resnet-lama-oxford-weights

🌸 Paint-Paint: Domain-Adaptive Image Inpainting Weights

This repository hosts the pre-trained model weights for the Paint-Paint research project, which investigates the efficacy of Fast Fourier Convolutions (FFCs) and adversarial fine-tuning for domain-adaptive image inpainting in data-scarce environments.

These weights are designed to be used in conjunction with the primary paint-paint GitHub repository.

---

📦 Model Artifacts

This repository contains three distinct model weights, representing the progression of the ablation study and the final adversarial pipeline:

• BaseLine_UNet_Master.pth
  • Architecture: Custom U-Net with Gated Convolutions.
  • Role: The Control Variable. Trained entirely from scratch to establish a baseline for spatial reconstruction and feature mapping without prior geometric knowledge.
• ResNetLama_Master.pth
  • Architecture: Hybrid ResNet-18 Encoder + Custom Fast Fourier Convolution (FFC) Bottleneck + Transposed Convolution Decoder.
  • Role: The Pre-Trained Prior. Trained in pure FP32 precision to prevent frequency-domain corruption. Optimized strictly using L1 (Mean Absolute Error) and High Receptive Field Perceptual Loss (HRFPL) via a frozen VGG-19 network. Exhibits strong spatial geometry but inherent textural oversmoothing due to the L1 objective.
• GAN_Generator_Final.pth
  • Architecture: Hybrid ResNet-LaMa (Generator).
  • Role: The Production Model. This model underwent adversarial fine-tuning against an advanced Spectral Dilated PatchGAN Discriminator. The adversarial objective successfully overcomes the L1 oversmoothing limitation, forcing the synthesis of high-frequency, photorealistic domain textures.

---

🛠️ Usage & Integration

To maintain a self-contained and reproducible environment, it is highly recommended to use the official huggingface_hub API for retrieval. Standard CLI tools like wget often fail to resolve Git Large File Storage (LFS) redirects, resulting in corrupted 1KB text pointers.

Automated Retrieval (Python)

import os
from huggingface_hub import hf_hub_download

# Ensure the target directory exists
os.makedirs("models/saved_weights", exist_ok=True)

# Download the final GAN weights directly from this Hugging Face repository
hf_hub_download(
    repo_id="Binamra00/resnet-lama-oxford-weights",
    filename="GAN_Generator_Final.pth",
    local_dir="models/saved_weights"
)