Research map / Image generation / Image restoration
FlowIE: Efficient Image Enhancement via Rectified Flow
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 FlowIE, a conditioned rectified flow framework that adapts a pretrained Stable Diffusion prior into a many-to-one flow for image enhancement, using midpoint "mean value sampling" to achieve high quality in under 5 steps.
Problem
Diffusion-based image enhancement methods leverage strong generative priors but require long, costly sampling, making them impractical; existing rectified flow assumes one-to-one noise-to-image pairs and synthetic data, which does not suit real-world enhancement.
Approach
Adapt rectified flow into a many-to-one mapping where any Gaussian noise point flows to a fixed real HQ image, conditioned on a coarse restoration from a pretrained deblurring model via a ControlNet-style branch; at inference, a velocity sampled at a learned midpoint (Lagrange mean value theorem) gives a more accurate direction in few Euler steps.
Key results
- FlowIE achieves FID 19.81 and IDS 0.69 on synthetic CelebA-Test, the best among compared methods.
- FlowIE reaches MANIQA 0.5953 on RealSRSet and 0.6087 on Collect-100, outperforming DiffBIR (0.5906 / 0.6022).
- FlowIE runs at 2.846 FPS versus 0.285 FPS for DiffBIR (about 10x faster) with inference in 4 steps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FlowIE instead of DiffBIR | BFR · CelebA-Test | FID: 20.19 -> 19.81 | Helped |
| Use FlowIE instead of DiffBIR | BSR · RealSRSet | MANIQA: 0.5906 -> 0.5953 | Helped |
| Remove rectified flow (direct distillation) (ablation) | FlowIE · CelebA-Test | FID: 19.81 -> 49.74 | Hurt |
| Remove rectified flow (direct distillation) (ablation) | FlowIE · RealSRSet | MANIQA: 0.5953 -> 0.5311 | Hurt |
| Remove mean value sampling (ablation) | FlowIE · CelebA-Test | FID: 19.81 -> 25.19 | Hurt |
| Remove mean value sampling (ablation) | FlowIE · RealSRSet | MANIQA: 0.5953 -> 0.5489 | Hurt |
| Remove the initial-stage model (ablation) | FlowIE · CelebA-Test | FID: 19.81 -> 27.76 | Hurt |
| Remove the initial-stage model (ablation) | FlowIE · RealSRSet | MANIQA: 0.5953 -> 0.5301 | Hurt |
Surprising findings
- Direct one-step distillation of the diffusion model (w/o flow) performs much worse (CelebA FID 49.74) than the rectified flow version.
Limitations
- Midpoint index k must be selected with a few test samples per task.
- Evaluated mainly on face restoration and BSR; general enhancement covered only through limited fine-tuning.
Experimental setup
Datasets: FFHQ ImageNet CelebA-Test LFW-Test CelebChild-Test WIDER-Test RealSRSet Collect-100
Benchmarks: CelebA-Test LFW-Test WIDER-Test CelebChild-Test RealSRSet Collect-100
Metrics: FID IDS PSNR SSIM LPIPS MANIQA FPS
Baselines: GPEN GCFSR GFPGAN VQFR RestoreFormer DMDNet CodeFormer DiffBIR Real-ESRGAN+ BSRGAN SwinIR-GAN FeMaSR
Builds on: rectified flow Stable Diffusion ControlNet SwinIR VQGAN
Closest papers by meaning
- Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-Aware Guidance (CVPR 2024 · similarity 0.82)
- Low-Light Image Enhancement via Generative Perceptual Priors (AAAI 2025 · similarity 0.81)
- Multi-Modal Synergistic Implicit Image Enhancement for Efficient Optical Flow Estimation (CVPR 2025 · similarity 0.81)
- DiffEditor: Boosting Accuracy and Flexibility on Diffusion-based Image Editing (CVPR 2024 · similarity 0.81)
- Reversing Flow for Image Restoration (CVPR 2025 · similarity 0.81)
- DPFlow: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework (CVPR 2025 · similarity 0.80)
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.