acceptodds
Withdrawn from ICLR 2027

LatentSwitch: One-Step Text-Guided Image Editing via Switchable Noisy Latents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Text-guided image editing aims to modify an input image according to a target prompt while preserving its structure and layout. While diffusion models provide strong generative priors for this task, most prior methods rely on iterative denoising, resulting in high inference cost and unintended changes to unedited regions. To address this, we propose LatentSwitch, a one-step diffusion framework based on the observation that the same latent can produce different semantic contents under different text conditions. Specifically, given a source image and a target prompt, LatentSwitch predicts a shared (switchable) latent that preserves source-specific visual information. Under the source prompt, this latent reconstructs the input image; by switching only the text condition to the target prompt, the pretrained diffusion denoiser directly produces the desired edit in a single step. The latent noise level is adaptively chosen based on the required editing strength, enabling low-noise latents to preserve source details for localized edits and higher-noise latents to support larger semantic changes. Experiments show superior editing quality with substantially lower inference cost than prior image editing methods.

Primary area:
generative models
Keywords:
Image Edit, Diffusion

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Gq3YOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes LatentSwitch, a one-step text-guided image editing method. Given a source image, a mask, and a target prompt, LatentSwitch predicts a scaled noise term and corresponding timestep by matching to the scheduler’s noise level. The resulting noisy latent is then denoised once by a frozen SD-Turbo under the target prompt. Training pairs are synthesized by branching a reference image and a target image from a shared noisy latent under two prompts, then compositing them with a mask. Training uses L1, perceptual, adversarial, a timestep loss, and a proposed distribution-matching loss L_dm. Results are reported on MagicBrush and HumanEdit.

Critical strengths

None

Critical weaknesses
  • Limited novelty over existing one-step editors. One-step editing with a network that predicts inversion noise for SD-Turbo is already established, for example by SwiftEdit. The new parts are the shared-latent pair synthesis, std-based timestep estimation, and L_dm. The paper does not show why these contributes to a significant advance rather than incremental design choices.
  • The construction of I_src is flawed and the mask is not well-defined.
    • Building I_src by pixel-blending I_tgt and I_ref (Eq. 1) creates artifacts whenever the target introduces an object that cannot blend seamlessly into the reference, for example turning a dog into a cat.
    • It is unclear whether M marks the new object or the old one.
  • The one-step claim is unclear. Section 3.2 only predicts a noisy latent. The paper never explains how that noisy latent is mapped to a clean Z_0 in a single evaluation. Meanwhile, Table 2 relies on multiple iterations to get its best results.
  • Results. The user study has only 13 participants and no protocol details.
  • Presentation.
    • Fig. 3 draws arrows labeled with text conditions from the VAE decoder to I_ref and I_tgt. This is misleading, since text conditions apply to the diffusion model, not the VAE.
    • Notation is redundant: M† (line 248, Eq. 7) is just 1−M.
Questions for authors

I don't think there are any questions I can ask the authors from which the authors can answer and change my decision as to me this work does not introduce any new technical contribution and there are several important flaws in method design (see critical weaknesses).

1 comment
Reviewer_Gq3YOpenReview ↗

Official Comment to Rebuttal Response

I thank the authors for providing the response. My main concerns about novelty and technical contributions have not been well addressed. Specifically, the rationale of predicting a noisy latents share the same idea with SwiftEdit with minor adjustment of predicting a noisy latent at particular time-step. However, there is no theoretical justification or empirical analysis to show that the inverted noise can match well with the target given that the network receives no information that reveals about the timestep (source latent, mask, and target prompt).

Also claiming one-step editing again is incorrect as the editing process is fundamentally multi-step denoising starting at a particular noisy latent timestep on the sampling trajectory.

In summary, the presentation and notation is minor and fixable; the author did well in actively refining and updating the manuscript. However, my biggest concern is still the paper's soundness and its key contribution in the context of one-step image editing. Hence, I maintain my original score as Clear rejection. I hope my feedback and comments would be useful for the authors to further improve the clarity as well as paper's soundness for future submission.

