SatGen-CV
Semantic Image Synthesis for Satellite Imagery from Road and Building Segmentation Masks
A Pix2Pix-based model for generating realistic satellite imagery from semantic segmentation masks containing roads and buildings.
Model Description
- Architecture: U-Net generator (~54.4M parameters) with PatchGAN discriminator
- Input: 256x256x2 (road + building segmentation masks)
- Output: 256x256x3 (RGB satellite image)
- Training: Massachusetts Roads Dataset + Inria Aerial Image Labeling Dataset
Usage
# Download the model
huggingface-cli download Guereak/SatGen-CV --local-dir ./checkpoints
import torch
from src.models.generator import Generator
# Load checkpoint
checkpoint = torch.load("checkpoints/satgen-model.pth", map_location="cpu")
config = checkpoint["config"]
# Initialize generator
generator = Generator(
in_channels=config["input_channels"],
out_channels=config["output_channels"],
features=config["generator_features"]
)
generator.load_state_dict(checkpoint["generator_state_dict"])
generator.eval()