ZZEdit: ZigZag Trajectories of Inversion and Denoising for Zero-shot Image Editing

1. The authors claim the essential of "locating a structure-preserving while sufficient-for-editing point as the editing pivot" and give the illustration in Fig.2. However, the searching process in lines 1-10 (Part I) in Algorithm 1 is not verified. Specifically, why simply using $||\epsilon^{tgt}-\epsilon^{null}||>||\epsilon^{src}-\epsilon^{null}||$ as the condition to judge the pivot timestep? The authors are advised to provide a theoretical analysis to prove it.

2. Similarly, I did not fully understand the potential motivation behind the ZigZag operation (inversion-denoising loops) around the pivot timestep since there is no derivation or supervision signal to regularize. Could the authors provide more detailed illustrations (like mathematical ones) to prove why this operation can handle the aforementioned problem?

3. Something I think that should be corrected is that the "inversion-and-then-editing" pipeline does not have to require achieving the final latent $z_T$, many methods including P2P simply perform inversion to around 0.7T and then denoising.

4. Since the authors have used large language models to evaluate the editing performance and determine the pivot, I advise the authors can add some LLM-related metrics [1] [2] in the comparison evaluation experiments. Also, more inversion methods should be included in the comparison such as [3],[4], etc.

[1] TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering

[2] Diffusion-Model Based Image Editing: A Survey

[3] Edict: Exact diffusion inversion via coupled transformations

[4] Effective real image editing with accelerated iterative diffusion inversion

• Results Assessment

  • The results do not appear promising. For instance, the dog in the main figure does not resemble a Pixar animation style.
  • Additionally, the road lane in ZZEdit with PnP + DDIM is not well-preserved, which the authors identify as a drawback of P2P + DDIM.
  • Furthermore, it is unclear whether this approach can handle cases requiring significant modifications, such as transforming a flamingo into a giraffe.

• Pilot Experiment and Pivot Point Concerns

  • The interpretation of the pilot experiment and their method for identifying a suitable pivot point are unconvincing. In my opinion, it’s hard to distinguish between the effects of the text and the inherent characteristics of the diffusion process, from coarse to fine detail. Early steps start with similar noise, making the images naturally alike, and in later steps, the fine details reduce differences again.
  • Even if this concern is addressed, it’s unclear how lines 240-241 relate to their method for identifying the pivot point. To me, this line suggests that the point where the guidance degree reaches its maximum should be the optimal pivot point. If that’s the case, why? Additionally, does the method you use to identify the optimal point—where there’s a stronger response to the target text condition—actually coincide with the point where the guidance degree is maximized? Further explanation would be helpful.

• Suggestions for Pivot Point Analysis

  • More analysis on designating pivot points would be helpful. Is the stronger response to the target text condition compared to the source text condition consistently observed beyond the initial point (i.e., the chosen pivot point), or is it stochastic?
  • I’m curious if the results in Figure 6 align with our intuition. For example, were images with a pivot close to 1.0T cases that required severe modification, while images with a pivot at 0.4T were cases where minimal modifications, like color changes, were sufficient? Such an analysis could potentially support the method for determining the pivot point.

• Zigzag Process Concerns

  • The motivation or theoretical explanation for the zigzag process is lacking.
  • I can see the clear improvements brought by the zigzag process, but why does it work? Further explanation would be helpful.
  • I’m curious about the results of combining this approach with the editing method used in [1](i.e., editing only through inversion&sampling) . If the guidance from the zigzag process is truly effective, it should work well in conjunction with the approach used in [1].

[1] CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models

1. Writing improvements are necessary. The paper’s structure is difficult to follow, and the purpose of the pilot analysis is unclear and doesn’t connect with the method. Additionally, the related work section could better emphasize the distinctions between previous research and the proposed approach to clarify the paper's positioning.
2. The experiments appear insufficient to fully support the proposed claims. Based on the results, it’s challenging to conclude that the proposed method enhances editability and fidelity.

1. In line 6 of the algorithm, is there a possibility that, in certain cases, the value of $P$ could be very small? If so, this would significantly restrict the editing space.

2. In Table 2, there is an unusual phenomenon: when using P2P, the performance of ZZEdit is much worse than PnP-inv. However, when switched to PnP, the performance of ZZEdit suddenly improves drastically, far surpassing PnP-inv. The reviewer requests a detailed and logical explanation for this discrepancy, as well as an in-depth analysis. From an ablation perspective, what is actually driving this improvement—PnP or ZZEdit? If ZZEdit is genuinely effective, why does it perform worse than PnP-inv when using P2P?

3. Almost all experiments focus on ablation studies or comparisons with baselines. The reviewer believes it is more important to include comparisons with state-of-the-art (SOTA) image editing methods, especially given the many training-based methods from 2024 that have already demonstrated mature and effective results. In this context, what is the motivation for introducing ZZEdit, which seems to exhibit only average performance? The reviewer suggests that the authors include direct comparisons with 4 to 5 SOTA methods from 2024, rather than selectively comparing with training-free methods.

4.The conclusions drawn from Table 1 do not strongly support the significance of the zigzag component. When zigzag is set to 0, the consistency with the original image is optimal. However, introducing zigzag causes a significant drop in 80% of the metrics, with only a minor improvement in the CLIP score. These results fail to justify the value of zigzag, as its gains seem minimal while its drawbacks are considerable. Since zigzag is the primary contribution of this work, without it, the paper mainly consists of engineering techniques.

5.The paper only addresses the simplest type of image editing, which is object replacement. It’s worth noting that image editing includes around a dozen types of tasks. For a method that claims to be a general-purpose image editing approach, it should at least demonstrate additional cases, such as object removal, action editing, and background replacement—three very common types of edits. If this method is only capable of object replacement, it should not be presented as a general-purpose editing solution.