LTOS

Layout-controllable Text-Object Synthesis via Adaptive Cross-attention Fusions

training

run CUDA_VISIBLE_DEVICES=0,1,2 python -m torch.distributed.launch --nproc_per_node=3 --master_port=22411 main.py --yaml_file=configs/text_object_control.yaml --DATA_ROOT=./DATA --batch_size=8 --name=cross_fusion_model --official_ckpt_name="checkpoint_generation_text.pth"

--name: saved filename --official_ckpt_name: pretrained model (you should download the GLIGEN checkpoint from (https://huggingface.co/gligen/gligen-generation-text-box/blob/main/diffusion_pytorch_model.bin))

inference

run python inference.py

dataset_contruction

in data_construction/README.md