Real-World Image Variation by Aligning Diffusion Inversion Chain

Recent diffusion model advancements have enabled high-fidelity images to be generated using text prompts. However, a domain gap exists between generated images and real-world images, which poses a challenge in generating high-quality variations of real-world images. Our investigation uncovers that this domain gap originates from a latents' distribution gap in different diffusion processes. To address this issue, we propose a novel inference pipeline called Real-world Image Variation by ALignment (RIVAL) that utilizes diffusion models to generate image variations from a single image exemplar. Our pipeline enhances the generation quality of image variations by aligning the image generation process to the source image's inversion chain. Specifically, we demonstrate that step-wise latent distribution alignment is essential for generating high-quality variations. To attain this, we design a cross-image self-attention injection for feature interaction and a step-wise distribution normalization to align the latent features. Incorporating these alignment processes into a diffusion model allows RIVAL to generate high-quality image variations without further parameter optimization. Our experimental results demonstrate that our proposed approach outperforms existing methods concerning semantic similarity and perceptual quality. This generalized inference pipeline can be easily applied to other diffusion-based generation tasks, such as image-conditioned text-to-image generation and stylization.

Paper

References (59)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 5k697/10 · confidence 3/52023-06-23

Summary

This paper proposes a framework RIVAL to generate variation of the real image without any tuning. It has two key components, the first one is a cross-image self-attention injection, where it first inverts the real image using DDIM, and then sample another random chain but uses a mix of real inverted chain’s value and denoised value in the denoising process. To mitigate the issue of out-of-distribution issue of the inverted latent of the real image, it further proposes a latent chain alignment process, where the random chain is initialised with the adaptively normalized Gaussian distribution of the inverted real latent. Experiments show improvement over baseline methods in terms of Text alignment and user preference. Overall the paper is well-written, the method is easy to follow.

Strengths

The proposed method does not require training or finetuning, although the inference time becomes twice as long as the normal inference, it is acceptable in practice. RIVAL is shown to work on other text-to-image applications such as test-driven image generation with real-image condition and example-based inpainting. Qualitative results show better identity preservation and style matching to the reference image compared to other baselines, plus, the editing results on text conditioning is impressive, considering no finetuning is needed. The authors also provide detailed ablation studies on the choice of alignment steps and framework components.

Weaknesses

The author does not provide analysis on the edibility of the proposed method, which I think is an important feature. How sensitive is the method to the CFG guidance weight? There is no ablation study on the choice of the CFG weight

Questions

In Figure 4, the person's identity seems to change over regeneration, any idea on how to better preserve the identity? Can the proposed method work with image conditional DM? Would be great to see some visual examples

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors provide limitation analysis, and see my questions above for further clarification

Reviewer 5k692023-08-21

Thank you for your response

The response resolved my concerns, I retain my original rating. I encourage the author to release code for the community.

Authorsrebuttal2023-08-21

Thank You for Recognizing RIVAL

Dear Reviewer 5k69, We are glad to see our rebuttal well addressed your issues. We greatly appreciate your consistently positive feedback on our work from beginning to end. Your comments are meaningful; for instance, considerations like identity preservation and the impact of CFG have led us to valuable reflections. As mentioned in our abstract, we will update the code after the final decisions are confirmed. We will release the code link and corresponding implementation in the final version (including new applications mentioned in the rebuttal, such as ControlNet). Thank you again for recognizing RIVAL!

Reviewer SLMY5/10 · confidence 4/52023-07-03

Summary

This paper works on the design of diffusion model to generate image variations given an image examplar as the source image. The basic idea is to align the image generation process to the image inversion process of source image. This is achieved by designing an cross-image self-attention injection for feature interaction between the source and generated images in the diffusion process. The results demonstrate better performance compared with the existing methods in image-conditioned text-to-image generation and examplar-based image inpainting.

Strengths

The idea to generate image variations given a source image is an interesting task. This paper proposed to align the source image inversion process and image generation process using the cross-image self-attention injection, which is a simple idea and shows good image generation results.

Weaknesses

1. The presentation and training of the parameters in "cross-image self-attention injection" are unclear to me. I understand that, in sect. 3.1, the features of examplar image and the generated image are interacted based on eqns. (2-4), in their inversion and generation chain respectively. However, how to train/determine the parameters in these equations? Are these parameters shared across different diffusion steps? And how do these parameters affect the final generation results? 2. Does this proposed generation method rely on the off-the-shelf diffusion model without any training? The chain alignment is conducted in the latent feature space or the image space? 3. An algorithmic description of this proposed method for image generation should be given for better understanding the training (possibly) and image generation process? 4. The shuffle operation in sect. 3.2 is described as "permutated sample from the inverted reference latent". Please clarify on this description, does this mean permutating the pixels of X_R^T? 5. In the Table I, please clarify on the metric of "Real-image Rank" and "Preference Rank". I know that lines 228-233 may describe them, however, it is still unclear to me on the difference between "Real-image Rank" and "Preference Rank".

Questions

I am interested in the problem to tackle and the proposed idea of feature interaction between source and generated image for inverse and generation chain alignment. However, I have many unclear points on the presentation of the proposed method (as mentioned in the weakness). I suggest authors to clarify the method in rebuttal and improve the writings and presentations of this manuscript.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The paper discussed on the limitation of this work.

