Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://skhu101.github.io/HumanLiff/

HumanLiff learns layer-wise 3D human with a unified diffusion process.
https://skhu101.github.io/HumanLiff/

Last synced: about 2 months ago
JSON representation

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

Host: GitHub
URL: https://skhu101.github.io/HumanLiff/
Owner: skhu101
License: other
Created: 2023-08-07T15:16:14.000Z (about 1 year ago)
Default Branch: main
Last Pushed: 2024-05-06T14:03:16.000Z (5 months ago)
Last Synced: 2024-05-07T08:56:45.675Z (5 months ago)
Language: Python
Homepage:
Size: 39.9 MB
Stars: 40
Watchers: 5
Forks: 0
Open Issues: 0
Metadata Files:
- Readme: README.md

Awesome Lists containing this project

Awesome-Text-to-3D - HumanLiff: Layer-wise 3D Human Generation with Diffusion Model - L587) | [site](https://skhu101.github.io/HumanLiff/) | [code](https://github.com/skhu101/HumanLiff) (Papers :scroll:)

README

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.

:open_book: 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_.

## :desktop_computer: Requirements

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

```bash
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](https://mycuhk-my.sharepoint.com/:u:/g/personal/1155098117_link_cuhk_edu_hk/EU8blUJ9UKZPj1XCITu2Hw0Bm82y_lWZoENX7qBBXYc7NQ?e=0IdvmG). 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](https://mycuhk-my.sharepoint.com/:u:/g/personal/1155098117_link_cuhk_edu_hk/EawbE2omiahLvSRYVwmC9gYBCo8mfCAXww02JJyk-tj96g?e=RbGb84). The folder structure should look like

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

#### Download Models
Register and download SMPL and SMPLX (version 1.0) models [here](https://smplify.is.tue.mpg.de/download.php). 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
```

## :train: Layer-wise Reconstruction

```bash
cd recon_NeRF
```

### Reconstruction code with Layer-wise SynBody dataset

#### Optimize shared decoder with 100 subjects
```bash
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
```bash
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
```bash
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
```bash
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
```bash
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
```bash
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.

## :train: Layer-wise Generation

```bash
cd ../human_diffusion
```

### Generation code with Layer-wise SynBody dataset

#### Training
```bash
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
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
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
bash triplane_scripts/TightCap_triplane_sample_layered_cond_controlnet_scale_256x256x27_tv_loss_nineplane.sh 1000 27 192 27 0 12889
```

## :newspaper_roll: License

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