Best Regards,

Reviewer Gq3Y.

Reviewer_hEAMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes LatentSwitch, which reduces text-guided image editing to a single diffusion prediction by predicting a switchable noisy latent that can switch between source and target conditions. The idea is simple, and the efficiency and user preference results are appealing. However, the paper does not provide sufficient ablations of its core mechanism, “adaptive noise level,” and the reconstruction quality under the source prompt is also not strong. More seriously, the CLIP values in a key ablation table contradict both the claims in the text and the results in the main table, directly affecting the credibility of the experiments.

Critical strengths
  1. The core idea is intuitive and has practical value. Interpreting the same latent as a shared representation whose semantic realization can be switched through text conditions provides a clear modeling perspective for one-step editing.
  2. Inference efficiency is clearly better than that of most iterative editing methods.
  3. Text alignment and subjective preference in the main results are competitive.
Critical weaknesses
  1. Table 5 contains a serious internal contradiction. The table reports a one-step CLIP score of 0.3098 “without distribution matching,” whereas the value printed “with distribution matching” is 0.2413; yet the text claims that this component improves performance, and the main table reports 0.3113 for the same final configuration.
  2. The incremental contribution over the closest prior work is insufficient, and the comparisons are incomplete. The core framework is essentially the same as SwiftEdit, while InverFill (CVPR 2026) and TurboFill (CVPR 2025), which likewise use “mask + one-step inversion + few-step model,” as well as few-step editing methods such as iCD and TurboEdit, are all absent from the experimental comparisons. Moreover, SwiftEdit does not rely on a user-provided mask by default, whereas LatentSwitch uses a mask, making a direct comparison unequal. The authors should compare against masked one-step/few-step methods under the same mask conditions and discuss precedents for target-aware inversion and timestep selection, such as ENM Inversion and TiNO-Edit.
  3. The paper does not provide decisive ablations of adaptive timesteps or noise-level prediction. At a minimum, it should compare fixed timesteps, random timesteps, adaptive timesteps, oracle timesteps, removal of timestep prediction, and different timestep ranges, and report how prediction error varies with editing strength.
Questions for authors
  1. “The switchable noisy latent reconstructs the input image under the source prompt” is an important part of the method's definition, but the reconstruction results under the source prompt are not strong. This leaves open the question of how much source-specific information is actually preserved.
  2. The one-step advantage does not establish an absolute lead in speed. A more accurate conclusion is therefore that this paper offers a competitive trade-off among speed, target alignment, and content preservation, rather than universally achieving the highest efficiency.
Reviewer_5ynxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes LatentSwitch, a framework for one-step text-guided image editing using a frozen diffusion model. The method trains a predictor model to estimate a noisy latent and its associated timestep from a source image, editing mask, and target prompt. Training examples are generated by denoising a shared noisy latent under different text conditions and blending the resulting images. The method reports competitive editing performance with reduced inference cost compared to several multi-step baselines.

Critical strengths

I like the general idea of training a predictor to find a noisy latent from which a frozen diffusion model can generate an edited image in a single step. I also like the synthetic training procedure, which avoids the need for a dataset of paired editing examples. The results are promising, especially in terms of balancing editing quality with preserving the original image.

In particular, I like the idea of having the model learn an appropriate timestep for each edit. Intuitively, different edits require different noise levels: some edits require stronger corruption to remove existing structure, while others need weaker corruption to preserve details. I also find the idea of approximating this through supervision on randomly sampled noise levels interesting. My understanding is that by exposing the predictor to a range of noise levels, some potentially too weak or too strong for a given edit, the hope is that the model will learn to associate the required editing strength with an appropriate timestep.

Critical weaknesses
  1. Insufficient justification of the adaptive timestep prediction mechanism:

My biggest concern is the motivation and supervised training of the adaptive timestep mechanism, which is presented as a central contribution. The method generates training examples by sampling a timestep and producing reference and target images from a shared noisy latent. The prediction network is then supervised to match the noise magnitude associated with this sampled timestep.

