CycleNet: Rethinking Cycle Consistency in Text-Guided Diffusion for Image Manipulation

Diffusion models (DMs) have enabled breakthroughs in image synthesis tasks but lack an intuitive interface for consistent image-to-image (I2I) translation. Various methods have been explored to address this issue, including mask-based methods, attention-based methods, and image-conditioning. However, it remains a critical challenge to enable unpaired I2I translation with pre-trained DMs while maintaining satisfying consistency. This paper introduces Cyclenet, a novel but simple method that incorporates cycle consistency into DMs to regularize image manipulation. We validate Cyclenet on unpaired I2I tasks of different granularities. Besides the scene and object level translation, we additionally contribute a multi-domain I2I translation dataset to study the physical state changes of objects. Our empirical studies show that Cyclenet is superior in translation consistency and quality, and can generate high-quality images for out-of-domain distributions with a simple change of the textual prompt. Cyclenet is a practical framework, which is robust even with very limited training data (around 2k) and requires minimal computational resources (1 GPU) to train. Project homepage: https://cyclenetweb.github.io/

Paper

Similar papers

Peer review

Reviewer sYAZ6/10 · confidence 4/52023-07-03

Summary

The paper proposes CycleNet, a method for unpaired image-to-image translation using pretrained diffusion models. The main idea is to reconstruct the conditional images through a reverse process. The idea seems reasonable, and the paper is well-organized, with sufficient experimental results.

Strengths

The idea seems reasonable, and the paper is well-organized, with sufficient experimental results.

Weaknesses

1. The concept of cycle construction has been extensively studied in the field, and this work follows a similar idea, which weakens its contribution. 2. In CycleGAN, the reconstruction constraint $L_{x \rightarrow y \rightarrow x}$ plays a vital role. However, in FastCycleNet, this constraint is omitted, yet the performance does not show significant drops. It would be helpful if the authors could provide an explanation for this observation. 3. For some variable notation (such as ($\bar{x}_o$), authors had better provide their definitions, It would be beneficial for readers to quickly understand these notations.

Questions

Please see in weakness

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

4 excellent

Presentation

4 excellent

Contribution

2 fair

Limitations

Yes

Reviewer yCsT2/10 · confidence 5/52023-07-06

Summary

The paper introduces CycleNet, a new method that enhances image manipulation by incorporating cycle consistency into diffusion models. The paper addresses the challenge of unpaired image-to-image translation and aims to provide a consistent and intuitive interface for this task.

Strengths

1.Originality: The paper introduces CycleNet, a new method that incorporates cycle consistency into diffusion models for image manipulation. This approach combines the concept of cycle consistency with text-guided diffusion models. 2. Quality: The paper provides a comprehensive evaluation of CycleNet's performance using various quantitative metrics such as FID, CLIP Score, LPIPS, PSNR, SSIM, and L2. The results also support the claims of the paper. 3. Clarity: The paper is well-written and presents the concepts, methodology, and findings.

Weaknesses

1.Novelty is limited. The concept of combing cycle-consistency constraint and diffusion model for image-to-image translation has been widely explored by other works, such as UNIT-DDPM, Dual Diffusion Implicit Bridges (DDIBs), cyclediffusion and so on. 2.Lack of Comparative Analysis between the proposed method with other existing cycle-consistency based diffusion models. The part of 6.3 is not sufficient to illustrate the key difference and novelty of the proposed method. 3.The results are low-quality and far from satisfaction. For the classic image-to-image translation tasks, i.e., winter-to-summer and horse-to-zebra, I do not find any improvement over existing ui2i methods. In addition, the results of CUT aand CycleGAN are questionable, especially for winter-to-summer task, they are too bad compared with the results from their original papers. 4.The results of coffee-to-empty also do not convince me as it seems to only transfer the rgb image to gray image. Similar results also happen in the horse-to-zebra.

