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 👆
| ProlificDreamer | DreamFusion | Magic3D | SJC | Latent-NeRF | Fantasia3D | TextMesh |
| Zero-1-to-3 | Magic123 |
| InstructNeRF2NeRF | Control4D |
Did not find what you want? Checkout threestudio-extension or submit a feature request here!
| Animate-124 | 4D-fy | DreamCraft3D | Gaussian Splatting | MVDream | Mesh-Fitting |
- 23/12/2023: Thank Yuyang Zhao for implementation of image-to-4D generation extensions Animate-124! Follow the instructions on the extensions website to give it a try.
- 18/12/2023: Implementation of 4D-fy for 4D generation and DreamCraft3D for high-quality image-to-3D generation as the custom extensions! Follow the instructions on the extensions website to give it a try.
- 13/12/2023: Implementation supporting Stable Zero123 for 3D generation from a single image! Follow the instructions here to give it a try.
- 30/11/2023: Implementation of MVDream, Gaussian Splatting as the custom extensions. You can also use neural representation to fit a mesh by Mesh-Fitting.
- 30/11/2023: Implementation of custom extension system and you can add your extensions in this project.
- 25/06/2023: Implementation of Magic123! Follow the instructions here to give it a try.
- 06/07/2023: Join our Discord server for lively discussions!
- 03/07/2023: Try text-to-3D online in HuggingFace Spaces or using our self-hosted service (GPU support from Tencent). To host the web interface locally, see here.
- 20/06/2023: Implementations of Instruct-NeRF2NeRF and Control4D for high-fidelity 3D editing! Follow the instructions for Control4D and Instruct-NeRF2NeRF to give it a try.
- 14/06/2023: Implementation of TextMesh! Follow the instructions here to give it a try.
- 14/06/2023: Implementation of prompt debiasing and Perp-Neg! Follow the instructions here to give it a try.
- 29/05/2023: An experimental implementation of using Zero-1-to-3 for 3D generation from a single image! Follow the instructions here to give it a try.
- 26/05/2023: Implementation of ProlificDreamer! Follow the instructions here to give it a try.
- 14/05/2023: You can experiment with the SDS loss on 2D images using our 2dplayground.
- 13/05/2023: You can now try threestudio on Google Colab!
- 11/05/2023: We now support exporting textured meshes! See here for instructions.
See installation.md for additional information, including installation via Docker.
The following steps have been tested on Ubuntu20.04.
- You must have an 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
# Newer pip versions, e.g. pip-23.x, can be much faster than old versions, e.g. pip-20.x.
# For instance, it caches the wheels of git packages to avoid unnecessarily rebuilding them later.
python3 -m pip install --upgrade pip
- Install
PyTorch >= 1.12
. We have tested ontorch1.12.1+cu113
andtorch2.0.0+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)
tiny-cuda-nn
installation might require downgrading pip to 23.0.1 -
(Optional, Recommended) The best-performing models in threestudio use the newly-released T2I model DeepFloyd IF, which currently requires signing a license agreement. If you would like to use these models, you need to accept the license on the model card of DeepFloyd IF, and login into the Hugging Face hub in the terminal by
huggingface-cli login
. -
For contributors, see here.
Here we show some basic usage of threestudio. First let's train a DreamFusion model to create a classic pancake bunny.
If you are experiencing unstable connections with Hugging Face, we suggest you either (1) setting environment variable TRANSFORMERS_OFFLINE=1 DIFFUSERS_OFFLINE=1 HF_HUB_OFFLINE=1
before your running command after all needed files have been fetched on the first run, to prevent from connecting to Hugging Face 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.
Multi-GPU training is supported, but may still be buggy. Note that data.batch_size
is the batch size per rank (device). Also remember to
- Set
data.n_val_views
to be a multiple of the number of GPUs. - Set a unique
tag
as timestamp is disabled in multi-GPU training and will not be appended after the tag. If you the same tag as previous trials, saved config files, code and visualizations will be overridden.
