Pixel-Art Sprite Generation & Restoration DCGAN + U-Net Autoencoder Author: Deivid Ribeiro · 2025

End-to-end pipeline to

1. Generate 64 × 64 px pixel-art sprites with a DCGAN
2. Restore blurry sprites with a U-Net auto-encoder
3. Assemble ready-to-use sprite-sheets for any 2-D game engine

Main Scripts

Script	Purpose
DCGAN_Deivid_Ribeiro_v0.4.py	Train the DCGAN and save 8 × 8 sample grids
AutoEncoders_Deivid_Ribeiro_v0.4.py	Train the U-Net auto-encoder (2 phases)
FinalPipeline_ApplyAutoencoder_v0.4.py	Restore GAN sprites, build triplets & sprite-sheets

Prerequisites

Python 3.10+ and the libraries below:

torch torchvision opencv-python pillow matplotlib tqdm

torch>=2.1.0 torchvision>=0.16.0 pillow>=9.5.0 opencv-python>=4.8.0 matplotlib>=3.7.1 tqdm>=4.66.1

Folder Layout

• data/ – real sprites for training   • 0/, 1/, 2/, 3/ – four poses (ImageFolder format, ≈ 3 648 PNGs)

• gan_aux/0/ – optional extra GAN sprites used to fine-tune the Auto-encoder

• runs/ – auto-generated by the training scripts   • YYYYMMDD_HHMMSS/ – DCGAN checkpoints + sample grids   • YYYYMMDD_HHMMSS_ae/ – Auto-encoder checkpoints + training plots

• sprites_split/ – 64 PNGs (64 × 64) cut from one DCGAN grid   • sprite_00.png … sprite_63.png

• gan_restored/ – final outputs   • sprite_00_triplet.png … sprite_63_triplet.png (original | blurred | restored)   • spritesheet_original.png – 8 × 8 originals   • spritesheet_borrada.png – 8 × 8 blurred   • spritesheet_restaurada.png – 8 × 8 restored   • spritesheet_tripla.png – 24 × 8 mega-sheet

• Scripts   • DCGAN_Deivid_Ribeiro_v0.4.py   • AutoEncoders_Deivid_Ribeiro_v0.4.py   • FinalPipeline_ApplyAutoencoder_v0.4.py

If your folders differ, edit DATA_DIR, GAN_IMG_DIR, etc. at the top of each script.

Step-by-Step

1 – Train the DCGAN

python DCGAN_Deivid_Ribeiro_v0.4.py

Input : data/0-3/
Output: runs/…/samples_epoch_XXX.png

Cut one 8 × 8 sample grid into 64 PNGs and drop them in sprites_split/ (any sprite-sheet slicer will do).

2 – Train the U-Net Auto-encoder

python AutoEncoders_Deivid_Ribeiro_v0.4.py

Phase 1 : real sprites · Phase 2 (optional) : gan_aux/0/
Output : runs/…_ae/autoencoder.pth + 3 training plots

3 – Restore sprites & build sprite-sheets

python FinalPipeline_ApplyAutoencoder_v0.4.py

Input : • sprites_split/ (64 PNG originals)
• autoencoder.pth from runs/…_ae/

What it does :

1. Creates a blurred copy (for comparison only)
2. Restores each sprite with the AE
3. Saves original | blurred | restored as sprite_NN_triplet.png
4. Auto-builds four sheets:
   spritesheet_original.png, spritesheet_borrada.png,
   spritesheet_restaurada.png, spritesheet_tripla.png

Tips & Tweaks

• All hyper-parameters live at the top of each script (EPOCHS, LR, …)
• CUDA is used automatically if available (DEVICE = "cuda").
• Increase EPOCHS (DCGAN) or EPOCHS_FT (U-Net) for higher quality.
• PNG sheets import directly into most retro-game engines (rows contiguous).

Happy hacking! 😊