0 system.prompt_processor.prompt="The leaning tower of Pisa" system.geometry.shape_init=ellipsoid system.geometry.shape_init_params="[0.3,0.3,0.8]"

Tips

• If you find the shape easily diverge in early training stages, you may use a lower guidance scale by setting system.guidance.guidance_scale=30..

More to come, please stay tuned.

• Dream3D
• DreamAvatar

If you would like to contribute a new method to threestudio, see here.

Prompt Library

For easier comparison, we collect the 397 preset prompts from the website of DreamFusion in this file. You can use these prompts by setting system.prompt_processor.prompt=lib:keyword1_keyword2_..._keywordN. Note that the prompt should starts with lib: and all the keywords are separated by _. The prompt processor will match the keywords to all the prompts in the library, and will only succeed if there's exactly one match. The used prompt will be printed to console. Also note that you can't use this syntax to point to every prompt in the library, as there are prompts that are subset of other prompts lmao. We will enhance the use of this feature.

Tips on Improving Quality

It's important to note that existing techniques that lift 2D T2I models to 3D cannot consistently produce satisfying results. Results from the great papers like DreamFusion and Magic3D are (to some extend) cherry-pickled, so don't be frustrated if you did not get what you expected on your first trial. Here are some tips that may help you improve the generation quality:

• Increase batch size. Large batch sizes help convergence and improve the 3D consistency of the geometry. State-of-the-art methods claims using large batch sizes: DreamFusion uses a batch size of 4; Magic3D uses a batch size of 32; Fantasia3D uses a batch size of 24; some results shown above uses a batch size of 8. You can easily change the batch size by setting data.batch_size=N. Increasing the batch size requires more VRAM. If you have limited VRAM but still want the benefit of large batch sizes, you may use gradient accumulation provided by PyTorch Lightning by setting trainer.accumulate_grad_batches=N. This will accumulate the gradient of several batches and achieve a large effective batch size. Note that if you use gradient accumulation, you may need to multiply all step values by N times in your config, such as values that have the name X_steps and trainer.val_check_interval, since now N batches equal to a large batch.
• Train longer. This helps if you can already obtain reasonable results and would like to enhance the details. If the result is still a mess after several thousand steps, training for a longer time often won't help. You can set the total training iterations by trainer.max_steps=N.
• Try different seeds. This is a simple solution if your results have correct overall geometry but suffer from the multi-face Janus problem. You can change the seed by setting seed=N. Good luck!
• Tuning regularization weights. Some methods have regularizaion terms which can be essential to obtaining good geometry. Try tuning the weights of these regularizations by setting system.loss.lambda_X=value. The specific values depend on your situation, you may refer to tips for each supported model for more detailed instructions.

VRAM Optimization

If you encounter CUDA OOM error, try the following in order (roughly sorted by recommendation) to meet your VRAM requirement.

• If you only encounter OOM at validation/test time, you can set system.cleanup_after_validation_step=true and system.cleanup_after_test_step=true to free memory after each validation/test step. This will slow down validation/testing.
• Use a smaller batch size or use gradient accumulation as demonstrated here.
• If you are using PyTorch1.x, enable memory efficient attention by setting system.guidance.enable_memory_efficient_attention=true. PyTorch2.0 has built-in support for this optimization and is enabled by default.
• Enable attention slicing by setting system.guidance.enable_attention_slicing=true. This will slow down training by ~20%.
• If you are using StableDiffusionGuidance, you can use Token Merging to drastically speed up computation and save memory. You can easily enable Token Merging by setting system.guidance.token_merging=true. You can also customize the Token Merging behavior by setting the parameters here to system.guidance.token_merging_params. Note that Token Merging may degrade generation quality.
• Enable sequential CPU offload by setting system.guidance.enable_sequential_cpu_offload=true. This could save a lot of VRAM but will make the training extremely slow.

Documentation

threestudio use OmegaConf to manage configurations. You can literally change anything inside the yaml configuration file or by adding command line arguments without --. We list all arguments that you can change in the configuration in our documentation. Happy experimenting!

Contributing to threestudio

• Fork the repository and create your branch from main.
• Install development dependencies:

pip install -r requirements-dev.txt

• If you are using VSCode as the text editor: (1) Install editorconfig extension. (2) Set the default linter to mypy to enable static type checking. (3) Set the default formatter to black. You could either manually format the document or let the editor format the document each time it is saved by setting "editor.formatOnSave": true.

• Run pre-commit install to install pre-commit hooks which will automatically format the files before commit.

• Make changes to the code, update README and DOCUMENTATION if needed, and open a pull request.

Code Structure

Here we just briefly introduce the code structure of this project. We will make more detailed documentation about this in the future.

• All methods are implemented as a subclass of BaseSystem (in systems/base.py). There typically are six modules inside a system: geometry, material, background, renderer, guidance, and prompt_processor. All modules are subclass of BaseModule (in utils/base.py) except for guidance, and prompt_processor, which are subclass of BaseObject to prevent them from being treated as model parameters and better control their behavior in multi-GPU settings.
• All systems, modules, and data modules have their configurations in their own dataclasses.
• Base configurations for the whole project can be found in utils/config.py. In the ExperimentConfig dataclass, data, system, and module configurations under system are parsed to configurations of each class mentioned above. These configurations are strictly typed, which means you can only use defined properties in the dataclass and stick to the defined type of each property. This configuration paradigm (1) natually supports default values for properties; (2) effectively prevents wrong assignments of these properties (say typos in the yaml file) or inappropriate usage at runtime.
• This projects use both static and runtime type checking. For more details, see utils/typing.py.
• To update anything of a module at each training step, simply make it inherit to Updateable (see utils/base.py). At the beginning of each iteration, an Updateable will update itself, and update all its attributes that are also Updateable. Note that subclasses of BaseSystem, BaseModule and BaseObject are by default inherit to Updateable.

Known Problems

• Gradients of Vanilla MLP parameters are empty in AMP (temporarily fixed by disabling autocast).
• FullyFused MLP may cause NaNs in 32 precision.

Citing threestudio

If you find threestudio helpful, please consider citing:

@Misc{threestudio2023,
  author =       {Yuan-Chen Guo and Ying-Tian Liu and Chen Wang and Zi-Xin Zou and Guan Luo and Chia-Hao Chen and Yan-Pei Cao and Song-Hai Zhang},
  title =        {threestudio: A unified framework for 3D content generation},
  howpublished = {\url{https://github.com/threestudio-project/threestudio}},
  year =         {2023}
}