skhu101 / HumanLiff

HumanLiff learns layer-wise 3D human with a unified diffusion process.

HumanLiff: Layer-wise 3D Human Generation with Diffusion Model

Shoukang Hu¹  Fangzhou Hong¹  Tao Hu¹  Liang Pan¹  Haiyi Mei²  Weiye Xiao²  Lei Yang²  Ziwei Liu¹ 

¹S-Lab, Nanyang Technological University  ²Sensetime Research

HumanLiff learns layer-wise 3D human with a unified diffusion process.

Figure 1. HumanLiff learns to generate layer-wise 3D human with a unified diffusion process. Starting from a random noise, HumanLiff first generates a human body and then progressively generates 3D humans conditioned on previous generation. We use the same background color to denote generation results from the same human layer.

📖 For more visual results, go checkout our project page

This repository contains the official implementation of HumanLiff: Layer-wise 3D Human Generation with Diffusion Model.

🖥️ Requirements

NVIDIA GPUs are required for this project. We recommend using anaconda to manage python environments.

    conda create --name humanliff python=3.8
    conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
    conda install -c fvcore -c iopath -c conda-forge fvcore iopath
    conda install pytorch3d -c pytorch3d (or pip install --no-index --no-cache-dir pytorch3d -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py38_cu113_pyt1110/download.html)
    pip install -r requirements.txt
    conda activate sherf

Set up Dataset

Layer-wise SynBody Dataset

Please download our rendered multi-view images of layerwise SynBody dataset from OneDrive. The folder structure should look like

./
├── ...
└── recon_NeRF/data/
    ├── SynBody/
        ├── 20230423

Layer-wise TightCap Dataset

Please download our rendered multi-view images of layerwise TightCap dataset from OneDrive. The folder structure should look like

./
├── ...
└── recon_NeRF/data/
    ├── TightCap/

Download Models

Register and download SMPL and SMPLX (version 1.0) models here. Put the downloaded models in the folder smpl_models. Only the neutral one is needed. The folder structure should look like

./
├── ...
└── recon_NeRF/assets/
    ├── SMPL_NEUTRAL.pkl
    ├── models/
        ├── smplx/
            ├── SMPLX_NEUTRAL.npz

./
├── ...
└── human_diffusion/assets/
    ├── SMPL_NEUTRAL.pkl
    ├── models/
        ├── smplx/
            ├── SMPLX_NEUTRAL.npz

🚋 Layer-wise Reconstruction

cd recon_NeRF

Reconstruction code with Layer-wise SynBody dataset

Optimize shared decoder with 100 subjects

python -m torch.distributed.launch --nproc_per_node 4 --master_port 12373 run_nerf_batch.py --config configs/SynBody.txt --data_root data/SynBody/20230423_layered/seq_000000 --expname SynBody_185_view_100_subject_triplane_256x256x27_tv_loss_1e-2_l1_loss_5e-4 --num_instance 100 --num_worker 3 --i_weights 50000 --i_testset 5000 --mlp_num 2 --batch_size 2 --n_samples 128 --n_importance 128 --views_num 185 --use_clamp --ddp 1 --lrate 5e-3 --tri_plane_lrate 1e-1 --triplane_dim 256 --triplane_ch 27 --tv_loss --tv_loss_coef 1e-2 --l1_loss_coef 5e-4

Optimize tri_plane parameters with fixed shared decoder for each subjec

python -m torch.distributed.launch --nproc_per_node 1 --master_port 12373 run_nerf_batch_ft.py --config configs/SynBody.txt --data_root data/SynBody/20230423_layered/seq_000000 --expname SynBody_185_view_100_subject_triplane_256x256x27_tv_loss_1e-2_l1_loss_5e-4 --num_instance 1 --num_worker 3 --i_weights 2000 --i_testset 8000 --mlp_num 2 --batch_size 8 --n_samples 128 --n_importance 128 --views_num 185 --use_clamp --ddp 1 --lrate 0 --tri_plane_lrate 1e-1 --ft_triplane_only --n_iteration 2000 --ft_path 200000.tar --multi_person 0 --start_idx 0 --end_idx 10 --triplane_dim 256 --triplane_ch 27 --start_dim 256 --tv_loss --tv_loss_coef 1e-2 --l1_loss_coef 5e-4

where start_idx and end_idx denote the start index and end index of subjects to be optimized.

