💃DreamWaltz: Make a Scene with Complex 3D Animatable Avatars

Project Page | Arxiv | Video

This repository contains the official implementation of NeurIPS 2023 paper:

DreamWaltz: Make a Scene with Complex 3D Animatable Avatars
Yukun Huang^1,2, Jianan Wang¹, Ailing Zeng¹, He Cao¹, Xianbiao Qi¹, Yukai Shi¹, Zheng-Jun Zha², Lei Zhang¹
¹International Digital Economy Academy ²University of Science and Technology of China

Introduction

DreamWaltz is a learning framework for text-driven 3D animatable avatar creation using pretrained 2D diffusion model ControlNet and human parametric model SMPL. The core idea is to optimize a deformable NeRF representation from skeleton-conditioned diffusion supervisions, which ensures 3D consistency and generalization to arbitrary poses.

Figure 1. DreamWaltz can generate animatable avatars (a) and construct complex scenes (b)(c)(d).

Installation

This code is heavily based on the excellent latent-nerf and stable-dreamfusion projects. Please install dependencies:

pip install -r requirements.txt

The cuda extension for instant-ngp is built at runtime as in stable-dreamfusion.

Prepare SMPL Weights

We use smpl and vposer models for avatar creation and animation learning, please follow the instructions in smplx and human_body_prior to download the model weights, and build a directory with the following structure:

smpl_models
├── smpl
│   ├── SMPL_FEMALE.pkl
│   └── SMPL_MALE.pkl
│   └── SMPL_NEUTRAL.pkl
└── vposer
    └── v2.0
        ├── snapshots
        ├── V02_05.yaml
        └── V02_05.log

Then, update the model paths SMPL_ROOT and VPOSER_ROOT in configs/paths.py.

Prepare Motion Sequences

You might need to prepare SMPL-format human motion sequences to animate the generated avatars. Our code provide a data api for AIST++, which is a high-quality dance video database with SMPL annotations. Please download the SMPL annotations from this website, build a directory with the following structure:

aist
├── gWA_sFM_cAll_d26_mWA5_ch13.pkl
├── gWA_sFM_cAll_d27_mWA0_ch15.pkl
├── gWA_sFM_cAll_d27_mWA2_ch17.pkl
└── ...

and update the data path AIST_ROOT in configs/paths.py.

Getting Started

DreamWaltz mainly consists of two training stages: (I) Canonical Avatar Creation and (II) Animatable Avatar Learning.

The following commands are also provided in run.sh.

1. SMPL-Guided NeRF Initialization

To pretrain NeRF using mask images rendered from canonical-posed SMPL mesh:

python train.py \
  --log.exp_name "pretrained" \
  --log.pretrain_only True \
  --prompt.scene canonical-A \
  --prompt.smpl_prompt depth \
  --optim.iters 10000

The obtained pretrained ckpt is available to different text prompts.

2. Canonical Avatar Creation

To learn a NeRF-based canonical avatar representation using ControlNet-based SDS:

text="a wooden robot"
avatar_name="wooden_robot"
pretrained_ckpt="./outputs/pretrained/checkpoints/step_010000.pth"
# the pretrained ckpt is available to different text prompts

python train.py \
  --guide.text "${text}" \
  --log.exp_name "canonical/${avatar_name}" \
  --optim.ckpt "${pretrained_ckpt}" \
  --optim.iters 30000 \
  --prompt.scene canonical-A

3. Animatable Avatar Learning

To learn a NeRF-based animatable avatar representation using ControlNet-based SDS:

text="a wooden robot"
avatar_name="wooden_robot"
canonical_ckpt="./outputs/canonical/${avatar_name}/checkpoints/step_030000.pth"

python train.py \
  --animation True \
  --guide.text "${text}" \
  --log.exp_name "animatable/${avatar_name}" \
  --optim.ckpt "${canonical_ckpt}" \
  --optim.iters 50000 \
  --prompt.scene random \
  --render.cuda_ray False

4. Make a Dancing Video

To make a dancing video based on the well-trained animatable avatar representation and the target motion sequences:

scene="gWA_sFM_cAll_d27_mWA2_ch17,180-280"
# "gWA_sFM_cAll_d27_mWA2_ch17" is the filename of motion sequences in AIST++
# "180-280" is the range of video frame indices: [180, 280]

avatar_name="wooden_robot"
animatable_ckpt="./outputs/animatable/${avatar_name}/checkpoints/step_050000.pth"

python train.py \
    --animation True \
    --log.eval_only True \
    --log.exp_name "videos/${avatar_name}" \
    --optim.ckpt "${animatable_ckpt}" \
    --prompt.scene "${scene}" \
    --render.cuda_ray False \
    --render.eval_fix_camera True

The resulting video can be found in PROJECT_ROOT/outputs/videos/${avatar_name}/results/128x128/.

Results

Canonical Avatars

Figure 2. DreamWaltz can create canonical avatars from textual descriptions.

Animatable Avatars

Figure 3. DreamWaltz can animate canonical avatars given motion sequences.

Complex Scenes

Figure 4. DreamWaltz can make complex 3D scenes with avatar-object interactions.

Figure 5. DreamWaltz can make complex 3D scenes with avatar-scene interactions.

Figure 6. DreamWaltz can make complex 3D scenes with avatar-avatar interactions.

Reference

If you find this repository useful for your work, please consider citing it as follows:

@article{huang2023dreamwaltz,
 title={DreamWaltz: Make a Scene with Complex 3D Animatable Avatars},
 author={Yukun Huang and Jianan Wang and Ailing Zeng and He Cao and Xianbiao Qi and Yukai Shi and Zheng-Jun Zha and Lei Zhang},
 year={2023},
 eprint={2305.12529},
 archivePrefix={arXiv},
 primaryClass={cs.CV}
}

@article{huang2023dreamtime,
 title={DreamTime: An Improved Optimization Strategy for Text-to-3D Content Creation},
 author={Yukun Huang and Jianan Wang and Yukai Shi and Xianbiao Qi and Zheng-Jun Zha and Lei Zhang},
 year={2023},
 eprint={2306.12422},
 archivePrefix={arXiv},
 primaryClass={cs.CV}
}

IDEA-Research/DreamWaltz

IDEA-Research

Reviews

Repository Details