# this results in an effective batch size of 4 (number of GPUs) * 2 (data.batch_size) = 8
python launch.py --config configs/dreamfusion-if.yaml --train --gpu 0,1,2,3 system.prompt_processor.prompt="a zoomed out DSLR photo of a baby bunny sitting on top of a stack of pancakes" data.batch_size=2 data.n_val_views=4
If you define the CUDA_VISIBLE_DEVICES
environment variable before you call launch.py
, you don't need to specify --gpu
- this will use all available GPUs from CUDA_VISIBLE_DEVICES
. For instance, the following command will automatically use GPUs 3 and 4:
CUDA_VISIBLE_DEVICES=3,4 python launch.py --config configs/dreamfusion-if.yaml --train system.prompt_processor.prompt="a zoomed out DSLR photo of a baby bunny sitting on top of a stack of pancakes"
This is particularly useful if you run launch.py
in a cluster using a command that automatically picks GPU(s) and exports their IDs through CUDA_VISIBLE_DEVICES, e.g. through SLURM:
cd git/threestudio
. venv/bin/activate
srun --account mod3d --partition=g40 --gpus=1 --job-name=3s_bunny python launch.py --config configs/dreamfusion-if.yaml --train system.prompt_processor.prompt="a zoomed out DSLR photo of a baby bunny sitting on top of a stack of pancakes"
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/dir/ckpts/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/dir/ckpts/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/dir/ckpts/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
# only load weights from saved checkpoint but dont resume training (i.e. dont load optimizer state):
python launch.py --config path/to/trial/dir/configs/parsed.yaml --train --gpu 0 system.weights=path/to/trial/dir/ckpts/last.ckpt
To export the scene to texture meshes, use the --export
option. We currently support exporting to obj+mtl, or obj with vertex colors.
# this uses default mesh-exporter configurations which exports obj+mtl
python launch.py --config path/to/trial/dir/configs/parsed.yaml --export --gpu 0 resume=path/to/trial/dir/ckpts/last.ckpt system.exporter_type=mesh-exporter
# specify system.exporter.fmt=obj to get obj with vertex colors
# you may also add system.exporter.save_uv=false to accelerate the process, suitable for a quick peek of the result
python launch.py --config path/to/trial/dir/configs/parsed.yaml --export --gpu 0 resume=path/to/trial/dir/ckpts/last.ckpt system.exporter_type=mesh-exporter system.exporter.fmt=obj
# for NeRF-based methods (DreamFusion, Magic3D coarse, Latent-NeRF, SJC)
# you may need to adjust the isosurface threshold (25 by default) to get satisfying outputs
# decrease the threshold if the extracted model is incomplete, increase if it is extruded
python launch.py --config path/to/trial/dir/configs/parsed.yaml --export --gpu 0 resume=path/to/trial/dir/ckpts/last.ckpt system.exporter_type=mesh-exporter system.geometry.isosurface_threshold=10.
# use marching cubes of higher resolutions to get more detailed models
python launch.py --config path/to/trial/dir/configs/parsed.yaml --export --gpu 0 resume=path/to/trial/dir/ckpts/last.ckpt system.exporter_type=mesh-exporter system.geometry.isosurface_method=mc-cpu system.geometry.isosurface_resolution=256
For all the options you can specify when exporting, see the documentation.
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.
Launch the Gradio web interface by
python gradio_app.py launch
Parameters:
--listen
: listens to all addresses by settingserver_name="0.0.0.0"
when launching the Gradio app.--self-deploy
: enables changing arbitrary configurations directly from the web.--save
: enables checkpoint saving.
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.
This is an unofficial experimental implementation! Please refer to https://github.com/thu-ml/prolificdreamer for official code release.
Results obtained by threestudio (Stable Diffusion, 256x256 Stage1)
prolificdreamer.mp4
prolificdreamer-scene.mp4
Results obtained by threestudio (Stable Diffusion, 256x256 Stage1, 512x512 Stage2+3)
prolificdreamer-full.mp4
Notable differences from the paper:
- ProlificDreamer adopts a two-stage sampling strategy with 64 coarse samples and 32 fine samples, while we only use 512 coarse samples.
- In the first stage, we only render 64x64 images at the first 5000 iterations. After that, as the empty space has been effectively pruned, rendering 512x512 images wouldn't cost too much VRAM.
- We currently don't support multiple particles.
