VeloEdit: Training-Free Consistent and Continuous Instruction-Based Image Editing via Velocity Field Decomposition

Zongqing Li; Zhihui Liu; Yujie Xie; Shansiyuan Wu; Hongshen Lv; Songzhi Su

Figure 1: VeloEdit constructs continuous editing trajectories for instruction-based image editing models. Our method empowers these models to achieve continuous and consistent control over edit effects without additional training.

Figure 2: Masks derived via the velocity field. These masks separate preservation regions from editing regions, serving as the foundation for enhancing consistency and enabling precise control over editing intensity.

Figure 3: Impact of early velocity replacement. Intervening in the initial one or two timesteps completely suppresses the editing effect.

Figure 4: Overview of the proposed pipeline. We derive a spatial mask by analyzing the velocity discrepancy between preservation and editing flows. Our method explicitly preserves high similarity regions while blending low similarity regions, thereby yielding a sequence of edited results with smooth semantic transitions.

Figure 5: Editing results of the high similarity velocity replacement strategy. By substituting predicted velocities in high similarity regions(

S_{t}>\tau

) with the preservation velocity, VeloEdit effectively maintains the structural integrity of non-edited regions.

Figure 6: Visual editing results on GPT-Image-Edit [ 51 ] and Subject200K [ 45 ] . VeloEdit achieves smooth control over both local and global editing intensities.

Method	PSNR $\uparrow$	SSIM ( $\times 10^{2}$ ) $\uparrow$	CLIP-Sim. $\uparrow$
P2P [ 14 ]	17.87	71.14	25.01
PnP [ 47 ]	22.28	79.05	25.41
MasaCtrl [ 4 ]	22.17	79.67	23.96
RF-Inversion [ 38 ]	20.82	71.92	25.20
RF-Solver [ 49 ]	22.90	81.90	26.00
UniEdit [ 18 ]	24.10	84.86	26.97

Table 1: Consistency evaluation results on PIEbench. Our method achieves competitive performance compared to methods based on inversion and attention map injection. Baseline results are cited from [ 30 ] , and the best and second-best results are indicated in bold and underlined, respectively.

Figure 7: Qualitative comparison results. VeloEdit generates edited results with superior consistency and continuity, effectively avoiding background detail alteration and drift.

Figure 8: Qualitative comparison with CFG-scale guidance. VeloEdit produces consistent and continuous edits, while CFG-scale fails to maintain continuity.

Figure 9: Qualitative comparison results. Our method is effective in both Flux.1 Kontext [ 23 ] and Qwen-Image-Edit [ 53 ] , and can generate continuous and smooth editing results.

Method	$\delta_{\text{smooth}}\downarrow$	CLIP-Dir. $\uparrow$	$L_{1}\downarrow$	$L_{2}\downarrow$
Training-Free
Freemorph [ 5 ]	0.354	0.147	0.142	0.211
CFG-scale [ 23 , 16 ]	3.362	0.379	0.140	0.209
Training-Based
KontinuousKontext [ 32 ]	0.280	0.219	0.083	0.132
VeloEdit	0.246	0.294	0.074	0.116

Table 2: Quantitative experiments on PIEbench. VeloEdit achieves the best or second-best metrics compared with training-free and training-based methods.

Method	$\delta_{\text{smooth}}\downarrow$	CLIP-Dir. $\uparrow$	$L_{1}\downarrow$	$L_{2}\downarrow$
$\text{VeloEdit}_{\text{Flux.1 Kontext}}$	0.246	0.294	0.074	0.116
$\text{VeloEdit}_{\text{Qwen-Image-Edit}}$	0.207	0.349	0.075	0.104

Table 3: Cross-model generalization performance on the PIEbench. VeloEdit demonstrates consistent effectiveness when integrated with Flux.1 Kontext and Qwen-Image-Edit.

$\tau$	Flux.1 Kontext				Qwen-Image-Edit
$\tau$	$\delta_{\text{smooth}}\downarrow$	CLIP-Dir. $\uparrow$	$\quad L_{1}\downarrow$	$\quad L_{2}\downarrow\quad$	$\delta_{\text{smooth}}\downarrow$	CLIP-Dir. $\uparrow$	$\quad L_{1}\downarrow\quad$	$\quad L_{2}\downarrow\quad$
0.0	0.038	0.079	0.034	0.052	0.024	0.149	0.023	0.038
0.2	0.751	0.168	0.041	0.063	0.921	0.195	0.025	0.040
0.4	0.439	0.249	0.055	0.088	0.432	0.279	0.036	0.056
0.6	0.303	0.285	0.067	0.107	0.242	0.334	0.058	0.084
0.8	0.246	0.294	0.074	0.116	0.207	0.349	0.075	0.104

Table 4: Ablation study on

\tau

. Performance of VeloEdit using different

\tau

values across Flux.1 Kontext and Qwen-Image-Edit.

