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.
Figures & tables
Figure 1: Comparison of SteerFlow and existing inversion-based methods. (a) Unconstrained generation causes over-editing with loss of source structure. (b) Prior approaches, represented by UniEdit [ 14 ] , either rely on computationally expensive second-order solvers, or heuristic feature injection/truncated inversion that often suffers from under-editing (as shown by the unchanged source subject) and limited architectural transferability. (c) In contrast, SteerFlow presents a principled, architecture-agnostic editing framework that achieves high source-fidelity and faithful target alignment.
Figure 2: SteerFlow Editing Framework. Note that adaptive masking is optional.
Figure 3: An illustration of the AFP Solver with increasing iterations.
Figure 4: Comparison of masking strategies. No masking causes editing leakage, while using only the SAM3 mask restricts editability. SteerFlow Adaptive Mask dynamically expands the SAM3 mask to accommodate appropriate structural changes.
Figure 5: Qualitative comparison of editing baselines and SteerFlow for various editing tasks on FLUX.1-dev model.
Method
NFE ↓
Rank ↓
Source Preservation
Target Alignment
Image Quality
DinoDist ↓
LPIPS ↓
SSIM ↑
PSNR ↑
CLIP ↑
ImgRew ↑
HPSv2 ↑
AS ↑
FLUX.1-dev
Inversion-Free Methods
FlowEdit [ 17 ]
48
3.62
2.61
10.50
84.04
22.37
25.27
0.654
28.78
6.74
Inversion-Based Methods
ODE Inv
30
4.50
10.38
30.64
64.98
14.52
26.67
1.275
31.25
6.98
Table 1: Quantitative comparison on PIE-Bench using FLUX.1-dev and SD3.5-Medium. Bold indicates best, underline second best. DinoDist, LPIPS, SSIM, and HPSv2 are scaled ×102 for readability. For SD3.5, UniEdit uses N=30 steps (default is 15) to be comparable to other methods. SteerFlow uses K=1 optimization steps.
Model
Method
Rank ↓
Conditional Reconstruction
Unconditional Reconstruction
PSNR ↑
LPIPS ↓
SSIM ↑
MSE ↓
PSNR ↑
LPIPS ↓
SSIM ↑
MSE ↓
FLUX.1-dev ( N=15 )
ODE Inv
5.00
15.32
41.11
55.97
3.52
15.65
48.04
48.11
3.08
RF-Solver [ 30 ]
3.38
19.32
20.80
73.34
2.13
19.39
21.47
72.01
1.80
FireFlow [ 8 ]
1.62
20.45
17.93
75.83
1.67
20.26
19.65
74.12
1.58
UniEdit [ 14 ]
3.25
19.51
18.66
76.12
3.27
18.98
21.21
73.12
2.06
SteerFlow
1.75
20.26
15.99
78.83
2.60
19.63
17.70
76.84
1.91
Table 2: Quantitative comparison on image reconstruction. Bold indicates the best performance, and underline indicates the second best. LPIPS, SSIM, and MSE are scaled ×102 for readability. SteerFlow uses K=8 optimization steps.
Figure 6: Multi-turn image editing with SteerFlow. Each turn applies one of three edit operations: object replacement, attribute modification, object addition.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Additional SteerFlow results on FLUX.1-dev model.
Figure 8: Additional SteerFlow results on Stable Diffusion 3.5-Medium model.
Figure 9: Additional SteerFlow results on Multi-Turn editing. First row is Flux.1-dev, second row is Stable Diffusion 3.5-Medium. For both models, SteerFlow enables consistent multi-turn editing and complex multi-object editing (in the last row).
Figure 10: Qualitative comparison of baseline editing methods and SteerFlow on Stable Diffusion 3.5-Medium. SteerFlow demonstrates strong source preservation and faithful target prompt alignment. Previous editing baselines struggle with pose and structural change (in the third and fourth rows). ODE Inversion suffers from over-editing.
Figure 11: Qualitative comparison on inversion & reconstruction performance of baselines and SteerFlow.
Figure 12: Comparision of SteerFlow multi-turn editing and single-turn editing. First row: single-turn editing. Second row: multi-turn editing.
Figure 13: Quantitative comparison of Standard Fixed-Point solver (first row) and AFP solver (second row) with increasing K .
Figure 14: Qualitative comparison of Standard Fixed-Point solver (first row) and AFP solver (second row) with increasing K .
Figure 15: Evaluation of AFP solver on editing performance with increasing K .
Figure 16: Impact of varying decay rate γ on editing performance.
Figure 17: Impact of varying CFG weight w on editing performance.
Figure 18: Illustrations of varying decay rate γ on editability.
Variant
DinoDist ↓
LPIPS ↓
CLIP ↑
ImgRew ↑
Fixed α=0.1
0.05
2.98
21.49
-0.64
Fixed α=0.5
0.85
10.90
23.91
0.12
Fixed α=0.9
7.76
40.31
26.57
1.08
Only time-decaying
3.10
24.55
25.32
0.70
Only Cossim
4.99
30.97
26.50
1.08
time-decaying + Cossim
2.20
19.87
25.12
0.60
Appendix
Table 3: SteerFlow editing performance with different α -schedulers
Method
Model
Hyperparameters
timesteps N
target CFG w
start step
injection steps
stop step
forward guidance γ
backward guidance η
ODE-Inversion
SD3.5
50
5.5
0/50
-
-
-
-
FLUX
28
3.5
0/28
-
-
-
-
RF-Inversion
FLUX
28
3.5
0/28
-
7/28
0.5
0.9
RF-Solver
FLUX
15
2.0
0/15
2
-
-
-
FireFlow
FLUX
15
2.0
0/15
2
-
-
-
Appendix
Table 4: Experiment settings of inversion-based baselines
Method
Model
Hyperparameters
timesteps N
source CFG ws
target CFG wt
start step
correction weight ζ
FlowEdit
SD3.5
50
1.5
5.5
17/50
-
FLUX
28
5.5
13.5
4/28
-
FlowAlign
SD3.5
33
-
13.5*
0/50
0.01
Appendix
Table 5: Experiment settings of inversion-free baselines
Method
Model
Hyperparameters
timesteps N
decay rate γ
CFG w
AFP iterations K
SteerFlow
SD3.5
30
5.5
3.5
1
FLUX
15
4.5
6.5
1
Appendix
Table 6: Experiment settings of SteerFlow
quantile ratio q
sigmoid temperature τ
kernel size k
0.95
15
5
Appendix
Table 7: SteerFlow Adaptive Mask setting
Figure 19: A sample result of inversion-based editing with a perfect latent that is used for the source image.