acceptodds
Withdrawn from ICLR 2027

TwinEdit: One-Step Image Editing as Single-State Latent Transport

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

One-step text-to-image models have brought image synthesis to real-time speeds, creating a natural foundation for equally fast image editing. Turning this speed into reliable editing remains challenging because a single update must introduce substantial target changes while retaining the structure and content of the source image. Recent one-step methods address this challenge through low-energy transport, post-hoc candidate selection, or localized adaptive editing. However, steering the edit and representing the edit are different problems. To address this gap, we propose TwinEdit, which formulates one-step editing as the transport of a single source-anchored latent state. Source- and target-conditioned latent responses estimate a target-directed path, while only the encoded source is transported along it. The transported state is refined by one target-conditioned denoising step and decoded into a target image. The final edit is then formed along the displacement from the source to this decoded target. Experiments on PIE-Bench show that TwinEdit produces clean, source-faithful edits with strong target alignment while retaining one-step inference efficiency and the lowest measured GPU-memory use in our comparison.

Primary area:
generative models
Keywords:
one-step image editing, text-guided image editing, training-free image editing, inversion-free image editing, single-state latent transport

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2SRaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TwinEdit studies fast image editing with one-step generative models, requiring neither additional training nor inversion. It estimates an editing displacement from source- and target-conditioned responses under antithetic noise, applies it to the source image latent, and performs target-conditioned denoising. Reference-guided spatial blending and coefficient optimization then control the spatial extent and strength of the final edit. Experiments on PIE-Bench evaluate two backbones and include ablations, efficiency analysis, and a user study.

Critical strengths

The paper distinguishes editing-direction estimation from the state carrying the edit. Comparisons across output stages and at approximately matched text alignment provide useful evidence for the role of source anchoring. Experiments on two backbones, random-seed analysis, and runtime and GPU-memory measurements also help assess the method’s practicality.

Critical weaknesses
  1. The novelty boundary is unclear. Existing methods also apply editing displacements to the source latent. The paper should clarify what distinguishes “single-state transport” and identify the individual contributions of antithetic estimation, target-conditioned denoising, and spatial blending.

  2. The benefit of the added complexity is not sufficiently established. The method combines multiple response evaluations, additional denoising, and spatial postprocessing, and runs slower than some one-step baselines. The results mainly demonstrate a preservation–text-alignment trade-off. More direct, controlled comparisons are needed to establish the additional value of these components.

Questions for authors
  1. What specifically distinguishes the method from source-state updating approaches such as ChordEdit? Does the advantage persist when only the displacement estimator changes and downstream processing remains fixed?

  2. How were the default hyperparameters selected? Was tuning performed on data independent of the final evaluation?

  3. Please consistently distinguish sampling steps, batched model invocations, and actual computational cost, and clarify the meaning of “one-step.”

[1] ChordEdit: One-Step Low-Energy Transport for Image Editing. CVPR 2026.

Reviewer_5M9COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper separates the estimation of the editing displacement from the latent state that carries it, applying the displacement estimated under positive and negative noise to the source-image encoding. The method achieve training-free, inversion-free one-step image editing.

Critical strengths
  1. Decoupling the editing displacement from its carrier state is a clear contribution. It makes where the displacement is applied an independently analyzable design choice.
Critical weaknesses
  1. The contribution boundary relative to existing one-step editing methods still needs clarification. The paper should compare against methods such as ChordEdit. Also, the experiments lack comparisons with the latest methods, such as DNAEdit (nips2025).
  2. The novelty of the full pipeline is separative: Positive/negative-noise responses, source-image anchoring, and endpoint-space harmonization together make up the method, but there is no unifying principle explaining why these modules must be combined in this way.
Questions for authors

Please refer to Weaknesses.

Reviewer_3mKnOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a method that formulates one-step image-editing as the transport of source latent state. In particular, source- and target-conditioned latent responses estimate the editing path, while the source latent state is transported along the path. In addition, the transported state is refined by one target-conditioned denoising step and decoded into a target image. Experiments on PIE-Bench show the effectiveness of the proposed method.

Critical strengths

(1) The motivation is well presented of separately considering editing path and editing itself.

(2) The illustrations are mostly clear and intuitive of the editing path estimation, the single-state latent transport and target-conditioned denoising.

Critical weaknesses

(1) The contribution is limited since the core idea of latent transport and opposite noise samples borrow from the literature.

(2) On Page 2, Line 82-84, the same latent displacement need not produce the same visible change when applied from different latent states, because the decoder is nonlinear. However, there is no causal connection between the nonlinear decoder and the phenomenon. In fact, a good decoder should have the property of semantic consistency and the image editing effect should also keep semantic consistency.

(3) On Page 3, Line 157-159, the authors use average of opposite noise samples but do not explain why. In addition, the average causes the problem of partial cancellation. There is also no evidence on how partial cancellation influence the editing process.

(4) On Page 1, Figure 1, the qualitative results of image editing are not good enough as claimed. In particular, consistency is not preserved for the un-edited area. For example, the bowl in row 1 column 4, the forest and fencing in row 2 column 2, the tree and white door in row 2 column 4, the skin texture of the cat in row 3 column 6.

(5) On Page 1, Figure 1, the authors present results of SD-Turbo in the top two rows and SwiftBrush-v2 in the bottom row. There is no

(6) On Page 6, Table 1, TwinEdit shows no strength over SwiftEdit or RRLS if the backbone is the same. In other words, if we split the table into two, one for SD-Turbo, one for SwiftBrush-v2, the performance of TwinEdit is worse than comparison methods. For example, TwinEdit-ET,I (SD-Turbo) & TwinEdit-PT,I (SD-Turbo) vs RRLST,I (SD-Turbo).

(7) On Page 8, Line 399-400, the authors present the quantitative results between TwinEdit and SwiftEdit, but give no explanation on the results.

Questions for authors

No.

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

Related papers

Discussion (0)

Sign in to comment.