Research map / Image generation / Super-resolution
One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation
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.
Distills the 12B FLUX.1-dev flow-matching model into a one-step Real-ISR model (FluxSR) via flow trajectory distillation on precomputed noise-image pairs, with TV-LPIPS and attention diversification loss to suppress periodic artifacts.
Problem
One-step diffusion Real-ISR methods are limited by teacher quality and distribution shift from fine-tuning T2I models, and distilling very large flow-matching models (12B) with an extra teacher is memory-prohibitive.
Approach
Derive the LR-to-HR velocity field from the noise-to-image (FLUX) and noise-to-LR flows via the linearity of rectified flow, so the original T2I flow is kept fixed; train on precomputed teacher noise-image pairs without the teacher online, and add a TV-augmented LPIPS loss and an attention diversification regularizer.
Key results
- On RealSR x4, FluxSR-s1 achieves the best no-reference scores among one-step methods (MUSIQ 68.95, MANIQA 0.5335, TOPIQ 0.6699, Q-Align 4.3781).
- On DIV2K-val x4, FluxSR-s1 reaches Q-Align 4.6128 and LPIPS 0.3397, while PSNR is lower than OSEDiff (22.30 vs 23.10).
- Ablation on RealSR: FTD raises MUSIQ from 56.02 to 67.84 and Q-Align from 3.517 to 4.147 at a cost of PSNR (26.33 to 24.67).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Flow Trajectory Distillation (ablation) | FluxSR · RealSR testset ×4 | MUSIQ: 56.02 -> 67.84 | Helped |
| Add Flow Trajectory Distillation (ablation) | FluxSR · RealSR testset ×4 | MANIQA: 0.3775 -> 0.5203 | Helped |
| Add Flow Trajectory Distillation (ablation) | FluxSR · RealSR testset ×4 | Q-Align: 3.5170 -> 4.1473 | Helped |
| Add Flow Trajectory Distillation (ablation) | FluxSR · RealSR testset ×4 | PSNR: 26.33 -> 24.67 | Hurt |
| Use TV-LPIPS instead of LPIPS (ablation) | FluxSR · RealSR testset ×4 | MUSIQ: 64.55 -> 65.04 | Helped |
| Use TV-LPIPS instead of LPIPS (ablation) | FluxSR · RealSR testset ×4 | MANIQA: 0.4937 -> 0.5113 | Helped |
| Use TV-LPIPS instead of LPIPS (ablation) | FluxSR · RealSR testset ×4 | Q-Align: 4.0515 -> 4.0927 | Helped |
| Use TV-LPIPS instead of LPIPS (ablation) | FluxSR · RealSR testset ×4 | PSNR: 23.10 -> 22.09 | Hurt |
| Add ADL to TV-LPIPS (ablation) | FluxSR · RealSR testset ×4 | MUSIQ: 65.04 -> 67.84 | Helped |
| Add ADL to TV-LPIPS (ablation) | FluxSR · RealSR testset ×4 | MANIQA: 0.5113 -> 0.5203 | Helped |
| Add ADL to TV-LPIPS (ablation) | FluxSR · RealSR testset ×4 | Q-Align: 4.0927 -> 4.1473 | Helped |
| Add ADL to TV-LPIPS (ablation) | FluxSR · RealSR testset ×4 | PSNR: 22.09 -> 24.67 | Helped |
Surprising findings
- Adding FTD lowers PSNR relative to reconstruction-only training, indicating a fidelity-realism trade-off.
- Training with the offline teacher-generated pairs avoids needing the 12B teacher online, while naive OSEDiff-style distillation on FLUX ran out of memory on 8x A800-80GB.
Limitations
- Large parameter count (12B) and high computational cost.
- Periodic high-frequency artifacts are not fully eliminated.
- PSNR/SSIM lower than some one-step baselines.
Experimental setup
Datasets: DIV2K-val RealSR RealSet65
Benchmarks: DIV2K-val RealSR RealSet65
Metrics: PSNR SSIM LPIPS DISTS MUSIQ MANIQA TOPIQ Q-Align
Baselines: StableSR DiffBIR SeeSR ResShift AddSR SinSR OSEDiff TSD-SR
Builds on: rectified flow OSEDiff DMD ADL
Closest papers by meaning
- Unleashing the Power of One-Step Diffusion based Image Super-Resolution via a Large-Scale Diffusion Discriminator (NeurIPS 2025 · similarity 0.92)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.91)
- TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.91)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.91)
- Solving Diffusion ODEs with Optimal Boundary Conditions for Better Image Super-Resolution (ICLR 2024 · similarity 0.91)
- Efficient Conditional Diffusion Model with Probability Flow Sampling for Image Super-resolution (AAAI 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.