Official Implementation of "Layout Free scene graph to image generation", accepted at the British Machine Vision Conference (BMVC), 2024, Glasgow.

For visual genome

accelerate launch SG_to_img_vg.py --pretrained_model_name_or_path="CompVis/stable-diffusion-v1-4" --output_dir=./Output_vg_concat --resolution=256 --train_batch_size=2 --gradient_accumulation_steps=1 --learning_rate=5e-6 --lr_scheduler="constant" --lr_warmup_steps=0 --num_train_epochs=1

For coco stuff

accelerate launch SG_to_img_coco.py --pretrained_model_name_or_path="CompVis/stable-diffusion-v1-4" --output_dir=./Output_coco_mmd --resolution=256 --train_batch_size=2 --gradient_accumulation_steps=1 --learning_rate=5e-6 --lr_scheduler="constant" --lr_warmup_steps=0 --num_train_epochs=1

Note:

1. The conditioning part of the code will be updated soon.
2. The updated environment file withoud dependency issues will be uploaded soon.