Recent diffusion and flow matching models have demonstrated strong capabilities in image generation and editing by progressively removing noise through iterative sampling. While this enables flexible inversion for semantic-preserving edits, few-step sampling regimes suffer from poor forward process approximation, leading to degraded editing quality. Existing few-step inversion methods often rely on pretrained generators and auxiliary modules, limiting scalability and generalization across different architectures. To address these limitations, we propose BiFM (Bidirectional Flow Matching), a unified framework that jointly learns generation and inversion within a single model. BiFM directly estimates average velocity fields in both image $\to$ noise" and noise → image" directions, constrained by a shared instantaneous velocity field derived from either predefined schedules or pretrained multi-step diffusion models. Additionally, BiFM introduces a novel training strategy using continuous time-interval supervision, stabilized by a bidirectional consistency objective and a lightweight time-interval embedding. This bidirectional formulation also enables one-step inversion and can integrate seamlessly into popular diffusion and flow matching backbones. Across diverse image editing and generation tasks, BiFM consistently outperforms existing few-step approaches, achieving superior performance and editability.
Figures & tables
Figure 1: Inversion-Based Image Editing. (a) In training-free inversion, the process is approximated by numerically reversing the generation steps, leading to accumulated approximation errors; (b) An auxiliary inversion network is introduced on top of a pretrained generator, improving fidelity but increasing complexity and reducing generalization across architectures. (c) Our method, BiFM, jointly learns generation and inversion within a single flow matching model, enabling consistent few-step inversion and editing.
Figure 2: Overview of BiFM. (a) Our one-step generation architecture built upon MMDiT based flow matching model. (b) A single MMDiT block showing how time embedding modulation impact model output. (c) Naive DDIM inversion reuses DDIM update in reverse time, causing departures from original ODE trajectory in few-step regime. (d) Tuning based inversion introduces an auxiliary network Φ(xt′,t′,t) (e) BiFM inversion (ours) learns a physically constrained bidirectional average velocity field.
Settings
Methods
NFE
Background Preservation
CLIP Semantics
LPIPS ×103↓
SSIM %↑
PSNR ↑
MSE ×104↓
Whole ↑
Edited ↑
Multi-Step
NT Inv [ 24 ] (CVPR23)
50
60.67
84.11
27.03
35.86
23.61
21.64
MasaCtrl [ 3 ] (ICCV23)
50
106.62
79.67
22.17
86.97
23.96
21.16
PnP Inv [ 15 ] (ICLR24)
50
49.25
84.86
27.22
32.87
25.83
22.39
EditFT [ 38 ] (CVPR25)
30
80.55
91.50
26.62
40.24
25.74
22.27
FlowEdit [ 17 ] (ICCV25)
28
112.19
83.08
21.96
94.99
25.25
22.58
Table 1: PIE-Bench Image Editing Performance. We compare with baselines under different sampling budgets with evaluation metrics from PIE-Bench [ 15 ] . More baselines can be found in Appendix.
Method
MSE ×104↓
LPIPS ×103↓
SSIM %↑
PSNR ↑
DDIM [ 29 ] (ICLR21)
224.43
210.84
70.96
17.76
NT Inv [ 24 ] (CVPR23)
97.64
102.25
82.33
27.82
PnP Inv [ 15 ] (ICLR24)
105.66
95.95
87.20
28.79
RF-Solver [ 34 ] (ICML25)
94.80
106.15
86.36
28.26
BiFM (ours)
87.72
89.49
88.03
30.32
Table 2: Image Reconstruction Performance . We use 50 inversion steps to generate results for all methods. BiFM’s learned inversion process greatly reduces reconstruction error.
Figure 3: Inversion and Reconstruction Quality. From left to right: original input image, PnP Inversion [ 15 ] , RF-Edit [ 34 ] , and BiFM (ours) . BiFM faithfully reconstructs image details, while RF-Edit exhibits semantic shift and PnP Inv fails to recover fine details in the source image.
Figure 4: Image Editing Visualization. Given a source image, a source prompt and a target prompt (left illustrates difference between source prompt and target prompt), BiFM generates edits which follow more faithfully the intended concept while better preserving the original layout and fine details than other baselines. For example, BiFM engraves a clear lion pattern on the latte art without distorting the background, swaps the Statue of Liberty’s torch for a flower without geometry distortion, and maintains the lighthouse structure.
Table 3: Ablation Study on 1-NFE Image Generation. FID computed from 50K samples is reported. Defaults are shown in bottom row.
Method
Model
FID ↓
U-Net [ 27 ]
Diffusion
7.32
U-ViT-S/2 [ 1 ]
Diffusion
5.95
U-ViT-S/2 (Deep) [ 1 ]
Diffusion
5.48
MMDiT
Flow Matching
6.05
MMDiT+REPA
Flow Matching
4.73
MMDiT+MeanFlow
Flow Matching
5.02
Table 4: MSCOCO-256 Text-to-Image Generation. FID computed from 40K samples is reported.
Setting
Model
FID ↓
NFE
Multi-Step
DDIM [ 29 ]
4.67
50
Flow Matching [ 20 ]
2.63
50
BiFM (ours)
2.17
50
One-Step
Rectified Flow [ 21 ]
4.85
1
sCT [ 22 ]
2.85
1
MeanFlow [ 11 ]
2.92
1
Table 5: Unconditional CIFAR-10 Results. We include both multi-step and one-step FID.
Model
Epochs
FID ↓
SiT-B/2
80
33.0
SiT-B/2 + BiFM
80
27.8
SiT-L/2
80
18.8
SiT-L/2 + BiFM
80
16.4
SiT-XL/2
80
17.2
SiT-XL/2 + BiFM
80
15.5
Table 6: ImageNet Performance across Model Size. Under multi-step sampling, BiFM training also consistently improves performance across model sizes.
Recent advances in flow-based generative models have enabled training-free, text-guided image editing by inverting an image into its latent noise and regenerating it under a new target conditional guidance. However, existing methods struggle to preserve source fidelity: higher-order solvers incur additional model inferences, truncated inversion constrains editability, and feature injection methods lack architectural transferability. To address these limitations, we propose SteerFlow, a model-agnostic editing framework with strong theoretical guarantees on source fidelity. In the forward process, we introduce an Amortized Fixed-Point Solver that implicitly straightens the forward trajectory by enforcing velocity consistency across consecutive timesteps, yielding a high-fidelity inverted latent. In the backward process, we introduce Trajectory Interpolation, which adaptively blends target-editing and source-reconstruction velocities to keep the editing trajectory anchored to the source. To further improve background preservation, we introduce an Adaptive Masking mechanism that spatially constrains the editing signal with concept-guided segmentation and source-target velocity differences. Extensive experiments on FLUX.1-dev and Stable Diffusion 3.5 Medium demonstrate that SteerFlow consistently achieves better editing quality than existing methods. Finally, we show that SteerFlow extends naturally to a complex multi-turn editing paradigm without accumulating drift.
Thinh Dao, Zhen Wang, Kien T. Pham +1
The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong · VinUniversity, Ha Noi, Viet Nam
We propose Perceptual Flow Matching (PFM), a simple yet effective framework for few-step generation in flow-matching models. Rather than performing velocity regression in the conventional VAE latent space, PFM supervises flow matching in a perceptual feature space using pretrained perceptual models. This simple change substantially improves the few-step generation capability of flow-matching models, reducing the number of sampling steps from 35-50 to 4-8 while preserving generation quality. Unlike existing acceleration and distillation approaches, PFM requires neither teacher models nor auxiliary score networks and can be integrated into standard flow-matching training pipelines with minimal modifications. Extensive experiments on image generation, video generation, and image editing tasks demonstrate that PFM consistently produces high-quality results while producing fewer artifacts than existing distillation-based methods. We further show that perceptual supervision shifts the regression minimizer from mean-seeking to mode-seeking, biasing predictions toward on-manifold modes that remain accurate under coarse few-step integration. Our results reveal that standard flow-matching training can naturally yield high-quality few-step generators when supervised in an appropriate representation space. We hope this insight inspires future research into representation-aware objectives for efficient generative modeling.
Chuyang Zhao, Yifei Song, Hongfa Wang +7
1Joy Future Academy · 2Fudan University · 3Tsinghua University +1
With recent advancements in large-scale pre-trained text-to-image (T2I) models, training-free image editing methods have demonstrated remarkable success. Typically, these methods involve adding noise to a clean image via an inversion process, followed by separate denoising steps for the reconstruction and editing paths during the forward process. However, since the reconstruction path is approximated using noisy latents from mismatched timesteps, existing methods inevitably suffer from accumulated drift, which fundamentally limits reconstruction fidelity. To address this challenge, we systematically analyze the inversion process within the flow transformer and propose DirectEdit, a simple yet effective editing method that eliminates the inherent reconstruction error without introducing additional neural function evaluations (NFEs). Unlike most prior works that attempt to rectify the inversion path, DirectEdit focuses on directly aligning the forward paths, enabling precise reconstruction and reliable feature sharing. Furthermore, we introduce a preservation mechanism based on attention feature injection and multi-branch mask-guided noise blending, which effectively balances fidelity and editability. Extensive experiments across diverse scenarios demonstrate that DirectEdit achieves efficient and accurate image editing, delivering superior performance that outperforms state-of-the-art methods. Code and examples are available at https://desongyang.github.io/Directedit.
Desong Yang, Mang Ye
National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University, Wuhan, China.