The advent of one-step text-to-image (T2I) models offers unprecedented synthesis speed. However, their application to text-guided image editing remains severely hampered, as forcing existing training-free editors into a single inference step fails. This failure manifests as severe object distortion and a critical loss of consistency in non-edited regions, resulting from the high-energy, erratic trajectories produced by naive vector arithmetic on the models' structured fields. To address this problem, we introduce ChordEdit, a model agnostic, training-free, and inversion-free method that facilitates high-fidelity one-step editing. We recast editing as a transport problem between the source and target distributions defined by the source and target text prompts. Leveraging dynamic optimal transport theory, we derive a principled, low-energy control strategy. This strategy yields a smoothed, variance-reduced editing field that is inherently stable, facilitating the field to be traversed in a single, large integration step. A theoretically grounded and experimentally validated approach allows ChordEdit to deliver fast, lightweight and precise edits, finally achieving true real-time editing on these challenging models.
Figures & tables
Figure 2: Comparing ChordEdit (SD-Turbo) against one-step , few-step , and multi-step editing methods on PIE-bench [ 11 ] , evaluating performance on background consistency (PSNR), semantic alignment (CLIP, referring to CLIP-Edited) [ 27 ] , and Runtime. Our method facilitates real-time text-guided editing while yielding highly competitive results.
Figure 3: One-Step Simple drift editing fails. ChordEdit preserves structure. Simple drifts, a direct drift-difference from a one-step model, induce a high-energy, non-smooth vector field, yielding two disqualifying failures: ( i ) severe object distortion and ( ii ) background breakup and spurious structures. Zoomed crops (bottom) highlight the distortions in Simple drifts versus the faithful, photorealistic result of ChordEdit .
Figure 4: Comparison of editing field stability. (a) Multi-step Simple Drift : In conventional multi-step diffusion, the iterative application of the simple drift Δv ensures a stable trajectory. (b) One-step Simple Drift : In distilled models, the naive field Δv(xt,t) is high-energy and volatile. A single, large integration step (solid arrow) accumulates significant error and deviates significantly, as the erratic underlying path (dashed) confirms. (c) Editing by ChordEdit (Ours) : We derive a stable, low-energy Chord Control Field by time-averaging the observable fields R(xτ,t) and R(xτ,t−δ) . This smoothed field facilitates an accurate, single-step transport (red arrow) that faithfully reaches the target xtar .
Figure 5: 2D Toy Example of Distribution transport. Naive residual fields are high-energy and unstable under coarse discretization. ChordEdit computes a low-energy field (Eq. ( 4.5 )) that drives particles straight to the target with minimal deviation, facilitating reliable one-step transport.
Figure 6: Effect of the Proximal Refinement. The refinement step enhances the target editing semantics.
Type
Method
Consistency
CLIP Semantics
Properties
Efficiency
PSNR ↑
MSE 103↓
LPIPS 103↓
Whole ↑
Edited ↑
T-free
I-free
Step ↓
NFE ↓
Runtime(s) ↓
VRAM(MiB) ↓
Multi-step ( ≥ 30 steps)
DDIM + MasaCtrl [ 30 , 3 ]
21.25
8.58
106.59
24.13
21.13
✓
✗
50
100
55.20
12272
Direct Inversion + MasaCtrl [ 11 , 3 ]
21.78
7.99
87.38
24.42
21.38
✓
✗
50
100
79.10
12272
DDIM + PnP [ 30 , 33 ]
21.26
8.42
113.58
25.45
22.54
✓
✗
50
100
28.01
9262
Direct Inversion + PnP [ 11 , 33 ]
21.43
8.10
106.26
25.48
22.63
✓
✗
50
100
28.03
9262
FlowEdit (SD3) [ 14 ]
22.17
7.69
104.81
26.64
23.69
✓
✓
33
33
7.22
17140
Table 1: Quantitative comparison on PIE-bench [ 11 ] . T-free : Training-free. I-free : Inversion-free. The best/second/third results in each numeric column are highlighted with yellow / orange / blue backgrounds, respectively. A comprehensive table with extended metrics (e.g., SSIM, Structure Distance) is available in Appendix.
