SMILE: Semantically-guided Multi-attribute Image and Layout Editing, ICCVW 2021.

Official PyTorch Implementation

[Paper 📰]   [Video 📹]   [Poster 📜]   [Slides 📌]

✨ Results

SMILE can manipulate a source image into ab output image reflecting the attribute and style (e.g., eyeglasses, hat, hair, etc.) of a different person. More high-quality videos can be found in this link.

Checkout the project page for additional visualizations.

Overview of the method

🔧 Download Pretrained Weights

bash download_weights.sh

⚡ Demo

python main.py --GPU=NO_CUDA --FAN --EYEGLASSES --GENDER --EARRINGS --HAT --BANGS --HAIR --TRAIN_MASK --MOD --SPLIT_STYLE --mode=demo --ref_demo Figures/ffhq_teaser --rgb_demo Figures/teaser_input.png --pretrained_model models/pretrained_models/smileSEM

This command should reproduce the teaser figure. Explanation of arguments:

• --FAN: Remove all shortcuts in the upsampling residual blocks and add skip connections with the adaptive wing based heatmap.
• --EYEGLASSES --GENDER --EARRINGS --HAT --BANGS --HAIR: The selected attributes to manipulate.
• --TRAIN_MASK: Use semantic maps instead of RGB.
• --MOD: Use modulated convolutions.
• --SPLIT_STYLE: Weight Gender with more dimensionality than the others.
• --ref_demo: Folder with reference images. During demo an attribute classifier is going to extract every attribute to impose them on the --rgb_demo. See teaser figure.

🌏 Citation

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

@InProceedings{Romero_2021_ICCV,
    author    = {Romero, Andres and Van Gool, Luc and Timofte, Radu},
    title     = {SMILE: Semantically-Guided Multi-Attribute Image and Layout Editing},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
    month     = {October},
    year      = {2021},
    pages     = {1924-1933}
}