PAL4VST

Project Page | Paper | Bibtex | Colab | Hugging Face

Perceptual Artifacts Localization for Image Synthesis Tasks
International Conference on Computer Vision (ICCV), 2023
Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, Jianbo Shi

This paper presents a study of Perceptual Artifacts Localization on multiple synthesis tasks.

Prerequisites

• Linux
• Python 3
• NVIDIA GPU + CUDA CuDNN

Table of Contents:

1. Setup - download pretrained models and resources
2. Test Images - quick usage with torchscript
3. Quick Usage - quick usage with torchscript
4. Datasets - download our train/val/test artifacts datasets (coming soon)
5. Checkpoints - download the checkpoints for all our models (coming soon)
6. Inference - inference with models/data (coming soon)
7. Training - training scripts (coming soon)

Setup

• Clone this repo:

git clone https://github.com/owenzlz/PAL4VST

• Install dependencies:

conda create --name pal4vst python=3.8 -y
conda activate pal4vst
pip install torch torchvision
pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"
pip install timm==0.6.5
pip install scikit-image
pip install -U openmim && mim install "mmpretrain>=1.0.0rc8"
pip install mmdeploy==1.3.0
cd mmsegmentation
pip install -v -e .

pip install transformers
pip install --upgrade diffusers[torch]

For more information, please feel free to refer to MMSegmentation: https://mmsegmentation.readthedocs.io/en/latest/

Test images

We randomly sampled 10 test images per synthesis tasks, in case you want to quickly try it out. Please check "./demo_test_data". The conditional inputs (i.e. prompt/edge/etc.) are included if any.

- demo_test_data
    |- anyresgan
    |- cvton
    |- edge2image
    |- stylegan2_ffqh
    |- ...

Quick Usage

Download torchscript checkpoint (swin-large_upernet_unified_512x512), and place it under ./deployment/pal4vst/swin-large_upernet_unified_512x512/.

• A snippet of inference on a single image

from utils import *
from PIL import Image
import numpy as np 
import torch

device = 0
torchscript_file = './deployment/pal4vst/swin-large_upernet_unified_512x512/end2end.pt'
img_file = './demo_test_data/stylegan2_ffhq/images/seed0417.jpg'

model = torch.load(torchscript_file).to(device)
img_tensor = prepare_input(np.array(Image.open(img_file).resize((512, 512))), device)
pal = model(img_tensor).cpu().data.numpy()[0][0] # prediction: Perceptual Artifacts Localization (PAL)

• Alternatively, quick inference with torchscript

python test_torchscript.py \
       --img_file ./demo_test_data/stylegan2_ffhq/images/seed0417.jpg \
       --torchscript_file ./deployment/pal4vst/swin-large_upernet_unified_512x512/end2end.pt \
       --out_pal_file pal.png \
       --out_vis_file img_with_pal.jpg

• Test over all images in demo_test_data folder

python test_torchscript_all.py

Note: This checkpoint is trained on the unified data, and produces reasonable predictions to all synthesis tasks, except for shadow removal which requires a specialist model as reported in the paper. You will need a Shadow specialist checkpoint. Specialist checkpoints can be downloaded below.

Checkpoints

We only released the 'unified' and 'shadow removal' checkpoints, due to extensive number of models. As discussed in the paper, this 'unified' has similar performance to all specialist models on individual tasks, except for 'shadow removal' task. If you need other specialist checkpoints, please reach out to me (zlz@seas.upenn.edu).

Model Name	Torchscript	Pytorch
unified	unified_torchscript	unified_pytorch
stylegan2	stylegan2_torchscript	stylegan2_pytorch
shadow removal	shadow_torchscript	shadow_pytorch

Application 1 - Artifacts Refinement

We implemented SD-XL inpainter to refine artifacts detected by our PAL model.