Figure 7: Comparison of edited results. Real images are in the first column. Prompts are noted under each row.
Figure 8: Qualitative Comparison and Energy Visualization. We compare ChordEdit (Ours) against the naive baseline ( δ=0 , Naive). The naive method’s high-energy field leads to artifacts and background corruption. Our ChordEdit derives a stable, low-energy field, resulting in high-fidelity edits that preserve object identity and non-edited regions. Results shown used SwiftBrush-v2 (first column) and SD-Turbo (second and third columns). Energy plots are computed as E=SC1∑s=1S∑channel(u^ts)2 .
Figure 9: (top) ChordEdit Stability as a function of Integration Steps. We compare ChordEdit ( δ=0.15 , red) against the naive baseline ( δ=0 , blue). The naive field’s energy spikes as step S→1 , confirming its unsuitability for large steps. ChordEdit’s energy remains low. Consequently, the naive method’s background consistency (PSNR) collapses, while ChordEdit maintains high PSNR. (bottom) Analysis of the Perceptual-Semantic Trade-off. Our method ( δ=0 , red) strictly Pareto-dominates the naive baseline ( δ=0 , blue), consistently achieving superior semantic alignment for any given level of perceptual distortion.
Figure 10: Effect of the number of noise samples. This figure qualitatively confirms that increasing the number of noise samples yields negligible marginal returns.
Figure 11: Pareto dominance and Seed robustness. Left: LPIPS–CLIP Pareto fronts [ 37 ] comparing ChordEdit (solid) to the naive baseline (dashed). Shaded regions denote the envelope across seeds. Fronts for ChordEdit with n=1…4 are nearly overlapping and dominate the naive counterparts, indicating negligible marginal returns from multi-noise. Right: histograms of CLIP-Edited and PSNR across 20 seeds for single-noise ( n=1 ). Both distributions are tight (CLIP CoV 0.20% , PSNR CoV 0.07% ), confirming that ChordEdit one noise is effectively insensitive to the random seed.
Method
Naive ( δ=0 )
Ours ( δ=0.15 )
NFE
PSNR ↑
CLIP-Edited ↑
PSNR ↑
CLIP-Edited ↑
w/o prox
21.89
20.83
23.89
21.87
1
w/ prox
21.38
21.96
22.20
22.96
2
Table 2: Ablation of Chord transport field and refinement. Our Chord field drives consistency (PSNR), while the prox step boosts semantics (CLIP-Edited). Full metrics are in the Appendix.
T2I Method
PSNR ↑
CLIP-Edited ↑
Naive
Ours
Naive
Ours
InstaFlow [ 19 ]
22.05
23.05
20.19
21.39
SwiftBrush-v2 [ 5 ]
20.52
22.04
21.06
22.58
SD-Turbo [ 28 ]
21.38
22.20
21.96
22.96
Table 3: Quantitative comparison on different T2I models. Our method (Ours) consistently outperforms the naive baseline across all tested models. Full details are provided in the Appendix.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 12: Raw hyperparameter sweep distribution for CCF analysis. We compare the naive baseline ( δ=0 , blue) against ChordEdit ( δ=0 , red). The plots visualize the trade-off between semantic alignment (CLIP-Edited score) and three standard background preservation metrics: (left) Mean Squared Error (MSE), (center) Peak Signal-to-Noise Ratio (PSNR), and (right) Structural Similarity Index (SSIM). In all three trade-off spaces, the ChordEdit samples (red) occupy a visibly superior region (e.g., lower MSE, higher PSNR/SSIM for a given CLIP score) than the naive baseline, which exhibits a wider, less stable, and inferior performance distribution.
