FaithDiff addresses faithful image super-resolution and restoration for applications like classic film rejuvenation, old photo revival, and social media enhancement. Targeting low-level vision researchers, it leverages diffusion priors for high-quality results on degraded imagery.

How It Works

This CVPR 2025 project implements FaithDiff for faithful image super-resolution, integrated within Hugging Face's diffusers. It uses diffusion model priors to reconstruct high-fidelity images. The approach supports FP8 inference and CPU offloading to drastically reduce memory, enabling high-resolution tasks on consumer hardware.

Quick Start & Requirements

• Gradio demo: python gradio_demo.py.
• Dependencies: Python, Conda, CUDA GPUs.
• Memory optimization: --cpu_offload and --use_fp8 crucial for low-VRAM GPUs (5GB VRAM for 2x upscale with both).
• Training: conda env create -f environment.yml, then run stage scripts.
• Inference: Download pre-trained models (FaithDiff, SDXL, LLaVA, etc.) to ./checkpoints, execute test.py.
• 8K+ restoration: Use --use_tile_vae flag.

Highlighted Details

• CVPR 2025 publication on faithful image super-resolution.
• Code integrated into Hugging Face diffusers.
• Significant memory reduction via FP8 inference and CPU offloading, enabling 8K+ restoration on 24GB GPUs.
• Introduces RealDeg dataset (238 real-world degraded images).

Maintenance & Community

• Developed by Junyang Chen, Jinshan Pan, Jiangxin Dong (IMAG Lab, Nanjing University of Science and Technology).
• Acknowledges diffusers, SUPIR, TLC, SimpleTuner.
• Contact: jychen9811@gmail.com.
• No explicit community channels or roadmap links provided.

Licensing & Compatibility

• The repository's license is not explicitly stated in the README. Clarification is needed for commercial use or integration into proprietary systems.

Limitations & Caveats

• Ultra-high-resolution results may still require substantial GPU resources (e.g., 24GB VRAM).
• Ongoing development for variants like "FaithDiff-SD3-Large" suggests potential incompleteness or experimental status.