Latent Diffusion Matters: Human-like Sketches in One-shot Drawing Tasks

1. Download the database

In this article we have used the QuickDraw database (available here) and the Omniglot database (available here )

2. Training RAEs

Once the databse are downloader you can train the RAE using the following command

python train_RAE.py --dataset_root YOUR_DATA_PATH --dataset quickdraw_clust --latent_size 128 --w_kl 0 --w_cl 0 --b_cons 1 --w_cons 0 --w_vq 0 --proto_cons 0 --device cuda:6 

In this script w_kl, w_vq, w_cl, proto_cons, w_cons, b_cons correspond to the beta coefficient for the KL, VQ, classification, prototype-based, SimCLR and Barlow regularizers, respectively.

3. Training Diffusion models

Once you have trained you RAEs, you can run the following script

python train_diffusion.py --dataset_root YOUR_DATA_PATH --dataset quickdraw_clust --ae_path YOUR_AE_PATH --ae_name YOUR_AE_NAME --timestep 1000 --diffuser att_mlp_1d --device cuda:6