Reproduce “Revisiting the Perception-Distortion Trade-off with Spatial-Semantic Guided Super-Resolution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dan Wang, Haiyan Sun, Shan Du, Z. Jane Wang, Zhaochong An, Serge Belongie, Xinrui Cui
Researchers developed SpaSemSR, a super-resolution framework that combines spatial-grounded textual guidance and semantic-enhanced visual guidance within a diffusion model. This approach generates images that are both perceptually realistic and faithful to the original input, outperforming existing methods on various metrics, including a 26.4% higher CLIP-IQA score over the second-best GAN-based model on DIV2K-Val.
Key insights
- Combining precise spatial-semantic textual cues with robust visual features from degraded inputs can effectively mitigate hallucinations in diffusion-based SR while maintaining high perceptual quality.
- Explicitly embedding universal statistical properties and degradation awareness into the guidance mechanisms helps diffusion models achieve higher fidelity in restoration tasks.
- An adaptive fusion of multimodal priors (spatial, semantic, and degradation) within the ControlNet and diffusion process allows for controlled generation that balances realism and faithfulness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.