$N$	Flux.1 Kontext				Qwen-Image-Edit
$N$	$\delta_{\text{smooth}}\downarrow$	CLIP-Dir. $\uparrow$	$\quad L_{1}\downarrow$	$\quad L_{2}\downarrow\quad$	$\delta_{\text{smooth}}\downarrow$	CLIP-Dir. $\uparrow$	$\quad L_{1}\downarrow\quad$	$\quad L_{2}\downarrow\quad$
1	0.246	0.294	0.074	0.116	0.207	0.351	0.078	0.108
2	0.220	0.207	0.063	0.098	0.133	0.271	0.059	0.082
3	0.206	0.167	0.057	0.089	0.153	0.215	0.048	0.068
4	0.229	0.138	0.052	0.083	0.259	0.170	0.041	0.059
5	0.290	0.116	0.048	0.077	0.331	0.106	0.034	0.050

Table 5: Ablation study on

N

. Performance of VeloEdit using different

N

values across Flux.1 Kontext and Qwen-Image-Edit.

Figure 10: Visualizations of extrapolation for

\alpha

. By employing blending intensities beyond the standard

[0,1]

range, VeloEdit can achieve either inverse semantic effects (

\alpha<0

) or intensified editing results (

\alpha>1

) relative to the prompt.

Figure 11: Additional visual results on Subject200K. VeloEdit generates consistent and continuous results across diverse editing tasks.

Figure 12: Additional visual results on GPT-Image-Edit. VeloEdit generates consistent and continuous results across diverse editing tasks.

Figure 13: By employing a larger

\alpha

, VeloEdit expands the editing boundaries of the foundation model and achieves more pronounced continuous editing outcomes.

Figure 14: Additional qualitative comparison with Kontinuous Kontext. Our method better preserves consistency in unedited regions and generates more continuous editing results.

Figure 15: Additional qualitative comparison with FreeMorph and CFG-scale. Our method better preserves consistency in unedited regions and generates more continuous editing results.

Figure 16: Additional visual results. Our method is effective in both Flux.1 Kontext and Qwen-Image-Edit, and can generate continuous and smooth editing results.

Edit Type	Method	$\delta_{\text{smooth}}\downarrow$	CLIP-Dir. $\uparrow$	$L_{1}\downarrow$	$L_{2}\downarrow$
Random	FreeMorph	0.302	0.177	0.150	0.223
	KontinuousKontext	0.119	0.251	0.089	0.136
	VeloEdit	0.151	0.321	0.089	0.135
Change Object	FreeMorph	0.410	0.208	0.159	0.253
	KontinuousKontext	0.379	0.236	0.091	0.155
	VeloEdit	0.277	0.384	0.086	0.151

Table 6: Performance comparison of VeloEdit and other continuous editing baselines across different editing tasks.

Method	Total time (s)	Intervention time (s)	Rate (%)
$\text{VeloEdit}_{\text{Flux.1 Kontext}}$	210.9	0.031	0.015
$\text{VeloEdit}_{\text{Qwen-Image-Edit}}$	737.3	0.041	0.005

Table 7: Analysis of computational overhead and efficiency. We report the total inference time for 100 images alongside the additional intervention cost introduced by VeloEdit. The marginal overhead confirms the efficiency of VeloEdit.

Figure 17: Visualizations of the ablation study on

\tau

. We illustrate the impact of varying

\tau

on the continuity of the generated editing trajectories.

Figure 18: Visualizations of the ablation study on

\tau

. We illustrate the impact of varying

\tau

on the continuity of the generated editing trajectories.

Figure 19: Visualizations of the ablation study on

N

. We illustrate the impact of varying

N

on the continuity of the generated editing trajectories.

Figure 20: Visualizations of the ablation study on

N

. We illustrate the impact of varying

N

on the continuity of the generated editing trajectories.

Figure 21: Failure cases. VeloEdit exhibits limited continuity in tasks involving object addition, object removal, and pose variation.

VeloEdit: Training-Free Consistent and Continuous Instruction-Based Image Editing via Velocity Field Decomposition

Abstract

Figures & tables

Training-Free Image Editing

Training-Free Image Editing with Visual Context Integration and Concept Alignment

Constrained Edit Fields for Training-Free Flow Editing

h-Flow: Flexible Flow-based Image Editing via Doob's h-Transform

Abstract

Training-Free Image Editing

Explore similar work

Training-Free Image Editing with Visual Context Integration and Concept Alignment

Constrained Edit Fields for Training-Free Flow Editing

h-Flow: Flexible Flow-based Image Editing via Doob's h-Transform