RefDrop: Controllable Consistency in Image or Video Generation via Reference Feature Guidance

There is a rapidly growing interest in controlling consistency across multiple generated images using diffusion models. Among various methods, recent works have found that simply manipulating attention modules by concatenating features from multiple reference images provides an efficient approach to enhancing consistency without fine-tuning. Despite its popularity and success, few studies have elucidated the underlying mechanisms that contribute to its effectiveness. In this work, we reveal that the popular approach is a linear interpolation of image self-attention and cross-attention between synthesized content and reference features, with a constant rank-1 coefficient. Motivated by this observation, we find that a rank-1 coefficient is not necessary and simplifies the controllable generation mechanism. The resulting algorithm, which we coin as RefDrop, allows users to control the influence of reference context in a direct and precise manner. Besides further enhancing consistency in single-subject image generation, our method also enables more interesting applications, such as the consistent generation of multiple subjects, suppressing specific features to encourage more diverse content, and high-quality personalized video generation by boosting temporal consistency. Even compared with state-of-the-art image-prompt-based generators, such as IP-Adapter, RefDrop is competitive in terms of controllability and quality while avoiding the need to train a separate image encoder for feature injection from reference images, making it a versatile plug-and-play solution for any image or video diffusion model.

Paper

Similar papers

Peer review

Reviewer Z6755/10 · confidence 4/52024-07-05

Summary

This paper targets the task of subject consistency in image and video generation. Previous consistency generation methods [53,68] are based on concatenated attention. The authors reformulate concatenated attention in a manner similar to classifier-free guidance, simplifying the constant C matrix to a constant c, leading to the proposed reference feature guidance (RFG). The authors applied RFG to multiple applications, including subject consistency in image and video generation. Qualitative and quantitative results show that RFG outperforms prior works such as IP-Adapter, BLIPD, and SDXL.

Strengths

1. The authors propose a simple yet effective method for consistent text-to-image generation. The relationship among the proposed method, cross-frame attention, and concatenated attention is an interesting observation. 2. The proposed method is shown to be useful in multiple image and video generation applications, including blending features of multiple images, using negative examples to increase diversity, improving temporal consistency in video generation, and preserving identity in video generation. This method is shown to be generalizable to different techniques. 3. The paper includes comprehensive visual examples demonstrating good visual quality.

Weaknesses

1. The generated images appear to have limited diversity. In Figure 13, many examples of humans have very similar poses and styles. In contrast, IP-Adapter generates more diverse images. 2. The paper discusses the relationship between concatenated attention and cross-frame attention, mentioning that the results have similar effects to recent works [53] and [68]. However, visual comparisons are not included. 3. Most results focus on humans. It would be beneficial to see different types of subjects/objects. 4. Although the proposed method successfully generates videos with temporal consistency, the generated videos contain less motion. Could the authors explain the possible reason for this and suggest ways to improve it?

Questions

The text alignment decreases in Figure 14. Could the authors provide comments on why this happens and suggest possible solutions to improve text alignment?

Rating

5

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

Yes, the authors have addressed the limitations and potential negative impact.

Reviewer c2sE6/10 · confidence 4/52024-07-12

Summary

The authors analyzed the current methods of consistent content generation and revealed that the concatenation of reference features in the self-attention block can be reformulated as a linear interpolation of image self-attention and cross-attention between synthesized content and reference features with a constant rank-1 coefficient. Motivated by the analysis, the paper identified the rank-1 coefficient is not necessary and proposes a simplified method by replacing the rank-1 coefficient with a scalar coefficient. Compared to the previous method, this method doesn’t require additional training, while achieving considerable performance, and allowing multiple applications, such as multi-references, concept suppression, and video generation.

Strengths

1. The method and related work are clear and well-written, easy to follow. 2. The proposed method is simplified based on previous methods and achieves effective results. 3. The method is training-free and the authors claim that it’s flexible and plug-and-play for diverse text-to-image models. 4. The paper exhibited many qualitative results and highlighted its effectiveness with comprehensive experiments for different applications.

