New discrete diffusion model enhances image super-resolution

Researchers have introduced DiMOO-SR, a novel framework for photo-realistic image super-resolution that leverages discrete diffusion models. This approach addresses challenges in discrete diffusion for super-resolution, specifically the under-representation of rare visual tokens and spatial inconsistencies during decoding. DiMOO-SR employs Inverse Frequency Sampling during training to prioritize important but under-represented tokens and Spatial Consistency Ranking during inference to refine token confidence and improve structural coherence. Experiments on standard benchmarks show that DiMOO-SR can achieve competitive perceptual quality with a limited number of parallel decoding steps. AI

IMPACT This research advances discrete diffusion models for image generation, potentially improving the quality and efficiency of super-resolution tasks.

RANK_REASON The item is an academic paper detailing a new method for image super-resolution. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• DiMOO-SR
• Gotit.pub
• Hugging Face
• Inverse Frequency Sampling
• Litmaps
• ScienceCast
• scite Smart Citations
• Spatial Consistency Ranking

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →