Craftify

Generate Minecraft Skin by using DCGAN. I trained over 900,000 skins(dataset: dataset: https://www.kaggle.com/datasets/sha2048/minecraft-skin-dataset/data)

This is my first project in Deep Learning. Forgive my poor code quality and lack of comments :D

I used to train the whole skin file(64x64), but the result is not good. So I only train the head part(8x8) of the skin. The model in pth_release/ folder is trained by the head part.

The dataset is 64x64 skin images, if you only want to train the head part like me, you can execute python3 utils.py after downloading and unziping the dataset and move it to data/skin folder.

|--data
|  |--skin
|     |--Skins

Further work

• Try SAGAN(https://export.arxiv.org/pdf/1805.08318)
• Train the full skin

Demo(Head part)

simple readme :p https://api.soulter.top/craftify

image generate: https://api.soulter.top/craftify/predict

Usage

Arguments

args.add_argument("--dataroot", type=str, default="data/skin/SkinsOnlyHead") # dataset root path
args.add_argument("--type", type=str, default="head") # head or full skin training
args.add_argument("--onlypredict", type=bool, default=False) # only predict. default model is pth_release/t1_g_head.pth
args.add_argument("--gmodelpath", type=str, default="pth_release/t1_g_head.pth") # generator model path

Download Dataset

The dataset is provided by SHA2048. You can download it from here.

And unzip to data/skin/ folder.

Train(Only Head)

python3 crop_head.py
python3 main.py --type head --dataroot data/skin/Skins_head_88

This command will automatically train the model and save the model to pth_release/ folder, and predict Minecraft skin(head) images to output/ folder.

Train(Full Skin Part)

python3 clean_data_full_skin.py
python3 main.py --type full --dataroot data/skin/Skins

This command will automatically train the model and save the model to pth_release/ folder, and predict Minecraft skin(head) images to output/ folder.