Pytorch implementation of few-shot semantic image synthesis

Last update: Sep 26, 2022

Related tags

Overview

Few-shot Semantic Image Synthesis Using StyleGAN Prior

Our method can synthesize photorealistic images from dense or sparse semantic annotations using a few training pairs and a pre-trained StyleGAN.

Prerequisites

Python3
PyTorch

Preparation

Download and decompress the file containing StyleGAN pre-trained models and put the "pretrained_models" directory in the parent directory.

Inference with our pre-trained models

Download and decompress the file containing our pretrained encoders and put the "results" directory in the parent directory.
For example, our results for celebaMaskHQ in a one-shot setting can be generated as follows:

python scripts/inference.py --exp_dir=results/celebaMaskHQ_oneshot --checkpoint_path=results/celebaMaskHQ_oneshot/checkpoints/iteration_100000.pt --data_path=./data/CelebAMask-HQ/test/labels/ --couple_outputs --latent_mask=8,9,10,11,12,13,14,15,16,17

Inference results are generated in results/celebaMaskHQ_oneshot. If you use other datasets, please specify --exp_dir, --checkpoint_path, and --data_path appropriately.

Training

For each dataset, you can train an encoder as follows:

CelebAMask

python scripts/train.py --exp_dir=[result_dir] --dataset_type=celebs_seg_to_face --stylegan_weights pretrained_models/stylegan2-ffhq-config-f.pt --start_from_latent_avg --label_nc 19 --input_nc 19

CelebALandmark

python scripts/train.py --exp_dir=[result_dir] --dataset_type=celebs_landmark_to_face --stylegan_weights pretrained_models/stylegan2-ffhq-config-f.pt --start_from_latent_avg --label_nc 71 --input_nc 71 --sparse_labeling

Intermediate training outputs with the StyleGAN pre-trained with the CelebA-HQ dataset. It can be seen that the layouts of the bottom-row images reconstructed from the middle-row pseudo semantic masks gradually become close to those of the top-row StyleGAN samples as the training iterations increase.

LSUN church

python scripts/train.py --exp_dir=[result_dir] --dataset_type=lsunchurch_seg_to_img --stylegan_weights pretrained_models/stylegan2-church-config-f.pt --style_num 14 --start_from_latent_avg --label_nc 151 --input_nc 151

LSUN car

python scripts/train.py --exp_dir=[result_dir] --dataset_type=lsuncar_seg_to_img --stylegan_weights pretrained_models/stylegan2-car-config-f.pt --style_num 16 --start_from_latent_avg --label_nc 5 --input_nc 5

LSUN cat

python scripts/train.py --exp_dir=[result_dir] --dataset_type=lsuncat_scribble_to_img --stylegan_weights pretrained_models/stylegan2-cat-config-f.pt --style_num 14 --start_from_latent_avg --label_nc 9 --input_nc 9 --sparse_labeling

Ukiyo-e

python scripts/train.py --exp_dir=[result_dir] --dataset_type=ukiyo-e_scribble_to_img --stylegan_weights pretrained_models/ukiyoe-256-slim-diffAug-002789.pt --style_num 14 --channel_multiplier 1 --start_from_latent_avg --label_nc 8 --input_nc 8 --sparse_labeling

Anime

python scripts/train.py --exp_dir=[result_dir] --dataset_type=anime_cross_to_img --stylegan_weights pretrained_models/2020-01-11-skylion-stylegan2-animeportraits-networksnapshot-024664.pt --style_num 16 --start_from_latent_avg --label_nc 2 --input_nc 2 --sparse_labeling

Using StyleGAN samples as few-shot training data

Run the following script:

python scripts/generate_stylegan_samples.py --exp_dir=[result_dir] --stylegan_weights ./pretrained_models/stylegan2-ffhq-config-f.pt --style_num 18 --channel_multiplier 2

Then a StyleGAN image (*.png) and a corresponding latent code (*.pt) are obtained in [result_dir]/data/images and [result_dir]/checkpoints.

Manually annotate the generated image in [result_dir]/data/images and save the annotated mask in [result_dir]/data/labels.
Edit ./config/data_configs.py and ./config/paths_config.py appropriately to use the annotated pairs as a training set.
Run a training command above with appropriate options.

Citation

Please cite our paper if you find the code useful:

@article{endo2021fewshotsmis,
  title = {Few-shot Semantic Image Synthesis Using StyleGAN Prior},
  author = {Yuki Endo and Yoshihiro Kanamori},
  journal   = {CoRR},
  volume    = {abs/2103.14877},
  year      = {2021}
}

Acknowledgements

This code heavily borrows from the pixel2style2pixel repository.

Pytorch implementation of few-shot semantic image synthesis

Related tags

Overview

Few-shot Semantic Image Synthesis Using StyleGAN Prior

Prerequisites

Preparation

Inference with our pre-trained models

Training

Using StyleGAN samples as few-shot training data

Citation

Acknowledgements

Owner

A hybrid framework (neural mass model + ML) for SC-to-FC prediction

[SIGGRAPH 2021 Asia] DeepVecFont: Synthesizing High-quality Vector Fonts via Dual-modality Learning

ChatBot-Pytorch - A GPT-2 ChatBot implemented using Pytorch and Huggingface-transformers

[ICLR'19] Trellis Networks for Sequence Modeling

"Learning and Analyzing Generation Order for Undirected Sequence Models" in Findings of EMNLP, 2021

Python script for performing depth completion from sparse depth and rgb images using the msg_chn_wacv20. model in Tensorflow Lite.

Efficient Training of Visual Transformers with Small Datasets

Creating Artificial Life with Reinforcement Learning

TensorFlow Implementation of Unsupervised Cross-Domain Image Generation

FairMOT for Multi-Class MOT using YOLOX as Detector

Show-attend-and-tell - TensorFlow Implementation of "Show, Attend and Tell"

DSTC10 Track 2 - Knowledge-grounded Task-oriented Dialogue Modeling on Spoken Conversations

DTCN SMP Challenge - Sequential prediction learning framework and algorithm

Keras implementation of Normalizer-Free Networks and SGD - Adaptive Gradient Clipping

Website for D2C paper

Official implementation of "Learning Proposals for Practical Energy-Based Regression", 2021.

SigOpt wrappers for scikit-learn methods

[ICCV 2021] Official Tensorflow Implementation for "Single Image Defocus Deblurring Using Kernel-Sharing Parallel Atrous Convolutions"

Playable Video Generation

Library extending Jupyter notebooks to integrate with Apache TinkerPop and RDF SPARQL.