# --------- Stage 1 (NeRF) --------- #
# object generation with 512x512 NeRF rendering, ~30GB VRAM
python launch.py --config configs/prolificdreamer.yaml --train --gpu 0 system.prompt_processor.prompt="a pineapple"
# if you don't have enough VRAM, try training with 64x64 NeRF rendering, ~15GB VRAM
python launch.py --config configs/prolificdreamer.yaml --train --gpu 0 system.prompt_processor.prompt="a pineapple" data.width=64 data.height=64 data.batch_size=1
# using the same model for pretrained and LoRA enables 64x64 training with <10GB VRAM
# but the quality is worse due to the use of an epsilon prediction model for LoRA training
python launch.py --config configs/prolificdreamer.yaml --train --gpu 0 system.prompt_processor.prompt="a pineapple" data.width=64 data.height=64 data.batch_size=1 system.guidance.pretrained_model_name_or_path_lora="stabilityai/stable-diffusion-2-1-base"
# Using patch-based renderer to reduce memory consume, 512x512 resolution, ~20GB VRAM
python launch.py --config configs/prolificdreamer-patch.yaml --train --gpu 0 system.prompt_processor.prompt="a pineapple"
# scene generation with 512x512 NeRF rendering, ~30GB VRAM
python launch.py --config configs/prolificdreamer-scene.yaml --train --gpu 0 system.prompt_processor.prompt="Inside of a smart home, realistic detailed photo, 4k"
# --------- Stage 2 (Geometry Refinement) --------- #
# refine geometry with 512x512 rasterization, Stable Diffusion SDS guidance
python launch.py --config configs/prolificdreamer-geometry.yaml --train --gpu 0 system.prompt_processor.prompt="a pineapple" system.geometry_convert_from=path/to/stage1/trial/dir/ckpts/last.ckpt
# --------- Stage 3 (Texturing) --------- #
# texturing with 512x512 rasterization, Stable Difusion VSD guidance
python launch.py --config configs/prolificdreamer-texture.yaml --train --gpu 0 system.prompt_processor.prompt="a pineapple" system.geometry_convert_from=path/to/stage2/trial/dir/ckpts/last.ckpt
This is a re-implementation, missing some improvements from the original paper(coarse-to-fine NeRF sampling, kernel smoothing). For original results, please refer to https://github.com/JunzheJosephZhu/HiFA
HiFA is more like a suite of improvements including image space SDS, z-variance loss, and noise strength annealing. It is compatible with most optimization-based methods. Therefore, we provide three variants based on DreamFusion, ProlificDreamer, and Magic123. We provide a unified guidance config as well as an SDS/VSD guidance config for the DreamFusion and ProlificDreamer variants, both configs should achieve the same results. Additionally, we also make HiFA compatible with ProlificDreamer-scene.
Results obtained by threestudio(Dreamfusion-HiFA, 512x512)
hifa_combined.mp4
Results obtained by threestudio(ProlificDreamer-HiFA, 512x512)
prolific-combined.mp4
Results obtained by threestudio(Magic123-HiFA, 512x512)
magic-combined.mp4
Example running commands
# ------ DreamFusion-HiFA ------- # (similar to original paper)
python launch.py --config configs/hifa.yaml --train --gpu 0 system.prompt_processor.prompt="a plate of delicious tacos"
python launch.py --config configs/experimental/unified-guidance/hifa.yaml --train --gpu 0 system.prompt_processor.prompt="a plate of delicious tacos"
# ------ ProlificDreamer-HiFA ------- #
python launch.py --config configs/prolificdreamer-hifa.yaml --train --gpu 0 system.prompt_processor.prompt="a plate of delicious tacos"
python launch.py --config configs/experimental/unified-guidance/prolificdreamer-hifa.yaml --train --gpu 0 system.prompt_processor.prompt="a plate of delicious tacos"
# ------ ProlificDreamer-scene-HiFA ------- #
python launch.py --config configs/prolificdreamer-scene-hifa.yaml --train --gpu 0 system.prompt_processor.prompt="A DSLR photo of a hamburger inside a restaurant"
# ------ Magic123-HiFA ------ #
python launch.py --config configs/magic123-hifa-coarse-sd.yaml --train --gpu 0 data.image_path=load/images/firekeeper_rgba.png system.prompt_processor.prompt="a toy figure of firekeeper from dark souls"
# We included a config for magic123's refine stage, but didn't really run it, since the coarse stage result already looks pretty decent.
Tips
- If the generated object's color seems oversaturated, decrease lambda_sds_img(or lambda_sd_img if using unified guidance).
- If the generated object looks cloudy, increase lamda_z_variance. If the shape becomes corrupted, decrease lambda_z_variance.
- If the generated object overall seems to have high luminance, increase min_step_percent.
- Make sure sqrt_anneal and use_img_loss are both set to True.
- Check out the original repo! The results are better.
- If you are using sqrt_anneal, make sure system.guidance.trainer_max_steps is equal to trainer.max_steps, so noise strength annealing works correctly
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 guidance 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.
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 guidance scale of 20 for DeepFloyd IF, while the paper uses 100 for eDiff-I.
