RealRestorer tops open-source image restoration benchmarks

RealRestorer represents a significant leap for open-source image editing, bringing professional-grade restoration to standard hardware through natural language control.

• –Uses a Diffusion in Transformer (DiT) backbone with a QwenVL text encoder, enabling users to direct restoration with specific prompts like "remove lens flare."
• –Currently holds the #1 spot for open-source models on the RealIR-Bench, trailing closed-source leaders by a negligible margin.
• –Trained on a two-stage process that balances synthetic and real-world data, reducing artifacts and maintaining authentic image textures.
• –The model's versatility across multiple degradation types (blur, rain, reflections) makes it a comprehensive tool for both hobbyist restoration and professional archival work.
• –Open-source availability via GitHub and Hugging Face facilitates rapid integration into creative tools and developer workflows.