Questions

1. Why there lack the results of strong baseline CycleGAN in Table 1? 2. Could the authors consider conducting a user study to evaluate the subjective quality and usability of CycleNet? This would provide insights into how users perceive the generated images and the overall user experience, further validating the practicality and effectiveness of CycleNet. 3. Could the authors provide more information about the computational efficiency of CycleNet? Specifically, details about the training and inference times, as well as the hardware specifications used, would be helpful in understanding the practicality of CycleNet for real-world applications.

Rating

2: Strong Reject: For instance, a paper with major technical flaws, and/or poor evaluation, limited impact, poor reproducibility and mostly 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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

1. Novelty is limited. 2. Poor results are far from satisfaction as the strong generation ability of the diffusion model.

Reviewer yCsT2023-08-19

Final comments after reading authors' response

My final rating will still be Strong Reject. The reasons mainly include three aspects: 1. limited novelty as the cycle-consistency has been widely explored in existing literatures; 2. poor image results as image editing quality does not improve with cycle-consistency and numerous edits appear unreasonable.; 3. the comparison results of existing methods are questionable.

Authorsrebuttal2023-08-19

Response to Reviewer yCsT's Comments

We thank the reviewer for their prompt response to us! We would like to confirm a few factual understandings so there is no miscommunication over the review process. ### Questionable Results **Could the reviewer confirm that you have checked (1) our updated results in the General response and (2) the original CUT paper** (https://arxiv.org/abs/2007.15651)? Our presented images for CUT and CycleGAN in Figure 2 are **exactly the original copy** of the original paper (Horse->Zebra: Fig 3 & Summer->Winter: Fig 11). While the reviewer commented that “they are too bad compared with the results from their original papers”, we did use the same copy. ### Novelty **Could the reviewer confirm that they have read our discussion in General Response 2 (expanded below)?** If so, we would love to receive clear pointers to additional research that we missed in the discussion. We hope to know if the reviewer agrees that CycleNet presents the first effort that explicitly enforces cycle consistency constraints in image editing using pre-trained diffusion models. **Q: Has cycle-consistency constraint been explicitly integrated into diffusion models?** Not yet, to the best of our knowledge. Reviewer yCsT mentioned that: > “The concept of **combing cycle-consistency constraint** and **diffusion model** for image-to-image translation has been widely explored by other works” We believe there is a factual misunderstanding. To the best of our knowledge, this is the first effort that **explicitly integrates cycle consistency constraint** in unpaired image-to-image translation using pre-trained diffusion models. We expand on what we have discussed in Section 6.3 regarding the several existing efforts to study cycle consistency in diffusion-based I2I translation. - UNIT-DDPM ([39], Apr 2021) made an initial attempt in the unpaired I2I setting, training two DPMs and two translation functions from scratch. Cycle consistency losses are introduced in the translation functions during training to regularize the reverse processes. At inference time, the image generation does not depend on the translation functions, but only on the two DPMs in an iterative manner, leading to sub-optimal performance. This preprint does not have any public implementation, and can not be reproduced given the limited information in the paper. - DDIB (Mar 2022): The key proposition states that exact cycle consistency is possible **assuming zero discretization error**. DDIB itself does not enforce any cycle consistency constraint: > “DDIBs first obtain latent encodings for source images with the source diffusion model, and then decode such encodings using the target model to construct target images.” DDIB proposed one of the first theoretic analyses on the possibility to perform unpaired I2I translation with diffusion models and is the theoretic foundation for various diffusion-based image translation models, including prompt2prompt, NullText, and CycleDiffusion. - CycleDiffusion ([46], Nov 2022) proposes a zero-shot approach for image translation based on DDIB’s observation that a certain level of consistency could emerge from DMs. There is again no explicit treatment to encourage cycle consistency. > “CycleDiffusion can be seen as an extension of DDIB approach to stochastic DPM” [46] > “The method is based on the observation that cycle consistency naturally emerges from diffusion models, i.e., the same random noise leads to similar images. Therefore, there is no training objective that explicitly encourages cycle consistency.” (https://github.com/ChenWu98/cycle-diffusion/issues/16) If there is any similar prior work the reviewers are aware of, please let us know. **Q: Since a certain level of cycle consistency could emerge in diffusion models, do we benefit from enforcing it as a constraint?** Yes. We echo the comments of reviewer rhuF: > “When we apply pre-trained DM models on I2I tasks, it is difficult or unpredictable to control the mask and attention maps to achieve the desired results.” While exact cycle consistency is possible assuming zero discretization error, the assumption generally doesn’t hold true. Pixel-level consistency remains to be challenging when there exists image conditioning, as can be shown in the failure cases of baselines presented in our paper, and the constantly changing backgrounds in image and video synthesis tasks. **Q: Why is it non-trivial to enforce cycle-consistency constraints in diffusion models?** Unlike GAN-based models, in which the translation between images is end-to-end, diffusion-based methods are iterative and lack such a simple and intuitive translation cycle. Also, for text-guided diffusion models, it is difficult to design a discriminator, which is a key component to ensure cycle consistency in GAN-based models. To address this limitation, we carefully formulated the problem in Section 2 and theoretically derived the training objectives in Section 3.