- In the coarse stage, we use analytic normal, while the paper uses predicted normal.
- In the coarse stage, we use orientation loss as in DreamFusion, while the paper does not.
- There are many things that are omitted 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, and 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.geometry_convert_from=path/to/coarse/stage/trial/dir/ckpts/last.ckpt
# if you're unsatisfied with the surface extracted using the default threshold (25)
# you can specify a threshold value using `system.geometry_convert_override`
# decrease the value if the extracted surface is incomplete, increase if it is extruded
python launch.py --config configs/magic3d-refine-sd.yaml --train --gpu 0 system.prompt_processor.prompt="a delicious hamburger" system.geometry_convert_from=path/to/coarse/stage/trial/dir/ckpts/last.ckpt system.geometry_convert_override.isosurface_threshold=10.
Tips
- For the coarse 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 and degrade geometry quality, so we use analytic normal instead.
- 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 with 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.
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 bysystem.subpixel_rendering=false
to save VRAM in validation/testing.
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/dir/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.guide_shape=load/shapes/teddy.obj system.prompt_processor.prompt="a teddy bear in a tuxedo" system.weights=path/to/latent/stage/trial/dir/ckpts/last.ckpt
Results obtained by threestudio (Stable Diffusion)
fantasia-3d.mp4
Results obtained by threestudio (Stable Diffusion, mesh initialization)
fantasia3d-mesh.mp4
Notable differences from the paper:
- We enable tangent-space normal perturbation by default, which can be turned off by appending
system.material.use_bump=false
.
Example running commands
# --------- Geometry --------- #
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]"
# or you can initialize from a mesh
# here shape_init_params is the scale of the shape
# also make sure to input the correct up and front axis (in +x, +y, +z, -x, -y, -z)
python launch.py --config configs/fantasia3d.yaml --train --gpu 0 system.prompt_processor.prompt="hulk" system.geometry.shape_init=mesh:load/shapes/human.obj system.geometry.shape_init_params=0.9 system.geometry.shape_init_mesh_up=+y system.geometry.shape_init_mesh_front=+z
# --------- Texture --------- #
# to train PBR texture continued from a geometry checkpoint:
python launch.py --config configs/fantasia3d-texture.yaml --train --gpu 0 system.prompt_processor.prompt="a DSLR photo of an ice cream sundae" system.geometry_convert_from=path/to/geometry/stage/trial/dir/ckpts/last.ckpt
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.
.
Results obtained by threestudio (DeepFloyd IF, batch size 4)
textmesh-if.mp4
Notable differences from the paper
- Most of the settings are the same as the DreamFusion model. Please refer to the notable differences of the DreamFusion model.
- We use NeuS as the geometry representation while the original paper uses VolSDF.
- We adopt techniques from Neuralangelo to stabilize normal computation when using hash grids.
- We currently only implemented the coarse stage of TextMesh.
Example running commands
# uses DeepFloyd IF, requires ~15GB VRAM
python launch.py --config configs/textmesh-if.yaml --train --gpu 0 system.prompt_processor.prompt="lib:cowboy_boots"
Tips
- TextMesh uses a surface-based geometry representation, so you don't need to manually tune the isosurface threshold when exporting meshes!
This is an experimental implementation of Control4D using threestudio! Control4D will release the full code including static and dynamic editing after paper acceptance.
Results obtained by threestudio (512x512)
origin_1.mp4
We currently don't support dynamic editing.
Download the data sample of control4D using this link.
Example running commands
# --------- Control4D --------- #
# static editing with 128x128 NeRF + 512x512 GAN rendering, ~20GB VRAM
python launch.py --config configs/control4d-static.yaml --train --gpu 0 data.dataroot="YOUR_DATAROOT/twindom" system.prompt_processor.prompt="Elon Musk wearing red shirt, RAW photo, (high detailed skin:1.2), 8k uhd, dslr, soft lighting, high quality, film grain, Fujifilm XT3"
Results obtained by threestudio
in2n.mp4
Download the data sample of InstructNeRF2NeRF using this link.
Example running commands
# --------- InstructNeRF2NeRF --------- #
# 3D editing with NeRF patch-based rendering, ~20GB VRAM
python launch.py --config configs/instructnerf2nerf.yaml --train --gpu 0 data.dataroot="YOUR_DATAROOT/face" data.camera_layout="front" data.camera_distance=1 data.eval_interpolation=[1,3,50] system.prompt_processor.prompt="Turn him into Albert Einstein"
Results obtained by threestudio (Zero123 + Stable Diffusion)
magic123.mp4
Notable differences from the paper
- This is an unofficial re-implementation which shares the same overall idea with the official implementation but differs in some aspects like hyperparameters.
