This work is a modification of LostGAN-v2, where partwise segmentation masks are combined with the ISLA-Norm at the ResBlocks in the generator.

Results

Comparison of the images generated by LostGAN-v2 and the modified LostGAN

---

Visualization of the partmasks for generated images

---

Quantitative comparison of different layout-to-image synthesis methods on COCO-dataset (128x128 images)

Installation

Check INSTALL.md for installation instructions.

1. Train models

python train.py --dataset coco --out_path outputs/

2. Run pretrained model

python test.py --dataset coco --model_path pretrained_model/G_coco.pth --sample_path samples/coco/

Reference

The code is majorly borrowed from LostGAN-v2.