acceptodds
Under review as a conference paper at ICLR 2027

One Evaluation Is Enough: Inversion-Free Image Editing with Trajectory Correction

Abstract

Real-image editing commonly relies on inversion and iterative sampling, while efficient one-step editors often stabilize a large update by estimating and averaging editing fields at multiple noise levels. We ask a stricter question: is a single temporal field evaluation sufficient for reliable inversion-free editing? Our answer rests on a previously underused duality in a variance-preserving (VP) one-step model: one source-target query at a shared noisy state simultaneously determines a local angular semantic tangent and a clean endpoint displacement. We integrate the tangent over the analytically known remaining VP angle to obtain a nominal one-step transport, and use the endpoint displacement to correct its long-horizon extrapolation through a constrained projection. This transport-first, correction-second construction yields Endpoint-Calibrated Transport (ECT), a closed-form, training-free editor requiring only one noise level, two text conditions packed into one denoiser batch, and one latent update. ECT requires no image inversion, trajectory integration, second temporal observation, or iterative optimization, and can be directly applied to pretrained epsilon-prediction VP one-step generators. Across the reported benchmark settings, ECT delivers stable gains in editing quality within an extremely short editing time.

Primary area:
generative models
Keywords:
Inversion-free image editing, Training-free editing,One-step diffusion models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Q7x8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Endpoint-Calibrated Transport (ECT), a text-guided image editing method built on pretrained one-step diffusion models such as SD-Turbo and SwiftBrush-v2. It enables efficient editing without training or inversion. On PIE-Bench and SNR-Bench, the authors report that ECT reduces inference latency by approximately 16%–20% relative to the ChordEdit configuration used in the comparison, while also lowering CPU utilization and resident set size (RSS). It additionally improves several background preservation metrics and image quality evaluations.

Critical strengths
  1. The method requires neither additional training nor backpropagation. It performs editing through forward predictions from pretrained one-step diffusion models and analytical calibration, and its effectiveness has been demonstrated on SD-Turbo and SwiftBrush-v2.

  2. Compared with baseline methods such as SwiftEdit, the proposed method offers higher computational efficiency.

Critical weaknesses
  1. Does ECT limit editing strength and flexibility, encouraging the model to make as few changes as possible? Could the improvement in background fidelity result from more conservative editing?
  2. The analytical derivation of ECT relies on one-step variance-preserving (VP) diffusion backbones that use epsilon prediction. Does this dependence limit its applicability to other models?
Questions for authors

Could the authors clarify how general ECT is intended to be? I would like to understand whether ECT can be adapted to models beyond SD-Turbo and SwiftBrush-v2 while maintaining good performance.

Reviewer_DYUkOpenReview ↗
Rating
3(+2)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ECT, a training-free, inversion-free editor that needs a single denoiser evaluation. At a shared noisy state z_t built from the source, the source and target epsilon-predictions give an angular tangent g_t and an endpoint displacement d_t that are collinear (Eq. 21). They integrate the frozen tangent over the remaining VP horizon to get a "transport" a_t, then apply a proximal correction projecting it onto a half-space, yielding a closed-form update (Eq. 22): x_edit = x_src + (tau_t/alpha_t - lambda)(x_hat_0^e - x_hat_0^s). Experiments on PIE-Bench and SNR-Bench with SD-Turbo and SwiftBrush-v2 against 12 baselines.

Critical strengths

The derivation is clean: an analytic single-evaluation editor from collinearity plus a KKT projection, with no numerical solver. The efficiency gains over ChordEdit are concrete and consistent with halving the condition branches (latency ~16-20%, CPU ~42-48%). Seed controls and baseline coverage are more thorough than I expected, and the authors are honest about trade-offs.