Inference code

python -m torch.distributed.launch --nproc_per_node 1 --master_port 12373 run_nerf_batch.py --config configs/SynBody.txt --data_root data/SynBody/20230423_layered/seq_000000 --expname SynBody_185_view_100_subject_triplane_256x256x27_tv_loss_1e-2_l1_loss_5e-4 --num_instance 100 --num_worker 3 --i_weights 50000 --i_testset 5000 --mlp_num 2 --batch_size 1 --n_samples 128 --n_importance 128 --views_num 185 --use_clamp --ddp 1 --lrate 5e-3 --tri_plane_lrate 1e-1 --triplane_dim 256 --triplane_ch 27 --tv_loss --tv_loss_coef 1e-2 --l1_loss_coef 5e-4 --test --ft_path 200000.tar --test_layer_id 1

where test_layer_id denotes the layer index.

Reconstruction code with Layer-wise TightCap dataset

Optimize shared decoder with 100 subjects

python -m torch.distributed.launch --nproc_per_node 4 --master_port 12373 run_nerf_batch.py --config configs/TightCap.txt --data_root data/TightCap/f_c_10412256613 --expname TightCap_185_view_100_subject_triplane_256x256x27_tv_loss_1e-2_l1_loss_5e-4 --num_instance 100 --num_worker 3 --i_weights 50000 --i_testset 5000 --mlp_num 2 --batch_size 2 --n_samples 128 --n_importance 128 --views_num 185 --use_clamp --ddp 1 --use_canonical_space --lrate 5e-3 --tri_plane_lrate 1e-1 --triplane_dim 256 --triplane_ch 27 --tv_loss --tv_loss_coef 1e-2 --l1_loss_coef 5e-4

Optimize tri_plane parameters with fixed shared decoder for each subject

python -m torch.distributed.launch --nproc_per_node 1 --master_port 12373 run_nerf_batch_ft.py --config configs/TightCap.txt --data_root data/TightCap/f_c_10412256613 --expname TightCap_185_view_100_subject_triplane_256x256x27_tv_loss_1e-2_l1_loss_5e-4 --num_instance 1 --num_worker 3 --i_weights 2000 --i_testset 5000 --mlp_num 2 --batch_size 8 --n_samples 128 --n_importance 128 --views_num 185 --use_clamp --ddp 1 --use_canonical_space --lrate 0 --tri_plane_lrate 1e-1 --ft_triplane_only --n_iteration 2000  --ft_path 200000.tar --multi_person 0 --start_idx 0 --end_idx 10 --triplane_dim 256 --triplane_ch 27 --start_dim 256 --tv_loss --tv_loss_coef 1e-2 --l1_loss_coef 5e-4

where start_idx and end_idx denote the start index and end index of subjects to be optimized.

Inference code

python -m torch.distributed.launch --nproc_per_node 1 --master_port 12373 run_nerf_batch.py --config configs/TightCap.txt --data_root data/TightCap/f_c_10412256613 --expname TightCap_185_view_100_subject_triplane_256x256x27_tv_loss_1e-2_l1_loss_5e-4 --num_instance 100 --num_worker 3 --i_weights 50000 --i_testset 5000 --mlp_num 2 --batch_size 1 --n_samples 128 --n_importance 128 --views_num 185 --use_clamp --ddp 1 --use_canonical_space --lrate 5e-3 --tri_plane_lrate 1e-1 --triplane_dim 256 --triplane_ch 27 --tv_loss --tv_loss_coef 1e-2 --l1_loss_coef 5e-4 --test --ft_path 200000.tar --test_layer_id 1

where test_layer_id denotes the layer index.

🚋 Layer-wise Generation

cd ../human_diffusion

Generation code with Layer-wise SynBody dataset

Training

bash triplane_scripts/SynBody_triplane_train_layered_cond_controlnet_scale_256x256x27_tv_loss_nineplane.sh 1000 4 27 192 27 8 1000 200000

Inference

bash triplane_scripts/SynBody_triplane_sample_layered_cond_controlnet_scale_256x256x27_tv_loss_nineplane.sh 1000 27 192 27 0 12889

Generation code with Layer-wise TightCap dataset

Training

bash triplane_scripts/TightCap_triplane_train_layered_cond_controlnet_scale_256x256x27_tv_loss_nineplane.sh 1000 4 27 192 27 8 107 200000

Inference

bash triplane_scripts/TightCap_triplane_sample_layered_cond_controlnet_scale_256x256x27_tv_loss_nineplane.sh 1000 27 192 27 0 12889

🗞️ License

Distributed under the S-Lab License. See LICENSE for more information.