We’re developing an AI-powered sky replacement system for our photography platform and need an expert in Computer Vision / Deep Learning to build and train a custom model for realistic sky replacement. The system must detect and replace sky regions automatically while preserving trees, people, and buildings — producing professional, natural-looking results with accurate lighting and color balance. --- Project Scope Input: Standard JPG/PNG photos Output: Realistic photo with replaced sky Core Pipeline: Sky segmentation using U²-Net, MODNet, or similar Foreground preservation (people, trees, etc.) Seamless blending (Poisson or multi-scale blending) Lighting/tone matching (histogram or GAN-based) Edge refinement (guided/bilateral filters) Batch inference with GPU support --- Custom Training We’ll provide ~2000 images (outdoor/indoor scenes). Freelancer must: Create sky masks (white = sky, black = others) — manually or semi-automatically Fine-tune/train the model on our dataset Deliver trained weights, inference scripts, and accuracy results --- Requirements Strong knowledge of PyTorch / TensorFlow, OpenCV, image segmentation, matting, and blending Experience with U²-Net / MODNet / UNet architectures Preferably able to deploy via FastAPI --- Deliverables Python source code + API endpoint Trained model + weights Example before/after images Documentation --- Timeline: 2–3 weeks