Reviewer aEiE6/10 · confidence 4/52023-07-06

Summary

This work aims at addressing the task of unpaired I2I translation with pre-trained diffusion models. Inspired by CycleGAN, authors incorporate cycle consistency into diffusion model to regularize process of image translation and proposed CycleNet. In addition it also contributes a multi-domain I2I translation dataset with object state change. Extensive experiments demonstrate that the proposed method can produce high-quality translation results with good consistency and is able to generalize to out-of-domain generation tasks by changing text prompts.

Strengths

1. It is interesting and reasonable to introduce cycle consistency into diffusion models for training unpaired I2I translation models. 2. A new multi-domain I2I translation dataset for object state change with text prompt is collected. 3. Extensive experiments demonstrate the effectiveness of the proposed method. 4. Capbility to generalize to out-of-domain data by simply changing text prompt.

Weaknesses

1. I wonder how CycleNet compares with CycleGAN. More comparison between CylceNet and CycleGAN should be provided. If their performance is similar, then why use a diffusion-based method instead of GAN-based for unpaired I2I translation task? 2. Authors are suggested to experiment with more diffusion models such as T2I Adapter to show its generalization on models. 3. Any explanation on why FastCycleNet perform better in several metrics? The coffee-to-empty translation result does not seem good, not only changing the color but also failing to become empty.

Questions

See weakness.

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

Authors have clearly showed failure cases of the proposed method in the main text. But they do not address potential negative socital impact of this work.

Reviewer Tt5a5/10 · confidence 4/52023-07-07

Summary

This paper presents a method that tackles the challenge of consistent image-to-image (I2I) translation using diffusion models (DMs) and unpaired image conditioning and text prompts. The proposed approach incorporates Cycle Consistency Regularization to ensure cycle consistency and Self-Regularization to generate images that align with the target domain. Furthermore, the paper introduces a multi-domain I2I translation dataset that encompasses object state changes.

Strengths

a) From the visual results, the proposed CycleNet can well preserving the structures of the input image. b) The author introduces a multi-domain I2I translation dataset for object state changes.

Weaknesses

a) The experimental results are quite strange. i. In Figure 2, Figure 3, Figure 5, and many images in the supplementary materials, The results of CycleNet show severe hue shifting. This is a serious issue, and I suggest the author analyze and address this problem. b) The author's comparison methods are not comprehensive. Many editing methods based on diffusion have been proposed, such as pix2pix-zero, masactrl, etc. I suggest the author compare these methods. c) Although CycleNet aims to handle unpaired I2I translation tasks, it still requires paired datasets for training. So, how well does CycleNet generalize to unseen domains (such as variations of other unseen species, etc.)?

Questions

Please refer to the weaknesses part.

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

2 fair

Limitations

Please refer to the weaknesses part.

Reviewer rhuF5/10 · confidence 4/52023-07-10

Summary

This paper introduces cycle consistency in diffusion models to achieve regularization in image manipulation. It allows out of domain image generation with text prompt modification, and can be trained with very little data (~2K) with minimal computational requirements (1GPU). When we apply pre-trained DM models on I2I tasks, it is difficult or unpredictable to control the mask and attention maps to achieve the desired results. Cyclic consistency regularization allows us to address this. Experiments are conducted on many datasets such as Yosemiti summer-winter, horse-zebra, etc. On standard metrics such as CLIP, FID, and LPIPS, the proposed method outperforms other baselines such as CycleGAN, TEX2LIVE, etc.

Strengths

1) There has not been much regularization in diffusion models except for guidance driven by the conditional text. It is nice to see the celebrated Cycle-Consistency loss developed for diffusion models. 2) The paper clearly motivates and states the problem. It is nice to see that the proposed method nicely inherits the controlnet architecture to incorporate the cycle-consistency loss function during training.

Weaknesses

1) While enforcing cyclic consistency we are feeding the actual input image as the conditional input. This seems a bit strong and it appears that the diffusion model may be overfitting to the conditional input. 2) It is not clear how one obtains the negative prompt c^{-1}. The method takes a prompt c_{text}=(c^{+}, c^{-}), where c^{+} is the prompt that drives the diffusion process towards the images that are associated with it, and the negative prompt c^{-} drives the process away from these images. The paper only mentions this once without sufficient details on how this is done. 3) The proposed method heavily relies on controlNet and uses the original image to impose the Cyclic-Consistency loss. It may be worth just testing with standard ControlNet with canny edge-map as conditional input and appropriate text prompt.

Questions

Overall, the paper looks at an important problem and brings cyclic consistency to diffusion models. Please see my concerns above.

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

3 good

Contribution

3 good

Limitations

No concerns.

Area Chair FHyR2023-08-18

Diverging scores

Dear reviewers, The paper got diverging scores. The authors provided their response. Please engage into the discussion with authors, fellow reviewers and provide your final recommendation? Reviewers Tt5a, yCsT are on the negative side. Do the other reviews and the authors' response change your mind? Reviewers rhuF, aEiE, and sYAZ are on the positive side. Do the comments from Tt5a, yCsT influence your recommendation? Please discuss. Your AC

Reviewer Tt5a2023-08-20

Official Comment by Reviewer Tt5a

I would like to convey my appreciation for the authors' dedicated efforts in responding to my inquiries. In response to the Hue-Shifting issue, the authors proposed a solution to mitigate this problem. From the attached PDF provided by the authors, it can be seen that the hue shifting has indeed been alleviated to some extent. In addition, the authors have also included experimental results for pix2pix-zero and masactrl. Therefore, I will raise my score to 5.

Authorsrebuttal2023-08-20

Response to Reviewer Tt5a's Comments

We thank the reviewer for your positive feedback and prompt response! We are pleased that we have addressed your concerns regarding the hue-shifting issues and missing baselines you mentioned. These updates will be certainly reflected in the next version. Your suggestions are invaluable in enhancing our work.

Reviewer rhuF2023-08-21

Acknowledging the rebuttal

I thank the authors for the rebuttal. It addresses most of my concerns including the ControlNet experiments. I would like to keep my positive rating.

Authorsrebuttal2023-08-21

Response to Reviewer rhuF's Comments

We thank the reviewer for your response! We are pleased that we have addressed your concerns and your evaluation remains positive. The updates will be reflected in the next version and your suggestions are invaluable for our work!

Reviewer aEiE2023-08-22

Response to rebuttal

I appreciate the efforts made by authors in addressing most of my concerns with more experimental results and explanation. So I would like to keep my original rating.

Authorsrebuttal2023-08-22

Response to Reviewer aEiE's Comments

We thank the reviewer for the positive feedback! We are pleased that we have addressed your concerns regarding the potential experiment settings and baselines, and that the reviewer found our explanations convincing. The updates will be reflected in the next iterations. Your suggestions are invaluable in enhancing our work.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC