[Arxiv] | [Project Page]

This repository is an official implementation for:

SCP-Diff: Photo-Realistic Semantic Image Synthesis with Spatial-Categorical Joint Prior

Authors: Huan-ang Gao*, Mingju Gao*, Jiaju Li, Wenyi Li, Rong zhi, Hao Tang, Hao Zhao

Semantic image synthesis (SIS) shows promising potential for sensor simulation. However, current best practices in this field, based on GANs, have not yet reached the desired level of quality. As latent diffusion models make significant strides in image generation, we are prompted to evaluate ControlNet, a notable method for its dense control capabilities. Our investigation uncovered two primary issues with its results: the presence of weird sub-structures within large semantic areas and the misalignment of content with the semantic mask. Through empirical study, we pinpointed the cause of these problems as a mismatch between the noised training data distribution and the standard normal prior applied at the inference stage. To address this challenge, we developed specific noise priors for SIS, encompassing spatial, categorical, and an innovative spatial-categorical joint prior for inference. This approach, which we have named SCP-Diff, has yielded exceptional results, achieving an FID of 10.53 on Cityscapes and 12.66 on ADE20K.

Use conda to create a new virtual enviroment. We use torch==2.2.0+cu121.

conda env create -f environment.yaml
conda activate scp-diff

We have prepared dataloaders for Cityscapes, ADE20K, and COCO-Stuff datasets in the directory datasets/.

You can download cityscapes dataset from https://www.cityscapes-dataset.com/.

The file organization of the Cityscapes dataset is as follows:

├── cityscapes
│   ├── leftImg8bit
│   │   ├── train
│   │   ├── val
│   │   ├── test
│   ├── gtFine
│   │   ├── train
│   │   ├── val
│   │   ├── test
...

class CityscapesDataset(Dataset):
    RESOLUTION = (512, 1024)
    BASE = '/path/to/your/cityscapes/dataset'

    def __init__(self):
        super().__init__()
    ...

You need to set the root directory of the cityspaes dataset and specify the resolution for generated images by modifying the BASE and RESOLUTION. We use 512*1024 as default.

You can download ade20k dataset from https://groups.csail.mit.edu/vision/datasets/ADE20K/.

The file organization of the ade20k dataset is as follows:

├── ade20k
│   ├── annotations
│   │   ├── training # Gray-level annotations
│   │   ├── ...
│   ├── images
│   │   ├── training # RGB Images
│   │   ├── ...
...

class ADE20KDataset(Dataset):
    BASE = '/path/to/your/ade20k/dataset'
    RESOLUTION = (512, 512) 

    def __init__(self):
        super().__init__()

You need to set the root directory of the ade20k dataset and specify the resolution for generated images by modifying the BASE and RESOLUTION. We use 512*512 as default.

You can download coco-stuff dataset from https://github.com/nightrome/cocostuff

The file organization of the coc-stuff dataset is as follows:

├── coco-stuff
│   ├── annotations
│   │   ├── train2017 # Gray-scale annotations
│   │   ├── ...
│   ├── images
│   │   ├── train2017 # RGB Images
│   │   ├── ...
...

class CocostuffDataset(Dataset):
    BASE = '/path/to/your/coco-stuff/dataset/'
    RESOLUTION = (384, 512) 

    def __init__(self):

You need to set the root directory of the coco-stuff dataset and specify the resolution for generated images by modifying the BASE and RESOLUTION. We use 384*512 as default.

You can custom your own dataset and write the dataloader following datasets/*.py You should have the original images, the gray-scale semantic maps and color-coded semantic maps. See examples from the above three datasets.

We provide a download link for our finetuned ControlNets and reduced statistics in our [huggingface] page.

For inference, we need to use categorial and spatial prior. You can run noise_prior_inference.py:

CUDA_VISIBLE_DEVICES=gpu_ids \
    python3 noise_prior_inference.py --dataset cityscapes/ade20k/coco-stuff --sample_size <inference samples> --diffusion_steps <ddpm steps> --seed 4 --save_dir /path/to/save/infer/results \
    --batch_size <batch size> \
    --resolution <h> <w> --ckpt /path/to/ckpt \
    --stat_dir /path/to/statistics --stat_name <save_name of statistics> --scale <guidance scale>

where sample_size denotes the total sample size, ddpm steps ranges from 0 to 999, resolution should be compatible with the ControlNet finetuning process.

For classifier-free guidance scale, we recommend you to set the scale to 4.0 for cityscapes dataset, and 7.5 for ade20k and cocostuff dataset.

You can refer to noise_prior_inference.py for more details.

Run tutorial_train.py for training.

python tutorial_train.py --batch_size <batch_size> --dataset <your dataset> --default_root_dir <log_root_dir> --gpus gpu_ids --resume_path <resume file path>

We can aggregate gather statistical information on the dataset's categorical and spatial information.

We can run noise_prior.py:

CUDA_VISIBLE_DEVICES=gpu_ids \
    python3 noise_prior.py --dataset ade20k --sample_size 10000 --save_name ade20k_10000 \
    --ckpt /path/to/ckpt --resolution 512 512 --save_dir ./statistics

You can refer to noise_prior.py for more details.

If you find this work useful for your research, please cite our paper:

@article{gao2024scp,
  title={SCP-Diff: Photo-Realistic Semantic Image Synthesis with Spatial-Categorical Joint Prior},
  author={Gao, Huan-ang and Gao, Mingju and Li, Jiaju and Li, Wenyi and Zhi, Rong and Tang, Hao and Zhao, Hao},
  journal={arXiv preprint arXiv:2403.09638},
  year={2024}
}

We build our codebase on ControlNet, a neural network structure to control diffusion models by adding extra conditions.