Reviewer aMYr7/10 · confidence 2/52023-07-04

Summary

The authors propose a tunning-free pipeline called RIVAL(Real-world Image Variation by Alignment) for generating diverse and high-quality variations of real-world images. In previous works, some models also generate images with novel concepts and styles but require additional training stages and data, and others directly incorporated images as the input condition results in suboptimal visual quality and content diversity. To tackle this issue, the authors suggest a pipeline that maintains the style and semantic content of the reference without additional training stages and data. The proposed method comprises the inverted latent chain alignment(step-wise latent normalization) and cross-image self-attention injection. The authors determine whether aligning the inverted latent chain and injecting the modified self-attention key and value by adopting $t_{align}$ and $t_{early}$. Unlike the other papers that use the cross-image self-attention injection method with an additional forward to calculate the cross-attention mask, the suggested approach only needs a single forward. The authors show that RIVAL enhances performance by applying the alignment process to other text-to-image tasks through experiments and ablation studies.

Strengths

- The authors claimed a novel plug-and-play pipeline to generate high-quality, real-world image variations without additional optimization through an inverted latent chain alignment and cross-image self-attention injection. - The proposed pipeline enhances the performance of text-to-image models, such as text-driven image generation with real-image conditions and example-based inpainting. The pipeline generates images fast and flexibly with diverse text prompts adopting a new self-attention injection approach with a single forward pass. - Compared to CFG, the advantage of the proposed pipeline is its ability to avoid misalignment between latents in two chains and its direct utilization. The authors decouple the two inference chains and rescale noise prediction during denoising inference. Furthermore, they formulate noise prediction as an adaptive normalization. - The quantitative results outperform the other pipelines provided in the paper except for one metric, and the qualitative results show that RIVAL generates images with diverse text inputs while preserving the reference style.

Weaknesses

The experiments need to be more extensive. - The authors proposed a plug-and-play pipeline, but there are no quantitative results compared with the other plug-and-play pipelines mentioned in the paper. RIVAL used a similar scheme as MasaCtrl[2] and Plug-and-Play[3] except the latents alignment. The authors employed similar schemes to the pipeline, so comparing the quantitative results with the others would be beneficial. - In the appendix, the proposed method allows fast and flexible text-to-image generation, but there are no results about how fast it is. It would be reasonable to compare quantitatively with other plug-and-play pipelines. The authors need to validate the ability of the pipeline to apply to other diffusion-based generation tasks. - The authors show the qualitative results of self-example image inpainting in Fig. 6, but the results lack experiments. It would be hard to explain the ability to extend RIVAL’s framework to other image editing tasks. (Minor) The notation of $t_{align}$ and $t_{early}$ is confusing. [1] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In *CVPR*, pages 10684–10695, 2022.[2] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. *arXiv preprint arXiv:2304.08465*, 2023.[3] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. *arXiv preprint arXiv:2211.12572*, 2022.

Questions

- In Table 1, I wonder if using specific $t_{align}$ and $t_{early}$ values for each metric would lead to better quantitative results. - In Fig. 6, the inpainting results show a different masked area. Can you show me the same results by overlaying the masked area at the source image like SD[1] in the paper? - Can you provide examples of other tasks to which RIVAL can be applied? [1] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolutionimage synthesis with latent diffusion models. In *CVPR*, pages 10684–10695, 2022.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

The paper needs more extensive experiments and analysis of the limitations of the proposed method.

Reviewer mFBx7/10 · confidence 5/52023-07-23

Summary

Generating real-world image variations is an important research task with practical applications such as image editing, image synthesis, and data augmentation. Past approaches include texture synthesis, neural style transfer, and generative models, among others. This study proposes a method called "Real-world Image Variation by ALignment", which generates high-quality image variations from a single image sample through the inference process of an aligned diffusion model. The paper provides experiments on different images.

Strengths

The overall effect is very good. The proposed method takes full advantage of the properties of the diffusion model.

Weaknesses

There are a lot of recent papers based on a very well trained diffusion generative model. The results of these papers are generally very good (after all, it is based on a very mature large model). These papers generally have some micro-designs, which make the academic meaning of the papers more prominent. However, such papers have a common problem, that is, they do not discuss the research issues in a very specific way. Often only stay in the effect. But this hardly accounts for the actual scholarly progress of these papers. Chances are it's just a manifestation of the capabilities of the base model itself. Taking this paper as an example, the paper proposes the method of Aligning Diffusion Inversion Chain. But the necessity and interpretability of the method is not fully discussed. In fact, there are many "tricky" ways to achieve the effect of this paper, and it is not necessary to go through the method described in Section 3 (By the way, the writing of Section 3 is very difficult to understand. In fact, a lot of useless mathematics is unnecessary. It is recommended that authors use more carefully set diagrams to explain their methods). I doubt the necessity of the proposed method. I hope the author can have a full ablation experiment, and compare and discuss it with a wide range of existing tricks. Of course, I also realize that such a requirement is a bit too harsh. Because many methods and tricks have not been written into papers or formally disseminated. The authors may not be able to adequately discuss every possibility. Interpretability is also another important topic. I want to know the specific physical meaning of each latent and representation in the inversion transfer or the point that it can be exploited in this method. This paper doesn't really explain it (or even try to). ===================== Post rebuttal: Thanks for the rebuttal. I am willing to rase my score to accept (7). But I recommend authors to add these new discussions in to the final version.