Refine artifacts (PAL) with SD-XL inpainter

python refine_artifacts.py \
       --torchscript_file ./deployment/pal4vst/swin-large_upernet_unified_512x512/end2end.pt \
       --img_file ./demo_test_data/mask2image/images/000000483531.jpg \
       --out_refine_file ./demo_results/refine.jpg

Note: In the paper, we tried DALL-E 2 inpainter, which gives better results. However, since DALL-E 2 is not free, we offer SD-XL as an alternative option here.

Application 2 - Image Curation

Given a folder of generated images, our PAL model can rank their quality using the area of detected artifacts region.

Rank the quality for a set of images.

python curate_images.py \
       --torchscript_file ./deployment/pal4vst/swin-large_upernet_unified_512x512/end2end.pt \
       --input_img_dir ./demo_test_data/stylegan2_ffhq/images \
       --rank_img_dir ./demo_results/stylegan2_ffhq_rank

Datasets

The dataset for each synthesis tasks can be downloaded here: specific_tasks.zip. Altenatively, you can download the merged dataset from all tasks here: unified_task.zip.

Training

An examplar training command line (multi-GPU training)

./tools/dist_train.sh configs/pal4vst/swin-large_upernet_unified_512x512.py 8 --work-dir work_dirs/pal4vst/swin-large_upernet_unified_512x512

Note: All checkpoints are trained in MMSeg 0x codebase. We haven't retrained/reproduced the models after the codebase update (MMSeg 0x → MMSeg 1x).

Testing

Inference using Pytorch model (which should have similar performance to Torchscript model).

python test.py \
       --config_file work_dirs/pal4vst/convnext-large_upernet_unified1_512x512/convnext-large_upernet_unified1_512x512.py \
       --checkpoint_file work_dirs/pal4vst/convnext-large_upernet_unified1_512x512/best_mIoU_iter_6500.pth \
       --img_dir ../data/pal4vst/demo_test_data/stylegan2_ffhq/images \
       --seg_dir ../data/pal4vst/demo_test_data/stylegan2_ffhq/images_with_pal

Citation

If you use this code for your research, please cite our paper:

@InProceedings{Zhang_2023_ICCV,
    author    = {Zhang, Lingzhi and Xu, Zhengjie and Barnes, Connelly and Zhou, Yuqian and Liu, Qing and Zhang, He and Amirghodsi, Sohrab and Lin, Zhe and Shechtman, Eli and Shi, Jianbo},
    title     = {Perceptual Artifacts Localization for Image Synthesis Tasks},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {7579-7590}
}

This work is inspired by our previous ECCV 22' paper:

@inproceedings{zhang2022perceptual,
  title={Perceptual artifacts localization for inpainting},
  author={Zhang, Lingzhi and Zhou, Yuqian and Barnes, Connelly and Amirghodsi, Sohrab and Lin, Zhe and Shechtman, Eli and Shi, Jianbo},
  booktitle={European Conference on Computer Vision},
  pages={146--164},
  year={2022},
  organization={Springer}
}

Acknowledgement

We thank the authors who open-sourced the research codes/data, which enables us to generate the images.

StyleGAN: https://github.com/NVlabs/stylegan3
StyleGAN-Human: https://github.com/stylegan-human/StyleGAN-Human
Latent Diffusion Models: https://github.com/CompVis/latent-diffusion
Anyresgan: https://github.com/chail/anyres-gan
Real-ESRGAN: https://github.com/xinntao/Real-ESRGAN
PITI: https://github.com/PITI-Synthesis/PITI
DALL-E 2: https://openai.com/product
Stable Diffusion: https://github.com/huggingface/diffusers
Latent Composition: https://github.com/chail/latent-composition
CVTon: https://github.com/benquick123/C-VTON
Portrait Shadow: https://ceciliavision.github.io/project-pages/portrait
LaMa: https://github.com/advimman/lama
CoMod-GAN: https://github.com/zsyzzsoft/co-mod-gan
BlobGAN: https://github.com/dave-epstein/blobgan
DiT: https://github.com/facebookresearch/DiT
Verstile Diffusion: https://github.com/SHI-Labs/Versatile-Diffusion