- Textual Inversion is not supported, which means a text prompt is needed for training.
Example running commands
First train the coarse stage NeRF:
# Zero123 + Stable Diffusion, ~12GB VRAM
# data.image_path must point to a 4-channel RGBA image
# system.prompt_proessor.prompt must be specified
python launch.py --config configs/magic123-coarse-sd.yaml --train --gpu 0 data.image_path=load/images/hamburger_rgba.png system.prompt_processor.prompt="a delicious hamburger"
Then convert the NeRF from the coarse stage to DMTet and train with differentiable rasterization:
# Zero123 + Stable Diffusion, ~10GB VRAM
# data.image_path must point to a 4-channel RGBA image
# system.prompt_proessor.prompt must be specified
python launch.py --config configs/magic123-refine-sd.yaml --train --gpu 0 data.image_path=load/images/hamburger_rgba.png system.prompt_processor.prompt="a delicious hamburger" system.geometry_convert_from=path/to/coarse/stage/trial/dir/ckpts/last.ckpt
# if you're unsatisfied with the surface extracted using the default threshold (25)
# you can specify a threshold value using `system.geometry_convert_override`
# decrease the value if the extracted surface is incomplete, increase if it is extruded
python launch.py --config configs/magic123-refine-sd.yaml --train --gpu 0 data.image_path=load/images/hamburger_rgba.png system.prompt_processor.prompt="a delicious hamburger" system.geometry_convert_from=path/to/coarse/stage/trial/dir/ckpts/last.ckpt system.geometry_convert_override.isosurface_threshold=10.
Tips
- If the image contains non-front-facing objects, specifying the approximate elevation and azimuth angle by setting
data.default_elevation_deg
anddata.default_azimuth_deg
can be helpful. In threestudio, top is elevation +90 and bottom is elevation -90; left is azimuth -90 and right is azimuth +90.
Installation
Download pretrained Stable Zero123 checkpoint stable-zero123.ckpt
into load/zero123
from https://huggingface.co/stabilityai/stable-zero123
Results obtained by threestudio (Stable Zero123 vs Zero123-XL)
Direct multi-view images generation If you only want to generate multi-view images, please refer to threestudio-mvimg-gen. This extension can use Stable Zero123 to directly generate images from multi-view perspectives.
Example running commands
- Take an image of your choice, or generate it from text using your favourite AI image generator such as SDXL Turbo (https://clipdrop.co/stable-diffusion-turbo) E.g. "A simple 3D render of a friendly dog"
- Remove its background using Clipdrop (https://clipdrop.co/remove-background)
- Save to
load/images/
, preferably with_rgba.png
as the suffix - Run Zero-1-to-3 with the Stable Zero123 ckpt:
python launch.py --config configs/stable-zero123.yaml --train --gpu 0 data.image_path=./load/images/hamburger_rgba.png
IMPORTANT NOTE: This is an experimental implementation and we're constantly improving the quality.
IMPORTANT NOTE: This implementation extends the Zero-1-to-3 implementation below, and is heavily inspired from the Zero-1-to-3 implementation in https://github.com/ashawkey/stable-dreamfusion! extern/ldm_zero123
is borrowed from stable-dreamfusion/ldm
.
Installation
Download pretrained Zero123XL weights into load/zero123
:
cd load/zero123
wget https://zero123.cs.columbia.edu/assets/zero123-xl.ckpt
Results obtained by threestudio (Zero-1-to-3)
ezgif-3-355a192487.mp4
IMPORTANT NOTE: This is an experimental implementation and we're constantly improving the quality.
IMPORTANT NOTE: This implementation is heavily inspired from the Zero-1-to-3 implementation in https://github.com/ashawkey/stable-dreamfusion! extern/ldm_zero123
is borrowed from stable-dreamfusion/ldm
.
Example running commands
- Take an image of your choice, or generate it from text using your favourite AI image generator such as Stable Diffusion XL (https://clipdrop.co/stable-diffusion) E.g. "A simple 3D render of a friendly dog"
- Remove its background using Clipdrop (https://clipdrop.co/remove-background)
- Save to
load/images/
, preferably with_rgba.png
as the suffix - Run Zero-1-to-3:
python launch.py --config configs/zero123.yaml --train --gpu 0 data.image_path=./load/images/dog1_rgba.png
For more scripts for Zero-1-to-3, please check threestudio/scripts/run_zero123.sh
.