Critical weaknesses
  1. The method reduces to one tuned scalar. Since g_t = -(1/alpha_t) d_t, both the transport and the correction are scalar multiples of the same d_t, so Eq. 22 is just x_src + s*(x_hat_0^e - x_hat_0^s). That is naive one-step editing with a tuned coefficient, and this trivial baseline is never in the main tables — so nothing shows the transport-then-correct framing beats picking s. Appendix A half-concedes it ("does not establish that rescaling always reduces its error").
  2. Narrow reach. The collinearity only holds for VP epsilon-prediction models; the paper admits no velocity-prediction or rectified-flow coverage, which excludes the backbones (FLUX/SD3) most editing work now uses.
  3. Alignment never improves. CLIP drops in every matched pair except a tie on PIE-Bench SB-v2 (29.70 vs. 29.70) — e.g., 30.60 -> 30.20 on PIE-Bench SD-Turbo and 24.88 -> 24.31 / 25.76 -> 25.32 on SNR-Bench — so preservation gains are partly bought with the metric that measures instruction-following. The qualitative evidence is also thin: Figure 3 has only four examples and all four reappear in the appendix Figure 9, and the flower edit misses the instruction (flowers land near the face/suit, not in the mouth).
Questions for authors
  1. Can you add a naive one-step baseline x_src + s*(x_hat_0^e - x_hat_0^s), with s tuned on a validation split, to Tables 1-2? Without it, how is any quality gap attributable to the derivation rather than scalar tuning?

  2. What is tau_t/alpha_t at the default tau_t = 0.75 for both backbones, and how close to 1? This would separate the "angular transport" from the lambda shrinkage.

  3. Does the collinearity in Eq. 21 hold even qualitatively for a velocity-prediction or rectified-flow model? If not, how much of current editing does ECT actually cover?

Reviewer_zbxkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

ECT is a training-free image editor for epsilon-prediction VP one-step generators. It evaluates the source and target prompts at the same noisy state, uses their prediction difference to estimate angular transport, and then shrinks the update using the predicted endpoint difference. Because these two signals are collinear, the final update is a scaled endpoint difference added to the source latent. The authors test it on PIE-Bench and SNR-Bench with SD-Turbo and SwiftBrush-v2, and report runtime, human ratings, and parameter and seed studies. The question is whether one noise level is enough to get a useful editing result at lower cost than a two-level editor.

Critical strengths
  1. Using one noise level instead of two is a useful simplification to investigate. ECT uses two conditional branches rather than four, and Tables 1 and 2 report latency reductions of 15.7–20.0% against ChordEdit on matched backbones. I appreciate that Appendix E counts calls, noise levels, branches, and latent updates separately. If the quality results use the stated budget, the measured speedup makes this practically relevant.

  2. The seed study is a useful check. Appendix J.1 tests eleven noise seeds on PIE-Bench with SD-Turbo, and every seed beats the matched ChordEdit reference on all four preservation metrics. The evaluation also covers two benchmarks and two backbones. The authors acknowledge the mixed preservation results on SwiftBrush-v2 and avoid claiming statistical significance from the human-study percentages alone. This gives a more useful picture than a few selected examples.

Critical weaknesses
  1. I couldn't reconcile Table 1 with Table 3(c). Section 4.1 and Appendix J.2 say the main results use one batched denoiser call, without extra-NFE refinement. But the “w/ extra NFE” rows in Table 3(c) match the main results at the displayed precision: PSNR/LPIPS/CLIP are 23.7/109.0/30.2 for SD-Turbo and 22.9/122.8/29.7 for SwiftBrush-v2. The “w/o extra NFE” rows give different numbers. Appendix J.2 says refinement is a separate experiment, but doesn't explain why these results coincide. This may be a labeling or configuration issue; the table alone doesn't show that the main runs used an extra call. Even so, I need this cleared up to judge the paper's central quality–efficiency claim.

  2. I'm not sure what the two-stage construction adds beyond choosing a scalar editing strength. Equation (22) gives an update of kappa times the endpoint difference, with kappa = lambda × (theta_t/sigma_t − zeta). At fixed t, the transport factor and zeta can be absorbed into the free coefficient lambda. Appendix B also says that fixing zeta amounts to choosing an implicit sample-dependent budget. Table 3(b) and the parameter sweeps compare different strengths, but don't show an advantage over directly tuning kappa on the same one-query residual. The single-noise-level editor may still be useful. What remains unclear is why its transport and correction need to be treated as separate contributions.

Questions for authors
  1. Which configuration, sample set, and denoiser-call count produced each of the corresponding rows in Table 1 and Table 3(c)? If there is a labeling error, which labels or values need correction?
  2. At fixed t, what does ECT offer over directly tuning kappa on the same endpoint difference? Does the parameterization give a better selection rule or transfer more reliably? A matched scalar-strength comparison would help answer this.
  3. How many participants and distinct cases were included in Figure 6, and how many ratings did each case receive? Appendix G describes the aggregation but leaves out these counts. I need them to judge how much weight to give the text-alignment preference.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.