Reproduce “Text-guided Explorable Image Super-resolution” on OpenResearch

About this paper

Kanchana Vaishnavi Gandikota, Paramanand Chandramouli

Zero-shot text-guided explorable image super-resolution methods are presented, adapting pre-trained text-to-image diffusion models to generate diverse, semantically controlled high-resolution images from low-resolution inputs. The approaches maintain high data consistency (LR PSNR > 45 dB) and demonstrate superior visual quality and semantic matching (lower NIQE, higher CLIP scores) compared to baselines, enabling users to explore variations in attributes like age and expression through natural language prompts.

Key insights

• Pre-trained text-to-image diffusion models can be effectively repurposed for zero-shot open-domain inverse problems like super-resolution by appropriately enforcing data consistency at each generation step.
• Natural language prompts offer a powerful and intuitive interface for semantically guiding image reconstruction, allowing users to explore diverse high-resolution solutions by specifying desired attributes or content.
• The combination of robust data consistency mechanisms (e.g., null-space projection) and advanced semantic guidance from large language-image models enables the generation of both data-faithful and semantically aligned high-resolution outputs for extreme super-resolution.

Summary by alphaXiv.