Figure 13: Pareto dominance of the Chord Control Field. These frontiers, derived from the data in Figure 12 , confirm the strict Pareto dominance of ChordEdit ( δ=0 , red) over the naive baseline ( δ=0 , blue) across all three preservation metrics: (left) MSE vs. CLIP, (center) PSNR vs. CLIP, and (right) SSIM vs. CLIP. In every case, the ChordEdit frontier achieves superior semantic alignment (higher CLIP) for any given level of perceptual fidelity, and vice-versa. This empirically validates that the temporal smoothing induced by δ>0 is key to resolving the inferior trade-off inherent in the unstable naive approach.
Figure 14: Noise sample analysis across multiple metrics. This figure expands the analysis in the main paper by plotting the semantic-preservation trade-off (CLIP-Edited vs. MSE/PSNR/SSIM) as a function of the number of noise samples ( n ). The overlapping distributions and tight confidence bands confirm that increasing n>1 provides negligible marginal returns. ChordEdit’s performance with n=1 is already highly stable and robust, validating our default setting.
Figure 15: Ablation study on temporal hyperparameters and step scale. We analyze the sensitivity of ChordEdit to key parameters. (Left two panels): Impact of the main chord time t . Increasing t generally improves semantic alignment (CLIP-Edited ↑ ) at the cost of background preservation (LPIPS-Unedit ↓ ). (Right two panels): Impact of the proximal refinement time tc . Increasing tc robustly improves CLIP score, but also monotonically increases LPIPS distortion. This analysis confirms a clear trade-off space, allowing for the principled selection of our default parameters which balance these competing objectives.
Figure 16: Qualitative analysis of the main chord time t .
Figure 17: Qualitative analysis of the proximal refinement time tc .
Figure 18: Quantitative joint-analysis of δ and λ . These 3D surface plots show the trade-off between semantic alignment (CLIP) and background preservation (LPIPS, MSE, SSIM, PSNR). The naive baseline ( δ=0 , the front edge of each plot) is Pareto-inferior, suffering from low CLIP scores and high distortion (high LPIPS/MSE). Increasing δ (our temporal smoothing) robustly and monotonically improves all preservation metrics. Increasing λ (step scale) robustly increases semantic strength (CLIP) at the cost of preservation. Our default parameters are chosen from this smooth, well-behaved trade-off space.
Figure 19: Qualitative analysis of the temporal window δ . We fix other parameters and vary δ . The naive baseline ( δ=0.00 ) consistently fails, producing severe artifacts, distortions, and structural collapse, especially on complex semantic changes (e.g., ’tiger → cat’ or ’apple → cat’). Our Chord Control Field ( δ>0 ) immediately stabilizes the edit, resolving these failures. A value of δ=0.15 demonstrates a robust balance between edit stability and semantic strength, validating its choice as our default.
Figure 20: Qualitative analysis of the step scale λ . We vary λ while keeping other parameters fixed. λ functions as an intuitive ’edit strength’ controller. Small values (e.g., λ=0.8 ) result in subtle, under-edited images (’mountain’). As λ increases, the intensity of the target semantic (’volcano’) becomes progressively stronger. This provides a simple and predictable knob for users to modulate the edit’s impact.
Type
Method
Struct.
Background Preservation
CLIP Semantics
Efficiency
Dist. 103↓
PSNR ↑
MSE 103↓
SSIM 102↑
LPIPS 103↓
Whole ↑
Edited ↑
Runtime(s) ↓
Step ↓
NFE ↓
VRAM(MiB) ↓
Multi-step ( ≥ 20 steps)
DDIM + MasaCtrl
28.79
21.25
8.58
80.11
106.59
24.13
21.13
55.19
50
100
12272
Direct Inversion + MasaCtrl
24.46
21.78
7.99
81.74
87.38
24.42
21.38
79.10
50
100
12272
DDIM + PnP
28.20
21.26
8.42
78.90
113.58
25.45
22.54
28.01
50
100
9262
Direct Inversion + PnP
24.27
21.43
8.10
79.52
106.26
25.48
22.63
28.03
50
100
9262
FlowEdit (SD3)
12.34
22.17
7.69
83.54
104.81
26.64
23.69
7.22
33
33
17140
Appendix
Table 4: Quantitative comparison of our method against other editing methods on PIE Bench.
Figure 21: User Study Results. Aggregated human preference rates from a four-way blind comparison, matching the data cited in the main paper. ChordEdit was the clear winner in both Semantic Alignment (42.5%) and Preservation Quality (48.3%), demonstrating its superior overall performance.
Symbol
Description
t∈[0,1]
Time variable for the probability flow. t=1 corresponds to the source data distribution, and t=0 to the target noise distribution.
xt
The image state at time t .
c
The text condition (prompt).
csrc,ctar
The source and target text prompts, respectively.
xsrc,xtar
The source image ( x1 ) and the desired target image ( x0 ), respectively.
pt(x∣c)
The probability distribution of xt at time t conditioned on c .
Recent one-step text-to-image (T2I) models enable efficient image synthesis, opening up new possibilities for interactive Web-based image editing. However, existing one-step editing methods often construct a global transport field through simple vector operations between conditional predictions. While such strong transport can effectively drive the image toward the target semantics, it may be released along unreliable directions or over non-edited regions, leading to structural distortions within the editing region and background drift elsewhere. More importantly, stabilizing the editing process through global smoothing or uniformly weakening the transport field often makes large semantic changes overly conservative. In this work, we revisit one-step image editing from the perspective of the tension between semantic strength and transport stability, and identify a key insight: strong semantic transport is not inherently unstable. What matters is whether it is applied along reliable directions and at reliable spatial locations. Motivated by this insight, we introduce PreciseEdit, a model-agnostic, training-free and inversion-free one-step editing framework that strengthens target-semantic alignment while preserving background fidelity and structural consistency. We further propose a dual-noise-scale role decomposition strategy: a higher-noise scale produces the strong target-directed transport used for actual editing, whereas a lower-noise scale jointly serves as a probe for directional agreement and spatial structure. Extensive experiments on PIE-Bench demonstrate that PreciseEdit achieves state-of-the-art performance and effectiveness on various editing tasks.
Jingyi Wang, Xiao Lu
Hainan University School of Computer Science and Technology Haikou, China · Shanghai University School of Mechatronic Engineering and Automation Shanghai, China
One-step image editing is important for making text-guided editing fast, practical, and easy to deploy, but its underlying mechanism is still not fully understood. We revisit ChordEdit through reproduction, ablation, and simplification. Our analysis shows that a) the chord window δ largely acts as an effective timestep shift from t to t−δ; b) chord transport acts on high-noise images and mainly performs low-frequency semantic editing; and c) proximal alignment acts on low-noise images and complements it by adding high-frequency target details. In this view, ChordEdit naturally decomposes editing into a coarse low-frequency transport stage and a fine high-frequency alignment stage. These findings suggest a path toward prompt-conditioned dynamic timestep selection for adaptive image editing. All code and results can be found at link.
Recent one-step text-to-image (T2I) models enable efficient image synthesis and provide new opportunities for real-time image editing. However, existing one-step editing methods primarily rely on text conditioning for semantic transformation, lacking explicit spatial control over where to edit. More importantly, even when spatial constraints are introduced, these methods often struggle to achieve strong and stable semantic modifications within the target regions. In this work, we revisit one-step image editing from a spatially controlled perspective and identify two key challenges: discovering editable regions and achieving effective localized semantic transformation. We reveal that existing methods perform global semantic transport, which limits high-intensity local editing under the one-step setting. To address this issue, we propose WhereEdit, a framework that reformulates one-step editing as localized adaptive editing. WhereEdit automatically identifies semantically relevant regions from internal model features and applies adaptive local modulation to enhance target-region editing while preserving non-target areas and structural consistency. Experiments on the PIE-Bench benchmark demonstrate that WhereEdit consistently outperforms existing one-step image editing methods, achieving superior editing quality while maintaining the efficiency of one-step generation. Additional experiments with region-level supervision further highlight the importance of explicit spatial reasoning for high-quality one-step image editing.
Ming Hu, Mingyu Dou, Jianfu Yin +5
Xi’an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Zhongnan Hospital of Wuhan University +1