Object Fusion via Diffusion Time-step for Customized Image Editing with Single Example

1. The paper is not well-written. The abstract and introduction involve too many details. The introduction should mainly describe the motivation and high-level idea of the proposed method.
2. The discussion on the existing works is insufficient. Subject-driven image editing is a very broad topic and the authors should have a more thorough discussion. For example, object addition/insertion is also known as image composition, which should be discussed in introduction or related work. And more image compostion baselines should be compared.
3. The proposed method is trivial and lacks technical novelty. The main novelty is time-specific and location-specific lora, which is also trivial and straightforward. Time-specific lora has been widely used in previous works. Since we need to combine source image and reference image, it is natural to use location-specific lora.
4. The user needs to specify the mask to be edited, which is laborious and troublesome. If the object shape needs to be slightly changed in the source image, how to address this issue?
5. The proposed method needs to optimize the fusion coefficients for each pair of source image and reference image, which is very tedious. What about the time cost of three steps? There is not efficiency comparision between the proposed method and baselines. Is the proposed method more time-consuming than the baselines? If we change a reference object, it is necessary to optimize the fusion coefficients again?
6. In the experiments, the comparision with existing methods is not adequate and more related methods should be compared. Moreover, the authors conduct experiments on the self-collected datasets, so the experimental results are not convincing.
7. For the evaluation metrics, the authors only report the similarity metrics (e.g., CLIP, DINO) between the inserted/replaced object and the reference object. However, the quality of the whole image should be measured from multiple aspects, for example, whether the inserted object is compatible with the background in terms of illumination, viewpoint, and so on. Therefore, the evaluation is not convincing.
8. There is no discussion on the failure cases. What about the ability scope of the proposed method? Based on my understanding, the proposed method has no ability to adjust the view or pose of inserted object. If so, the ability of proposed method is very limited. There is no showcase for the challenging cases.

• The proposed TimeFusion seems more like a trick than a legitimate paper or piece of work. First, it can only handle object fusion by training object-centric LORAs and combining them, which is too limited. In contrast, previous LORA techniques can manage both object and style fusion. Additionally, there are numerous methods focused on object generation and editing through inpainting, which can also achieve visually pleasing results. Given this perspective, I am uncertain why the authors emphasize LORAs for this task. Second, there are many follow-ups to Anydoor that maintain fidelity well, such as CustomNet[1] and IMPRINT[2]. I wonder why the authors did not compare their work to these methods. Third, from my perspective, this work essentially performs a spatially aligned copy-and-paste from the reference image to the target image. For example, in Fig. 3, the ‘clock’ in the target image nearly occupies the same spatial position as in the reference image.
• I doubt the possibility of combining multiple LORAs together, either in the case of spatial alignment or spatial misalignment.

[1] Multi-LoRA Composition for Image Generation

[2] CustomNet: Zero-Shot Object Customization with Variable-Viewpoints in Text-to-Image Diffusion Models

[3] IMPRINT: Generative Object Compositing by Learning Identity-Preserving Representation

1. Although the motivation is that time variant weighting is important, the actually weight is learned. Can one design a rule and just vary the weighting during sampling? i.e. t is large for ref during beginning and slowly decrease? It will be nice to show such results.
2. No quantitative results for ZipLORA, Lora Switch etc.
3. It would be beneficial to see why using reconstruction objective can lead to good editing sampling.

1. The experiments primarily concentrate on basic tasks such as object addition or replacement, and do not include complex scenarios like dynamic backgrounds or scenes with multiple overlapping objects. This could limit the applicability of the results in more complex situations.
2. Although the computational requirements are discussed, there is no analysis of their practical implications. For example, it's unclear whether the computational load could prevent the method from being used in real-time or large-scale applications.
3. The extension from existing LoRA methods is not fundamentally new, as it mainly modifies the fusion process. The introduction of time-step-specific coefficients is a minor improvement and may not warrant a separate method unless its broader applicability is demonstrated.