The paper explains intuitively that the noise level is generally proportional to the level of semantic divergence in generation. The paper further suggests that for each edit, there is an appropriate noise level at which to corrupt the original image and continue the generation with the target prompt. An edit that modifies structure might need a high noise level, while one that modifies fine detail might need a lower noise level. However, my concern is with how the model is trained to predict the appropriate timestep. During training, the timestep is sampled randomly, and the source and target images are then generated from the same noisy latent using different text prompts. The predictor is trained to match the noise level that was used to generate this pair. I understand the intuition: since the sampled timestep influences how different the two images can be, learning to predict it might help the model associate larger edits with higher noise levels. The problem is that the timestep used to generate a training pair is not necessarily the best timestep for performing that edit. For example, an edit that only requires changing the color of an object could still be generated using a high noise level, even though a lower noise level might be more appropriate for preserving the original image. Similarly, an edit that requires a substantial structural change could be generated using a noise level that is too low. In both cases, the predictor is supervised to match the sampled timestep, even though we don't know whether that timestep is appropriate for the edit. While I understand why this might work as an approximation, I don't think the paper sufficiently explains why learning to predict the timestep used to generate a training pair should lead the model to select an appropriate timestep during inference. Even if the predictor learns to estimate the generating timestep accurately, it is not clear why that estimate should be close to the best timestep for performing the edit. Figure 6 does show the expected relationship between the predicted noise level and the amount of semantic change required by the edit, which is encouraging. However, I don't think this is enough to validate the supervision strategy. It shows that the model has learned to associate larger edits with higher noise levels, but does not establish whether this supervision actually improves editing performance. I would like to see a stronger justification for why this training strategy should lead to the desired behavior. Additionally, there are critical ablation experiments that are missing. What would happen if we omitted the timestep supervision entirely? Would the predictor network learn on its own the optimal noise level for each edit, or would the training collapse? A comparison with a fixed-timestep baseline would also help establish whether the adaptive timestep mechanism is actually contributing to the reported improvements. Moreover, the sensitivity to the timestep-sampling distribution is not studied. Since the training target depends directly on the sampled timestep, the choice of sampling distribution could significantly affect the learned behavior. For example, perhaps most edits require moderate noise levels, and therefore a logit-normal distribution that focuses more mass around intermediate timesteps would lead to a better approximation of the optimal timestep for each edit. I consider these missing controls particularly important because they concern the central mechanism of the proposed method. The ablations provided on the mask generation method and distribution matching loss are secondary. Bottom line: I would like to see two things:

a. An ablation removing the timestep supervision loss, to establish whether this supervision is actually necessary for the method to work. A fixed-timestep baseline would also be valuable for showing whether adapting the timestep actually helps.

b. A stronger justification for why supervising the randomly sampled timestep should lead the model to select a timestep suitable for the required edit. Without these, I don't think the paper sufficiently justifies one of its central contributions.

  1. Missing relevant prior work and incomplete comparisons

ChordEdit (CVPR 2026 - feb 2026) also addresses one-step text-guided image editing and evaluates with SD-Turbo, making it an especially relevant method for comparison. I would like to see a comparison with ChordEdit, or at least an explanation of why it was not included.

  1. Insufficient reporting of the user study

The user study involves only 13 participants. I would like to see the study extended to include more participants. Additionally, there is little to no information about the study itself beyond stating that participants were asked to select the best image among five outputs. The authors should provide the number of evaluated editing examples, how the examples were selected, and a representative subset of the questions or editing examples.

Questions for authors

I find the idea and results promising, but I am currently leaning toward a borderline rejection because the paper does not isolate the contribution of the adaptive timestep mechanism. The most important evidence for me would be an ablation removing the timestep supervision, together with a stronger justification for the supervision strategy. I would also like the authors to address the missing ChordEdit comparison and provide a better-documented user study. Convincing results on the timestep ablations and the comparison with ChordEdit could move me toward acceptance.

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

Related papers

Discussion (0)

Sign in to comment.