sprite-dx-anti-corruption-v1

SpriteDX Anti-Corruption Model v1.1.3

Pixel art restoration model that performs both quantization (denoising/sharpening) and optional background matting.

Model Description

• Architecture: U-Net with residual connections
• Input: 128x128 RGBA (4 channels)
• Output: 128x128 RGBA (4 channels)
• Training: Mixed dataset strategy with 50/50 split between:
  • Background matting task (composited images)
  • Quantization task (transparent images)

Capabilities

1. Pixel Art Quantization: Removes blur, noise, and interpolation artifacts
2. Background Matting: Extracts transparency from composited images
3. Unified Processing: Handles both tasks in a single forward pass

Usage

import torch
from PIL import Image
import numpy as np

# Load model (you'll need the model architecture from train_v1.1.3.py)
from train_v1_1_1 import AntiCorruptionUNet

model = AntiCorruptionUNet()
checkpoint = torch.load("model.safetensors")
model.load_state_dict(checkpoint)
model.eval()

# Process image
image = Image.open("input.png").convert("RGBA")
image_tensor = torch.from_numpy(
    np.array(image, dtype=np.float32).transpose(2, 0, 1) / 255.0
).unsqueeze(0)

with torch.no_grad():
    output = model(image_tensor)
    output = model.post_process(output)

# Convert back to PIL
output_np = output[0].permute(1, 2, 0).numpy()
output_np = np.clip(output_np * 255, 0, 255).astype(np.uint8)
output_image = Image.fromarray(output_np, mode='RGBA')

Training Details

• Corruption pipeline: nearest neighbor upscaling → noise → blur → translation → bicubic downscaling
• Loss: Weighted RGB + Alpha loss with contrast and whiteness weighting
• Data augmentation: Color shifts, flips, grayscale conversion

Limitations

• Input must be 128x128 pixels
• Best results on pixel art and sprites
• May struggle with very complex transparency patterns

License

MIT