Previous Zero-1-to-3 weights are available at https://huggingface.co/cvlab/zero123-weights/
. You can download them to load/zero123
as above, and replace the path at system.guidance.pretrained_model_name_or_path
.
Guidance evaluation
Also includes evaluation of the guidance during training. If system.freq.guidance_eval
is set to a value > 0, this will save rendered image, noisy image (noise added mentioned at top left), 1-step-denoised image, 1-step prediction of original image, fully denoised image. For example:
If you would like to contribute a new method to threestudio, see here.
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 the 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.
It's important to note that existing techniques that lift 2D T2I models to 3D cannot consistently produce satisfying results. Results from great papers like DreamFusion and Magic3D are (to some extent) cherry-pickled, so don't be frustrated if you do 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 claim 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 use 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 settingtrainer.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 nameX_steps
andtrainer.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 regularization 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. - Try debiasing methods. When conventional SDS techniques like DreamFusion, Magic3D, SJC, and others fail to produce the desired 3D results, Debiased Score Distillation Sampling (D-SDS) can be a solution. D-SDS is devised to tackle challenges such as artifacts or the Janus problem, employing two strategies: score debiasing and prompt debiasing. You can activate score debiasing by just setting
system.guidance.grad_clip=[0,0.5,2.0,10000]
, where the order isstart_step, start_value, end_value, end_step
. You can enable prompt debiasing by settingsystem.prompt_processor.use_prompt_debiasing=true
. When using prompt debiasing, it's recommended to set a list of indices for words that should potentially be removed bysystem.prompt_processor.prompt_debiasing_mask_ids=[i1,i2,...]
. For example, if the prompt isa smiling dog
and you only want to remove the wordsmiling
for certain views, you should set it to[1]
. You could also manually specify the prompt for each view by settingsystem.prompt_processor.prompt_side
,system.prompt_processor.prompt_back
andsystem.prompt_processor.prompt_overhead
. For a detailed explanation of these techniques, refer to the D-SDS paper or check out the project page. - Try Perp-Neg. The Perp-Neg algorithm can potentially alleviate the multi-face Janus problem. We now support Perp-Neg for
stable-diffusion-guidance
anddeep-floyd-guidance
by settingsystem.prompt_processor.use_perp_neg=true
.
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
andsystem.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 tosystem.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.
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!
To enable the (experimental) wandb support, set system.loggers.wandb.enable=true
, e.g.:
python launch.py --config configs/zero123.yaml --train --gpu 0 system.loggers.wandb.enable=true`
If you're using a corporate wandb server, you may first need to login to your wandb instance, e.g.:
wandb login --host=https://COMPANY_XYZ.wandb.io --relogin
By default the runs will have a random name, recorded in the threestudio
project. You can override them to give a more descriptive name, e.g.:
python launch.py --config configs/zero123.yaml --train --gpu 0 system.loggers.wandb.enable=true system.loggers.wandb.name="zero123xl_accum;bs=4;lr=0.05"
- 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.
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
(insystems/base.py
). There typically are six modules inside a system: geometry, material, background, renderer, guidance, and prompt_processor. All modules are subclass ofBaseModule
(inutils/base.py
) except for guidance, and prompt_processor, which are subclass ofBaseObject
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 theExperimentConfig
dataclass,data
,system
, and module configurations undersystem
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) naturally 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
(seeutils/base.py
). At the beginning of each iteration, anUpdateable
will update itself, and update all its attributes that are alsoUpdateable
. Note that subclasses ofBaseSystem
,BaseModule
andBaseObject
are by default inherited toUpdateable
.
- Gradients of Vanilla MLP parameters are empty in AMP (temporarily fixed by disabling autocast).
- FullyFused MLP may cause NaNs in 32 precision.
threestudio is built on the following amazing open-source projects:
- Lightning Framework for creating highly organized PyTorch code.
- OmegaConf Flexible Python configuration system.
- NerfAcc Plug-and-play NeRF acceleration.
The following repositories greatly inspire threestudio:
Thanks to the maintainers of these projects for their contribution to the community!
If you find threestudio helpful, please consider citing:
@Misc{threestudio2023,
author = {Yuan-Chen Guo and Ying-Tian Liu and Ruizhi Shao and Christian Laforte and Vikram Voleti and Guan Luo and Chia-Hao Chen and Zi-Xin Zou and Chen Wang 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}
}