Weaknesses

1. As the paper mentions the proposed method is plug-and-play, if a subsection is added to show the flexibility of the method across different models and potential limitations/considerations when applying it, that would be great. 2. The quantitative comparisons can be difficult for this task but is it possible to use CLIP or other models to evaluate the consistency of the reference and generated images quantitatively? 3. A more comprehensive discussion on its limitations, failure cases, or social impact would be great.

Questions

1. How many generated images are used for human evaluation, and what do the values for the vertical axis (such as over 300 for consistency score) mean? 2. It seems that the major difference between the proposed method and the previous method is whether to use a rank-1 matrix or a scalar for the interpolation of attention. Could the authors elaborate more on why the previous method requires training while the proposed method does not? How is the rank-1 matrix obtained in the previous method? 3. How to align the parameters for the methods using the rank-1 matrix and the scalar so that the control of the reference is comparable across methods?

Rating

6

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors provide a concise discussion of limitations and it would be better to include a more comprehensive one and provide discussion of potential societal impact.

Authorsrebuttal2024-08-07

Continue the rebuttal

> A more comprehensive discussion on its limitations, failure cases, or social impact would be great. > Limitations: - For consistent image generation, our method sometimes fails to replicate the exact hairstyle and outfit from the reference image. This could be a critical requirement for tasks like advertisement video production. - For diverse image generation, our method currently cannot precisely control aspects of diversity, such as generating diverse subjects but not styles, or vice versa. Our method does not disentangle these two aspects. - For blending multiple images, the coefficient needs careful tuning to achieve a good balance among multiple reference images, and it heavily depends on the specific case. - For improving video temporal consistency, as noted by reviewer **Z675**, although our method successfully generates videos with temporal consistency, the generated videos contain less motion. Potential societal impact: The proposed method has several potential societal impacts across various domains: - For media and advertising, it can enhance the design of key frames in advertisements or movie videos, allowing companies to create more visually compelling content. - Individual artists may leverage this technique to blend multiple images, potentially sparking new creative inspirations. - RefDrop could be used to improve temporal consistency in short videos, which may be particularly valuable for social media platforms seeking to enhance AI-generated video content.

Reviewer c2sE2024-08-11

Thanks for the detailed information from the authors. My concerns are addressed and I will increase my rating from 5 to 6. Please make sure to include the clarified information in the revised paper.

Authorsrebuttal2024-08-12

Thank you very much for taking the time to read our rebuttal and raising the score! We will include the clarified information in the revised paper.

Reviewer NJob5/10 · confidence 4/52024-07-12

Summary

This paper focuses on the challenge of ensuring consistency in the generation of images and videos. Deep learning and artificial intelligence techniques are utilized in image and video generation for generating new images or generate video frames based on given inputs, such as text prompts or reference images. Ensuring consistency is necessary for maintaining coherence between the subject or object in the generated images or video frames. The previous methods mainly consisted of the IP (image prompt)-Adapter and concatenated attention approaches. The IP-Adapter technique retrieves characteristics from reference images and incorporates them into the generation process to improve coherence. More precisely, the method employs an image encoder that has been trained independently to extract features from the reference image. These features are then incorporated into the generation process using attention mechanisms, which involve both self-attention and cross-attention. Nevertheless, this method necessitated distinct training of the encoder and utilized unchanging weights, leading to restricted adaptability. Furthermore, it is difficult to maintain the coherence between the textual content and the accompanying visual references. This paper introduces a novel approach called RefDrop to address these challenges. The major contributions are as follows: 1. A review of methods for generating consistency. The paper demonstrates that current concatenated attention methods can be understood as linear combinations of self-attention and cross-attention. This discovery indicates these techniques merge features from reference images using linear interpolation. 2. Reference Feature Guidance (RFG): The RFG method is presented as an approach for directly managing the impact of reference images by linearly combining self-attention and cross-attention. Users can enhance adaptability by adjusting the influence of reference images using a scalar coefficient 𝑐. 3. No prior training necessary: RefDrop provides an adaptable method to efficiently utilize features from reference images without requiring separate training of the encoder. RefDrop is a novel method that utilizes features from reference images to effectively control consistency in the generation of images and videos. This method controls the impact of reference images by using linear interpolation of self-attention and cross-attention. More precisely, features are derived from both the input and reference images, and subsequently, self-attention and cross-attention are executed independently. The outcomes of these considerations are subsequently combined through linear interpolation to generate the ultimate features, employing a scalar coefficient c to equitably distribute the influences from each attention mechanism. This enables adaptable manipulation of the impact of reference images. RefDrop efficiently employs reference features without requiring additional training, making it suitable for a wide range of generation tasks. Furthermore, it guarantees temporal consistency in video generation by utilizing the initial frame as a reference image, thereby preserving coherence throughout the following frames.

Strengths

- Integration of RefDrop into Diffusion Models Without Additional Training - Controlling Reference Image Impact in RefDrop for Efficient Image Generation - High-Quality Image Generation with RefDrop Using Single and Multiple References - Temporal Consistency in Video Generation with RefDrop

Weaknesses

- The lack of feature-based comparisons using metrics like Kullback-Leibler (KL) divergence in the experiments

Questions

This approach may introduce undesired background elements, complicating the management of specific image components, isn't it?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

- The lack of attention between the reference images and text, which can result in feature entanglement

Reviewer 5p1r6/10 · confidence 4/52024-07-12

Summary

This paper presents RefDrop, a method that allows users to control the influence of reference context in a direct and precise manner. More specifically, the proposed method is training-free, which means it can be used plug-and-play without the need to train a separate image encoder for feature injection from reference images.

Strengths

- The paper is well written and easy to follow. - This paper presents an intuitive way for consistent image/video generation by fusing the features of objects into the diffusion process. - Experiments show the effectiveness of proposed method on preserving object appearance.

Weaknesses

- One problem of the proposed approach is that it does not dis-entangle spatial control with appearance control. A stronger guidance scale not only means higher appearance similarity, but also higher spatial similarity between the generated image/video and the reference image. This could be an intrinsic drawback of the proposed method.

Questions

- Can the authors provide some gifs for generated videos? I'm a little bit concerned that if we enforce high guidance scale, the generated video will probably be quite static without large movements. - I guess reference-ControlNet is also relevant? It would be nice to include it as a baseline.

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

See weakness.

Authorsrebuttal2024-08-12

Thank you for reading our rebuttal! We are more than happy to share our thoughts on your questions. > Other than Figure 3, do you have some intuitive explanations for this? > Our approach is largely inspired by recent advancements in fine-grained control for style transfer using SDXL UNet [1,2,3,4]. We hypothesize that the blocks processing low-resolution image feature maps—particularly those near the bottleneck in the UNet—exert more control over the overall coarse structure or core formulation of the image. In contrast, the blocks processing high-resolution image feature maps play a crucial role in controlling high-frequency details, a task that is difficult for low-resolution blocks to manage. Regarding why the first upsampling block influences the spatial layout the most, rather than a middle or final downsampling block, we conjecture it is because the other blocks are not yet sufficiently close to the final output image to have a significant impact. We conducted several ablation studies and found empirically that excluding the first upsampling block is the most effective in generating more diverse poses. > And has such idea of excluding the first upsampling block explored in previous works? > We are indeed inspired by a series of works that have demonstrated how different blocks of the SDXL UNet affect the final image generation. These works include, but are not limited to, InstantStyle [1], B-Lora [2], and discussions within the open-source community [4]. Most of these studies [1, 3, 4] focus on injecting style from a reference image and investigating which blocks most influence style generation. Meanwhile, B-Lora also analyzes which blocks have the greatest impact on content generation. In particular, B-Lora identifies the first upsampling block as the most influential for content generation, significantly controlling the spatial layout—a finding that aligns with our observation. However, these studies are primarily concerned with style transfer for text-to-image models. The key difference is that they aim to preserve the spatial layout or style from the reference image as much as possible, while we seek to exclude the spatial layout information from the reference image as much as possible. **Regarding consistent generation, we have not yet observed any work that uses this method to diversify object poses.** Previous consistent generation method Consistory [5] introduced several techniques to avoid the issue of similar poses, such as self-attention dropout and using vanilla query features. We believe that our approach of excluding the first upsampling block offers a simpler and more principled solution in practice. > Do you have some quantitative analysis to prove that removing the first upsampling block is more effective than removing other blocks? > We also include the quantitative results here. Specifically, we conducted 11 groups of experiments, where in each group, we excluded one of the 11 blocks of the UNet from applying Reference Feature Guidance. And in each group, we generated 20 sets of consistent objects, with each set containing 5 images. This resulted in a total of $11\times 20 \times 5 = 1100$ images for metric calculation. We then used `DreamSim` and `LPIPS` to measure the distance between the generated images and the reference image, and report the mean and standard deviation. Higher values of these metrics should indicate more diverse poses, as these scores tend to favor similar layouts [5]. From the table, we can see that excluding the first upsampling block greatly boosts the spatial layout diversity. | Excluded block | DreamSim distance | LPIPS distance | | --- | --- | --- | | Down1 | 0.2283 ± 0.0102 | 0.5484 ± 0.0052 | | Down2 | 0.2261 ± 0.0118 | 0.5484 ± 0.0043 | | Down3 | 0.2298 ± 0.0082 | 0.5475 ± 0.0058 | | Down4 | 0.2332 ± 0.0095 | 0.5551 ± 0.0072 | | Mid | 0.2484 ± 0.0048 | 0.5673 ± 0.0070 | | Up1 | **0.3077 ± 0.0092** | **0.5880 ± 0.0047** | | Up2 | 0.2567 ± 0.0087 | 0.5603 ± 0.0077 | | Up3 | 0.2440 ± 0.0088 | 0.5580 ± 0.0032 | | Up4 | 0.2441 ± 0.0079 | 0.5611 ± 0.0074 | | Up5 | 0.2368 ± 0.0106 | 0.5540 ± 0.0055 | | Up6 | 0.2455 ± 0.0103 | 0.5639 ± 0.0075 | **Reference** [1] Wang, Haofan, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. "Instantstyle: Free lunch towards style-preserving in text-to-image generation." *arXiv preprint arXiv:2404.02733* (2024). [2] Frenkel, Yarden, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. "Implicit Style-Content Separation using B-LoRA." *arXiv preprint arXiv:2403.14572* (2024). [3] Jeong, Jaeseok, Junho Kim, Yunjey Choi, Gayoung Lee, and Youngjung Uh. "Visual Style Prompting with Swapping Self-Attention." *arXiv preprint arXiv:2402.12974* (2024). [4] https://github.com/huggingface/diffusers/discussions/7534 [5] Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consistent text-to-image generation. arXiv preprint arXiv:2402.03286, 2024.

Reviewer 5p1r2024-08-12

Thanks for this detailed explanation from the authors, my question has been well-addressed, really appreciate it. I will increase my rating from 4 to 6. Please make sure to include these discussions in the revised version, since intuitions and observations sometimes is more useful for other researchers.

Authorsrebuttal2024-08-13

Thank you very much for your response! We quite enjoy this rebuttal process because of your insightful feedbacks, and we are already working on including these discussions in the revised version.

Reviewer 5p1r2024-08-12

Thanks for the authors' response! Here are one remaining question after reading your rebuttal: - I personally don't find it very intuitive to exclude the first upsampling block. Other than Figure 3, do you have some intuitive explanations for this? And has such idea of excluding the first upsampling block explored in previous works? You mentioned that "the first upsampling block predominantly influences the spatial layout.", do you have some quantiative analysis to prove that removing the first upsampling block is more effective than removing other blocks? Thanks!

Reviewer Z6752024-08-14

Many thanks to the authors for their detailed reply and efforts. My concerns have been well addressed.

Reviewer NJob2024-08-14

Many thanks for your feedback. 1) I understand what the authors did for quantitative evaluation. 2) To demonstrate the reason of less background leakage, I think that visualizing the attention output can be one of solutions.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC