Research map / Image generation / Fast diffusion sampling
BiFM: Bidirectional Flow Matching for Few-Step Image Editing and Generation
Yasong Dai, Zeeshan Hayder, David Ahmedt-Aristizabal, Hongdong Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train jointly for generation and inversion (BiFM) vs training-free inversion at 50 steps | Stable Diffusion 3 editing · PIE-Bench, 50 NFE | LPIPS x10^3: 49.25 -> 47.01 (vs PnP Inv) | Improved |
| BiFM few-step inversion at 4 steps vs strongest baseline InstantEdit | Stable Diffusion 3 editing · PIE-Bench, 4 steps | LPIPS x10^3: 44.39 -> 67.25 | Hurt |
| BiFM vs SwiftEdit in one-step editing | one-step editing · PIE-Bench, 1 step | SSIM %: 81.05 -> 85.88 SSIM; LPIPS 91.04 -> 92.30 | Mixed |
| learned bidirectional inversion vs baseline inversion methods | image reconstruction · PIE-Bench reconstruction | PSNR: 28.79 -> 30.32 (vs PnP Inv) | Improved |
| add BiFM training to MMDiT | MMDiT · MSCOCO-256 | FID: 6.05 -> 4.57 | Improved |
| add BiFM training to SiT models | SiT-B/2, L/2, XL/2 · ImageNet-256, 80 epochs | FID: 33.0 -> 27.8; 18.8 -> 16.4; 17.2 -> 15.5 | Improved |
| BiFM vs MeanFlow for one-step generation | CIFAR-10 U-Net · CIFAR-10, 1 NFE | FID: 2.92 -> 2.75 | Improved |
| condition on (t, t'-t) instead of (t, t') (ablation) | BiFM 1-NFE · 1-NFE image generation | FID: 59.37 -> 55.22 | Improved |
| add discrete direction flag to time input (ablation) | BiFM 1-NFE · 1-NFE image generation | FID: 55.22 -> 69.01 | Hurt |
| use warm-up weighting on consistency term instead of linear (ablation) | BiFM 1-NFE · 1-NFE image generation | FID: 67.37 -> 55.22 | Improved |
| sample (t,t') from lognorm(-0.4, 1.0) instead of uniform (ablation) | BiFM 1-NFE · 1-NFE image generation | FID: 63.25 -> 55.22 | Improved |
| use robust loss norm (p=1.0 default, p=0.5) instead of squared L2 (ablation) | BiFM 1-NFE · 1-NFE image generation | FID: 59.66 -> 55.22 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 3 on most shared results.
- InstantEdit — InstantEdit: Text-Guided Few-Step Image Editing with Piecewise Rectified Flow (ICCV 2025) · better on 2 of 4 shared results
- SwiftEdit — SwiftEdit: Lightning Fast Text-Guided Image Editing via One-Step Diffusion (CVPR 2025) · better on 1 of 2 shared results
- DNAEdit — DNAEdit: Direct Noise Alignment for Text-Guided Rectified Flow Editing (NeurIPS 2025) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations (ICLR 2025 · similarity 0.89)
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.89)
- SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing (NeurIPS 2025 · similarity 0.89)
- Unveil Inversion and Invariance in Flow Transformer for Versatile Image Editing (CVPR 2025 · similarity 0.89)
- Taming Rectified Flow for Inversion and Editing (ICML 2025 · similarity 0.89)
- PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code (ICLR 2024 · similarity 0.88)
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.