Questions

Two general questions: 1. Ablation study. 2. Interpretability. Please see Weaknesses for full details.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes

Reviewer 4HxC6/10 · confidence 4/52023-07-26

Summary

The paper investigates the domain gap between generated images and real-world images, as it is challenging in generating high-quality variations of real-world images. It mainly contains two contributions, the first one is the cross-image self-attention injection for generating images that correspond to the given exemplar, and the second one is the latent alignment, which can help to eliminate the biased tone and semantics. The comparison with stable diffusion, SD variation, and ELITE validates the effectiveness of each component. Lastly, the ablation studies analyze in depth the contribution of each submodule, showing the positive impact of each part.

Strengths

+ The paper is well-written and easy to follow for most parts, except when introducing the details of the shuffle operation in equation 8. It is not easy to directly understand the operation. + The paper has a nice flow of descriptions and figures to illustrate the qualitative comparison. + The diagrams are easy to understand and are well-made. + The main contribution of the proposed cross-image self-attention injection and latent alignment are intuitive and well-executed in this paper. Each design is reasonable to handle the corresponding challenges, like object custom and style consistency. + The ablation study is sufficient to validate the effectiveness of each component in designing the whole network, the concatenation of cross-image features in different steps, and the alignment of two features. + Experiments show superior performance in comparison with recent works.

Weaknesses

- Generally speaking, this work is more like the recent customized text-to-image generation, except that this work considered style consistency, like the analyses in Figure 4 (2nd-row). Since these competing methods did not consider this style factor, so they easily have the tones inconsistency problem. Directly comparing with them seems not very fair, as they can generate great results, except for the inconsistent tone distribution. In my opinion, it seems not a crucial issue. - The concatenation of v_G and v_R in equation 3 is also like some training-free video generation, which adopts the cross-frame interaction. The second proposed latent alignment is widely used in style transfer tasks.

Questions

Some deep analyses of the failed cases seem not enough. The provided examples cover very wide ranges, like semantic bias, complex scenes, and hard concepts. What is the difference between custom image generation and variation generation? It seems that it only has the style difference, as this work needs to keep the style consistent with the given exemplar. In this case, I think the authors should discuss these related works, and compare with them by considering style transfer factors.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer mFBx2023-08-11

Raising My Score

Thanks for the rebuttal. I am willing to rase my score to accept (7). But I recommend authors to add these new discussions in to the final version.

Authorsrebuttal2023-08-11

Thank You for Recognizing RIVAL

Dear Reviewer mFBx, Thank you for recognizing our work and the content of our rebuttal. The questions you raised in the review are essential for explaining the effectiveness of RIVAL and have helped us to delve deeper into the underlying mechanisms of this work. We will add all relevant experiments and discussions in the revised version. Besides, we will organize the content in the method section to make the article more understandable and academically meaningful. Thank you again for acknowledging our work!

Reviewer 4HxC2023-08-16

I appreciate the authors' detailed response and my concerns are addressed well. Considering other reviews, I would like to give weak accept.

Authorsrebuttal2023-08-16

Thank You for Recognizing RIVAL

Dear Reviewer 4HxC, Thank you very much for recognizing our work and the effort we put into our rebuttal. Especially in explaining this paper's novelty, your questions have helped us engage in a more comprehensive discussion and comparison of RIVAL's mechanism. We will integrate the pertinent discussions, especially concerning the comparison's fairness and the method's novelty, into the subsequent revision.

Reviewer SLMY2023-08-17

Thanks for the responses

I appreciate the responses for clarifying my questions and they are mostly clear, and the paper should consider these clarifications in the revised version, especially on the Q1 and Q2. I retain my rating and it is ok to me if accepting the paper.

Authorsrebuttal2023-08-17

Thank You for Your Comments and Feedback

Dear Reviewer SLMY, Thank you for acknowledging our clarification of issues in the rebuttal. As mentioned in Q6, based on your valuable feedback, we will revise the presentation in the paper's method section of the version to clarify our method better. We also appreciate your positive evaluation of this paper and your endorsement of our approach. As mentioned in the **global response Q1&Q2**, we will continue exploring broader applications of our method. We expect it can make contributions to the community. We also hope that the explorations we've made in our response can better demonstrate the effectiveness of RIVAL to you. Thank you again for your review and feedback on our response!

Authorsrebuttal2023-08-18

Thank You for Recognizing RIVAL

Dear Reviewer aMYr, Thank you very much for recognizing our work and rebuttal. Especially in exhibiting the applicability to other tasks and comprehensive quantitative comparisons, your questions have helped us engage in a more comprehensive discussion and comparison of RIVAL's mechanism. We will integrate the pertinent discussions into the revision. Thank you again for your review and feedback on our response!

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC