Organizations: National Taiwan University · Institute of Information Science, Academia Sinica · University of California, Davis · National Tsinghua University
Text-guided image editing requires realizing the requested semantic change while preserving content unrelated to the instruction. FlowEdit provides an efficient inversion-free solution, but its source-anchored coupling can become overly restrictive, causing zero-edit, where the output remains close to the source and the requested change fails to emerge. We find that source coupling directly governs the preservation--editing trade-off, while a fixed coupling schedule cannot adapt to timestep-specific editing needs. Motivated by this observation, we propose V2C-Edit, a training-free VLM-steered velocity control framework. V2C-Edit constructs complementary coupling-controlled velocities and dynamically selects among them using Instruction-Level Similarity (ILS), which evaluates whether the predicted transformation follows the editing instruction. Adaptive spatial masking further preserves regions unrelated to the edit. Experiments on MagicBrush and PIE-Bench with FLUX and Stable Diffusion 3 show that V2C-Edit consistently improves instruction alignment over FlowEdit while maintaining competitive structural preservation, and that ILS better reflects editing effectiveness than target-caption CLIP similarity. Our code is available at https://github.com/Kuanshen77/V2C-Edit.
Figures & tables
Figure 1: Coupling controls the preservation-editing trade-off in FlowEdit. As tcs decreases, the coupled stage is retained for longer, increasing PSNR to the source image over the annotated unedited region but weakening the requested semantic change. In the example of “Add a dog,” strong coupling eventually leads to zero-edit , where the output remains nearly unchanged despite high preservation.
Figure 2: Failure of target-caption CLIP similarity for evaluating image editing. Although the ground-truth edit correctly changes the clock face to English numbers, it receives a lower CLIP similarity than the unedited source image. In contrast, the instruction-level GPT score correctly favors the edited result, showing that target-caption similarity may fail to reflect whether the requested edit is actually realized.
Figure 3: Overview of the proposed V2C-Edit pipeline. (1) Coupling-Controlled Velocities: Construct four editing velocities with different coupling behaviors. (2) Instruction-Level VLM Steering: Select the optimal velocity at each timestep using ILS. (3) Two-Stage Adaptive Masking: Apply spatial masking when semantic alignment is sufficient and the masked update maintains ILS. The VLM query templates are provided in Appendix F .
Method
Venue
MagicBrush
PIEbench
CLIP ↑
IAGPT↑
IAGemini↑
BPPSNR↑
BPSSIM↑
CLIP ↑
IAGPT↑
IAGemini↑
BPPSNR↑
BPSSIM↑
FLUX
RF-Edit [ 42 ]
ICML’25
0.2721
1.48
0.83
21.49
0.6171
0.2359
1.83
0.96
31.85
0.9199
RF-Inversion [ 36 ]
ICLR’25
0.2851
4.12
4.10
17.16
0.5301
0.2562
4.14
4.12
20.58
0.7087
DNAEdit [ 43 ]
NeurIPS’25
0.2825
5.67
4.61
19.07
0.5932
0.2566
5.37
4.53
25.31
0.8831
FlowEdit [ 19 ]
ICCV’25
0.2865
5.56
4.76
17.48
0.5694
0.2618
5.32
4.78
21.82
0.8411
Table 1: Quantitative comparison on MagicBrush and PIEbench. The best is marked in bold and the second best is underlined.
Figure 4: Qualitative comparisons of text-based image editing methods. Our method achieves the highest fidelity to the editing instruction while maintaining strong background consistency with the source image.
Method
DNAEdit [ 43 ]
FlowEdit [ 19 ]
FlowAlign [ 15 ]
V2C-Edit (Ours)
Percentage
11.0%
16.7%
18.5%
53.8%
Table 2: User study results. Percentage indicates the proportion of participant selections received by each method.
Vtcpl
Vtdec
Vtext
Vtpdec
CLIP ↑
IAGPT↑
BPPSNR↑
BPSSIM↑
✓
✗
✗
✗
0.2864
5.69
17.78
0.5787
✓
✓
✗
✗
0.2888
5.91
17.78
0.5750
✓
✓
✓
✗
0.2879
5.91
17.57
0.5694
✓
✓
✓
✓
0.2891
6.07
17.44
0.5647
Table 3: Ablation study on editing direction candidates.
Metric
Token
Layer
Anomaly Rate ↓
Mean Score Gap ↑
CLIP
-
-
35.88%
0.0122
ILS
First
First
100%
0
Final
First
100%
0
Final
Final
16.82%
0.0533
First
Final
2.43%
0.2044
Table 4: Comparison of metrics across token positions and VLM layers on the MagicBrush. Anomaly rate measures the percentage of cases where the source scores at least as high as the ground truth. Mean score gap measures the average score difference.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Qualitative effect of the semantic threshold τ . Top: Lower thresholds ( τ≤0.6 ) trigger premature masking, leading to incomplete edits, whereas the default τ=0.7 successfully reduces the number of eggs to three. Bottom: Although all settings satisfy the instruction, disabling masking ( τ=∞ ) alters the person’s appearance and clothing, while τ=0.7 better preserves the source content.
Confidence τ
CLIP ↑
IAGPT
BPPSNR↑
BPSSIM↑
0.0
0.2877
5.72
18.36
0.5864
0.4
0.2877
5.76
18.36
0.5864
0.5
0.2876
5.74
18.34
0.5860
0.6
0.2881
5.96
17.91
0.5762
0.7
0.2891
6.07
17.44
0.5647
inf
0.2886
6.07
17.08
0.5535
Appendix
Table 5: Ablation on the semantic threshold. τ=0.0 applies masking at every timestep, τ=inf disables masking entirely, and intermediate values gate masking based on the semantic score.
Figure 6: Qualitative effect of progressively adding velocity candidates. Top: the coupled velocity produces a zero-edit result, while introducing a relaxed velocity enables the dog to be added. Middle: local one-step relaxation remains insufficient to replace the phone with a table lamp, whereas persistent decoupling successfully completes the edit; extrapolation also achieves the requested change but introduces larger unintended modifications. Bottom: extrapolation is required to correctly convert the clock numerals to English digits. Combining all candidates allows the controller to select the appropriate degree of coupling relaxation for different editing requirements.
Figure 7: Qualitative examples of ILS on MagicBrush source-ground truth pairs. For each example, the VLM generates a target edit instruction from the source image and instruction, and an inferred edit instruction from either the source image itself (source-to-source, no actual edit) or the ground truth image (source-to-GT, the intended edit).
Figure 8: Distribution of IA GPT scores on FLUX. Per-sample IA GPT distributions of FlowEdit and V2C-Edit on MagicBrush and PIE-Bench are shown using histograms and violin plots. On both benchmarks, V2C-Edit substantially reduces the concentration of samples in the low-score region and redistributes more samples toward medium-to-high IA GPT scores. On MagicBrush, this shift is mainly reflected in fewer low-alignment cases and increased density across the intermediate and high-score ranges, although the number of samples receiving the maximum score of 10 is slightly lower than that of FlowEdit. On PIE-Bench, V2C-Edit similarly reduces low-score outcomes while showing a broader concentration in the upper-score region. The violin plots reflect the same redistribution, with less density near the lower end and greater concentration at higher IA GPT values.
Method
CLIP ↑
IAGPT↑
BPPSNR↑
BPSSIM↑
FlowEdit ( tcs=0 )
0.2865
5.56
17.48
0.5694
FlowEdit (Brute force tcs )
0.2892
6.22
16.18
0.5190
V2C-Edit (Ours)
0.2891
6.07
17.44
0.5647
Appendix
Table 6: Comparison with brute-force tcs search on FLUX. For each test sample, all possible values of tcs are exhaustively evaluated, and the result with the highest ILS score is selected, representing the best achievable outcome of FlowEdit’s original mechanism when paired with ILS.
Method
Time (s) ↓
VRAM (GB) ↓
FlowEdit ( tcs=0 )
4.00
36
FlowEdit (Brute force tcs )
79.99
53
V2C-Edit (Ours)
26.44
53
Appendix
Table 7: Average inference time and peak VRAM usage per image, measured on an NVIDIA H100 80GB GPU over 50 test samples.
Figure 9: Qualitative comparison with Brute-Force tcs search. From left to right: source image, Brute Force, and Ours. Both methods satisfy the instruction, but Brute Force introduces noticeably more changes to the background, while our method preserves the unedited regions much more faithfully.
Figure 10: User study interface. Participants select the best result from four randomly ordered edits based on instruction alignment and preservation of unchanged content, given the source image and editing instruction.
Figure 11: Additional qualitative results on FLUX.
Instruction-based image editing aims to modify source content according to textual instructions. However, existing methods built upon flow matching often struggle to maintain consistency in non-edited regions due to denoising-induced reconstruction errors that cause drift in preserved content. Moreover, they typically lack fine-grained control over edit strength. To address these limitations, we propose VeloEdit, a training-free method that enables highly consistent and continuously controllable editing. VeloEdit dynamically identifies editing regions by quantifying the discrepancy between the velocity fields responsible for preserving source content and those driving the desired edits. Based on this partition, we enforce consistency in preservation regions by substituting the editing velocity with the source-restoring velocity, while enabling continuous modulation of edit intensity in target regions via velocity interpolation. Unlike prior works that rely on complex attention manipulation or auxiliary trainable modules, VeloEdit operates directly on the velocity fields. Extensive experiments on Flux.1 Kontext and Qwen-Image-Edit demonstrate that VeloEdit improves visual consistency and editing continuity with negligible additional computational cost. Code is available at https://github.com/xmulzq/VeloEdit.
Text-guided image editing aims to perform a desired edit while preserving source content unrelated to it. Pretrained rectified-flow models enable training-free editing of real images through modifications to their sampling trajectories. However, responses at locations unrelated to the desired edit can still accumulate along the editing trajectory and become visible in the final result. To overcome this, we propose Constrained Edit Fields (CEF), which assigns each spatial location a continuous edit responsibility that quantifies its relevance to the desired edit. CEF estimates edit responsibility directly from the source image when the relevant content is present. For edits whose target content is absent from the source, CEF first generates an unconstrained proposal to reveal its realized spatial support and then estimates responsibility from that proposal. At each editing step, CEF decomposes the base edit field into prompt-induced and trajectory-induced components, enabling edit responsibility to preserve instruction-relevant updates while suppressing unintended trajectory-induced changes. Evaluated on all 700 PIE-Bench examples, CEF achieves state-of-the-art Structure Distance, background LPIPS, and background MSE with both Stable Diffusion 3.5 Medium and FLUX, while retaining competitive instruction alignment. On Stable Diffusion 3.5 Medium, it reduces these metrics over the previous best results by 10.2%, 21.2%, and 48.0%, respectively.
Editing images with pre-trained text-to-image flow models typically requires carefully balancing target alignment with the desired prompt and source consistency with the original image. Existing approaches either rely on inversion-based pipelines or heuristic source-to-target trajectory constructions, which often depend on architecture-specific designs or are sensitive to hyperparameters. In this paper, we propose h-Flow, a training-free and theoretically grounded flow-based editing framework. Inspired by Doob's h-Transform, we reformulate image editing as conditional generation under multiple terminal events corresponding to source consistency and target alignment. We first extend the classical h-Transform from SDE-based models to the deterministic RF framework by constructing an equivalent SDE with identical marginals. Within this formulation, we design dedicated h-functions for source consistency and target alignment, yielding closed-form reconstruction guidance and velocity-based semantic editing signals. We further introduce a velocity orthogonal decomposition to decouple reconstruction and editing directions, enabling a controllable trade-off between the two objectives. Extensive experiments demonstrate that h-Flow achieves effective, robust, and flexible editing across diverse scenarios. The code will be released soon.
Zehui Guo, Zhen Wang, Junwei Shu +3
East China Normal University, Shanghai, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China