Research map / Image generation / Super-resolution

SRSR: Enhancing Semantic Accuracy in Real-World Image Super-Resolution with Spatially Re-Focused Text-Conditioning

NeurIPS 2025 · Image generation › Super-resolution · approach: Perceptual and real-world SR · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SRSR, an inference-only plug-and-play framework that grounds DAPE-extracted tags with Grounded SAM, re-focuses cross-attention to tag masks (SRCA), and disables text CFG on ungrounded pixels (STCFG) to reduce hallucinations in diffusion-based real-world super-resolution.

Problem

Text-conditioned diffusion super-resolution suffers semantic misalignment and hallucinations because cross-attention leaks text tokens onto irrelevant pixels, degraded-image prompt extraction is error-prone, and tags do not cover the full image.

Approach

Use visually grounded tag-mask pairs (DAPE tags filtered by Grounded SAM 2) to mask cross-attention per token (SRCA), and apply classifier-free guidance only on grounded pixels while using unconditional noise prediction on ungrounded pixels (STCFG), all at inference time without training.

Key results

• SRSR-SeeSR-s50 reaches 26.40 PSNR and 0.7632 SSIM on RealSR versus 25.18 and 0.7216 for SeeSR, with best LPIPS (0.2718) and DISTS (0.2092) on that dataset.
• SRSR-SeeSR achieves the best PSNR, SSIM, LPIPS and DISTS on DrealSR (29.50 PSNR, 0.8128 SSIM, 0.2866 LPIPS), while on DIV2K-Val it has best PSNR and SSIM (24.72, 0.6416) but second-best LPIPS and DISTS.
• SRSR-OSEDiff-s1 improves PSNR and SSIM over OSEDiff on all three datasets (e.g., RealSR 24.43 to 24.53 PSNR) with slight perceptual trade-offs.

Measured effects reported in the paper

Surprising findings

• No-reference metrics (NIQE, MUSIQ, MANIQA, CLIPIQA) often improve for hallucinated outputs and drop when hallucinations are removed, so the authors deem them unreliable for SR semantic fidelity.
• Adding semantic segmentation tags (Mask2Former or DINO-X) to DAPE tags degrades performance relative to DAPE alone, suggesting inaccurate tags are more harmful than incomplete coverage.

Limitations

• Relies on the accuracy of DAPE tags and Grounded SAM; conservative grounding can drop partially relevant tags and reduce coverage.
• Inference-only design; integrating SRSR into training is left for future work.
• No-reference metrics are unreliable for judging hallucinated semantic content, limiting evaluation.

Experimental setup

Datasets: StableSR test set DIV2K-Val RealSR DRealSR

Benchmarks: DIV2K-Val RealSR DrealSR

Metrics: PSNR SSIM LPIPS DISTS FID NIQE MUSIQ MANIQA CLIPIQA

Baselines: StableSR DiffBIR PASD ResShift SinSR OSEDiff SeeSR

Builds on: SeeSR OSEDiff DAPE Grounded SAM classifier-free guidance

Closest papers by meaning

• Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach (CVPR 2025 · similarity 0.90)
• Text-Aware Real-World Image Super-Resolution via Diffusion Model with Joint Segmentation Decoders (NeurIPS 2025 · similarity 0.90)
• The Power of Context: How Multimodality Improves Image Super-Resolution (CVPR 2025 · similarity 0.89)
• SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution (CVPR 2024 · similarity 0.89)
• PRISM: Prior Rectification and Uncertainty-Aware Structure Modeling for Diffusion-Based Text Image Super-Resolution (NeurIPS 2026 · similarity 0.88)
• CoSeR: Bridging Image and Language for Cognitive Super-Resolution (CVPR 2024 · similarity 0.88)