Research map / Image generation / Super-resolution

Noise-Started One-Step Image Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement

NeurIPS 2026 · Image generation › Super-resolution · approach: Diffusion-based super-resolution

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 SMFSR, a noise-started one-step real-world super-resolution model that learns an LR-conditioned average-velocity (SplitMeanFlow) mapping from noise to HR latent, followed by a GAN/VSD refinement stage.

Problem

Diffusion-based real-world image super-resolution faces an efficiency-quality trade-off: multi-step noise-started methods are slow, while recent one-step methods restore directly from LR input, losing stochasticity and fine detail synthesis.

Approach

Learn a one-step noise-to-HR average-velocity field conditioned on the LR latent via Interval Splitting Consistency (SplitMeanFlow, initialized from a flow-matching SD3.5-Medium-based teacher with DiT4SR-style LR injection), then refine with a DINOv3-based GAN discriminator, variational score distillation, and reconstruction loss.

Key results

• SMFSR achieves the highest MUSIQ, MANIQA, and CLIPIQA among compared methods on DIV2K, DRealSR, RealSR, and RealLQ250 (e.g., CLIPIQA 0.7772 and MUSIQ 72.21 on RealLQ250).
• SMFSR runs in 0.22 s per 512x512 output with one step on an RTX 5880, slightly slower than OSEDiff (0.19 s).
• SMFSR has lower PSNR and SSIM than several baselines (e.g., 22.18 PSNR on DIV2K vs. 24.87 for CTMSR), which the authors attribute to the perception-distortion trade-off.

Surprising findings

• Multi-step inference does not improve the stage-1 student beyond one step, indicating the remaining gap is representational rather than step-count limited.
• The student after stage 1 surpasses the teacher in LPIPS but stays behind in CLIPIQA and MANIQA.

Limitations

• Lower full-reference fidelity (PSNR/SSIM) than many baselines.
• Larger trainable parameter count (366.2M) than several one-step baselines.

Experimental setup

Datasets: LSDIR FFHQ DIV2K-Val RealSR DRealSR RealLQ250

Benchmarks: DIV2K-Val RealSR DRealSR RealLQ250

Metrics: PSNR SSIM LPIPS MUSIQ MANIQA CLIPIQA inference time trainable parameters

Baselines: BSRGAN SwinIR StableSR SUPIR PASD SeeSR ResShift S3Diff OSEDiff SinSR CTMSR InvSR

Builds on: SplitMeanFlow MeanFlow DiT4SR OSEDiff variational score distillation DINOv3

Closest papers by meaning

• SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.91)
• Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model (CVPR 2025 · similarity 0.91)
• Adversarial Diffusion Compression for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.91)
• One-Step Effective Diffusion Network for Real-World Image Super-Resolution (NeurIPS 2024 · similarity 0.91)
• Iterative Token Evaluation and Refinement for Real-World Super-resolution (AAAI 2024 · similarity 0.90)
• Arbitrary-steps Image Super-resolution via Diffusion Inversion (CVPR 2025 · similarity 0.90)