Inserting Anybody in Diffusion Models via Celeb Basis

Exquisite demand exists for customizing the pretrained large text-to-image model, $\textit{e.g.}$, Stable Diffusion, to generate innovative concepts, such as the users themselves. However, the newly-added concept from previous customization methods often shows weaker combination abilities than the original ones even given several images during training. We thus propose a new personalization method that allows for the seamless integration of a unique individual into the pre-trained diffusion model using just $\textbf{one facial photograph}$ and only $\textbf{1024 learnable parameters}$ under $\textbf{3 minutes}$. So as we can effortlessly generate stunning images of this person in any pose or position, interacting with anyone and doing anything imaginable from text prompts. To achieve this, we first analyze and build a well-defined celeb basis from the embedding space of the pre-trained large text encoder. Then, given one facial photo as the target identity, we generate its own embedding by optimizing the weight of this basis and locking all other parameters. Empowered by the proposed celeb basis, the new identity in our customized model showcases a better concept combination ability than previous personalization methods. Besides, our model can also learn several new identities at once and interact with each other where the previous customization model fails to. The code will be released.

Paper

Similar papers

Peer review

Reviewer RQjb5/10 · confidence 5/52023-07-05

Summary

This paper describes a new method for personalizing pre-trained text-to-image models to generate images of unique individuals using just one facial photograph. The method analyzes and builds a celeb basis from the embedding space of the pre-trained text encoder, generating the new identity's embedding by optimizing the weight of this basis and locking all other parameters. The proposed method shows better concept combination ability than previous methods and can learn several new identities at once. The authors plan to release their code.

Strengths

1. The paper is well-written and easy to read. 2. It is an interesting approach to constructing a celeb basis from the text embedding space and optimzizing its weight to generate unique embeddings. 3. The results demonstrate the effectiveness of this approach and highlight the potential for generating highly customized images.

Weaknesses

1. Using a Celeb name basis may limit the expressiveness of the personalized model, as the generated images may be constrained by a predefined set of celebrity names. However, it is unclear what distribution of face, age, and gender was used. This information is crucial to assessing the robustness and generalizability of the proposed method, 2. The paper mentions the importance of using PCA to find the Celeb basis. However, the paper seems to lack an ablation study without using PCA to evaluate the impact of this step on the performance of the personalized model. Without such a comparison, it is difficult to assess the role of PCA in the proposed method and whether it is necessary for achieving the desired results. 3. The generated faces in the proposed method, as shown in Figure 5 for the "Anne Hathaway" example, have low quality and distortion. It could be better to provide more analysis.

Questions

see weakness

Rating

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

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

2 fair

Limitations

N/A

Reviewer Qr1K6/10 · confidence 4/52023-07-05

Summary

The manuscript with title: Inserting Anybody in Diffusion Models via Celeb Basis presents an interesting and novel way to insert identity to pre-trained text-to-image models. The paper is well-written and has clearly delivered the main concept. This approach only requires a single image with few trainable parameters, which opens a new avenue for personalized text-to-image translation.

Strengths

1. This paper presents an interesting and novel way to perform personalized text-to-image diffusion model with only one image. The key idea behind this is to map the image features to a pre-defined basis from Celebrity dataset, which is novel and works well. 2. The authors conducted a fairly amount of ablation studies, indicating the roles of each submodules, I appreciate that. 3. The results demonstrate improved visual quality compared to DreamBooth and Custom diffusion.

Weaknesses

1. One key point that I'm concerned about is how sensitive is the method to different face positions or image quality. To construct the basis, I assume that the dataset faces are well-aligned and more uniform. However, in the real cases, those conditions not necessarily hold, a face can be tilted, not well aligned, however most of the examples in the paper and the supplementary material have the input face with good alignment. I wonder if the authors can share some face generation results with poor alignments. and how to handel this issue. 2. Also I noticed that quite a large amount of the inference results are also from the celebrity dataset (correct me if I got it wrong), would appreciate more results from more personalized images - also curious about the results that are different from the celebrity dataset, for example photos of babies. 3. Lack ablation studies on the number of images. 4. I assume this framework can be generalized to not face only text-to-image generation, as long as we have a good basis, can you elaborate on this? or showing some examples of non-face image generation, for example can this framework be generalized to personalized car generation?

Questions

1. My key question is that if we want to apply this framework to a more generalized text-2-image generation, besides face generation, what do we need to do, how important is the celebrity basis. - lets say a simple experiments, what if we input a car image, where we can no longer accurately represented by the basis, what will the results be looking like?

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

1. The main limitation of this work is that it only focuses on face generation, with good alignment. This limitation may plaque the application of this work.

Reviewer MXrx7/10 · confidence 5/52023-07-07

Summary

This paper introduces an innovative method for generating highly personalized images utilizing a single reference image and distinct textual prompts.The central idea revolves around identifying a subspace within the text encoder embedding space, which, when combined, can represent any facial images. This is accomplished by applying PCA to the CLIP text embedding of hundreds of celebrity names. Subsequently, a single MLP is optimized to compute the combination weights using the features of a new individual, leveraging the diffusion reconstruction loss. The proposed approach's effectiveness is demonstrated through the results on both single and multiple subject generation tasks. The method significantly outperforms previous baselines in user studies, delivering superior qualitative results with notably enhanced diversity

Strengths

S1: The paper's innovative approach to identifying a text embedding subspace stands out. This subspace allows for an effective projection of subject-specific facial data while remaining compact, thereby avoiding overfitting to the reference subject. S2: The experimental sections of the paper are comprehensive, providing detailed comparisons via quantitative, qualitative, and user studies, along with insightful ablation studies. S3: The proposed method achieves a good balance between identity preservation and prompt consistency, outperforming popular baselines such as textual inversion and DreamBooth.

Weaknesses

W1: The authors compute the textual inversion subspaces via PCA on textual embedding of celebrity names. However, celebrities might not be a representative sample of ordinary people and could inherently carry biases related to racial and geographical factors. This might bias the subsequent image generation process. W2: The method employs subject inversion by optimizing weights in the celeb-basis. Nevertheless, the diffusion model remains static, which may result in generated samples occasionally deviating from the reference subject. W3: The method is still optimization based, and requires costly back propagation, which might be infeasible as the diffusion model gets larger and larger.

Questions

Despite the shortcomings identified above, I appreciate the overall simplicity and effectiveness of the proposed approach. Consequently, I am inclined to recommend acceptance of the paper.

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

See weakness sections.

Reviewer zJJ75/10 · confidence 4/52023-07-07

Summary

The paper presents a method to invert input portrait images to latent diffusion model (LDM)'s embedding space and then use the same to create text-based identity preserving compositional portrait images. To enable image inversion, the authors create a novel active appearance model "Celeb Basis" (image-only identity PCA) from StyleGAN-2 style frontal images. Later, this "Celeb Basis", and a textual embedding are jointly used to train CLIP for identity preserving (ArcFace style loss) image synthesis. At test time, new images of the inverted target image (represented by the Celeb Basis) can be generated based on textual prompts. The authors presents ablation studies wrt prompt and identity preservation (and detection) and compare against SOTA methods "An image is worth one word" (Textural Inversion), Dreambooth and Multi-concept customization of text-to-image diffusion (Custom diffusion) and achieve good results for prompt-based facial images. Note, that while other methods are not specific for portrait images, the current Celeb Basis has been built for portrait style images only.

Strengths

While the idea of creating an active appearance model, Cootes et al. (Celeb basis) is quite established in the vision community; this paper attempts to "rig" (e.g. StyleRig, Tewari et al.) the LDM space to create new prompted facial image given target image (identity preservation only). This is the main contribution of the paper and is novel in that regard, while the idea of controlled prompt-based image synthesis via LDM has been explored before. The celeb basis thus enables more control over image synthesis, where the semantic of "identity" preservation helps with a more _semantic_ control and results in better compositional results (two target images used together). These compositional results analysis (Fig 5, last two row) are mostly visual since I look at the overall image to argue if it looks more "believable". The paper is generally easy to follow and extensive ablation studies justify the various modeling decisions made.

Weaknesses

- The range of "Celeb basis": Since the method used PCA and uses a limited set of image to create the basis, it's limited by the choice of images, which is further limited by the choice of named celebrities in the LDM model itself. For example, it's unclear how well the model will represent certain age, pose and expression ranges. - Table 1 shows that while the users prefer the author generated images on Identity and Detect metics, Textual Inversion outperforms them, while that method is not specific to facial images as such. The reason argued is that Textual Inversion outfits. Given that Textual Inversion is a generic method, while the current method is meant for facial images, this result is surprising. - Comparison against FastComposer: Since most results presented against a concurrent work "FastComposer" seem cherry picked, the presented comparison seems biased. For example, the result presented in Fig 1 (FastComposer) seem better than those presented in Supp (Figure 5). Moreover, the current method does not explicitly model expression control in their Celeb basis, and their Fig 6 comparison seems less principled.

Questions

- The data augmentation technique is essentially used color space augmentations. Did the authors consider any geometric augmentations? Since, the image inversion techniques in the StyleGAN space are quite mature (eg HyperStyle), adding augs that change the pose or expression would make the method further robust? - How were the results shown in the paper selected? Was it the first image that the method created used or several prompt tries with subtle textual changes made to pick the results shown in the paper? - Did the authors consider the use of multi-view images for identity reconstruction?

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

2 fair

Limitations

Novelty: The main novelty of the approach is introduction of celeb-basis, while the rest of the method seems straight forward. Since the results are only shown for a facial model, it scope limits the contribution of the work. Please rename the work from "Inserting Anybody" to "Inserting Anybody's Portrait". As such control over bodies, etc. is not part of the paper's contribution. Expression control: While the basis does not model expressions in the 3DMM sense or provides a basis for expressions, the method seems to provide control expression control via prompts (supp. comparison against FastComposer). This seems to be an after thought and it's unclear how well the expressions are enabled by the method. In order to fully understand the expression control capabilities, please use a ground truth dataset (such as https://www.cs.cmu.edu/afs/cs/project/PIE/MultiPie/Multi-Pie/Home.html) and provide comparison against ground truth expression vs those generated by prompts. This would also help understand how well the expressions can be re-created. Overall, I think this is an interesting paper and the results are generally good; the paper's novelty is limited since it enables only identity inversion, and later dependent on CLIP/LDM's capabilities to create pose, expression or semantic variations (e.g. an older V_1), etc. So, there's merit in the contribution. Besides the points/questions mentioned earlier, the authors should expand the limitations, move FastComposer results to the main paper.

Reviewer ABqC7/10 · confidence 2/52023-07-11

Summary

The paper proposes a new method to personalize the pre-trained text-to-image model on humans. The propose a novel idea called Celeb basis and it simplifies the problem to reconstructing the coefficients of any new identity. Both quantitative and qualitative results demonstrate the effectiveness of the proposed method on concept composition compared with the baselines.

Strengths

+ It's a very smart idea to utilize Celeb Basis from the text embedding space of the celebrities’ names in the text-to-image model. With the low dimension subspace, the learning problem boils down to very few coefficients, which is simple yet effective. + The experimental results speak for the paper itself. It shows more stable concept composition abilities than previous works. I especially like fig 5 where the motivation of the proposed method is clearly demonstrated and easy to see where the previous methods fall short on.

Weaknesses

N/A

Questions

will the codebase be publicly available and when?

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

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

I think authors have adequately discussed the limitations in terms of the artifacts of the real human faces of the original stable diffusion model and other species. Potential negative impact has also been discussed.

Reviewer RQjb2023-08-12

Thanks for the rebuttal, which do help clarify some of the issues. Howver, I think Figure 2 in rebuttal pdf shows more examples only?

Authorsrebuttal2023-08-12

Thanks for your carefully check. The `Figure 2 in Rebuttal PDF` (attached in the **Global Response** part) is a **histogram plot** showing the $691$ celeb names' frequency statistics in LAION-2B text samples (LAION-2B/5B dataset is used to train Stable Diffusion [1], which contains $\sim5.5$ billion image-text pairs), where the x-axis is the index of the name, the y-axis is the appearing time of this name in the dataset. This plot indicates that our collected celeb names correspond to thousands image-text paired samples in average. There are only $25$ names whose frequencies are less than $100$ and $173$ nemes whose frequencies are larger than $10000$. This frequency distribution shows that the distribution of the training is relatively fair. We would also be very willing to provide the detailed analysis of the distributions of each attribute, including age, gender, and race as you mentioned. However, we find it needs tones of time to crawl all the coorresponding celeb **___images___** from LAION dataset. For example, analyzing all the **text names** from LAION dataset will require $1$ day longer. And it will cost over $7$ days to download the corresponding **images** (just downloading, without analyzing the attributes), which seems impossible to be accomplished in such a short rebuttal time. We will also continue to download images and analysis these attributes and put them in the revised paper or our GitHub pages in the future to support our claim. Although without providing the age, gender, and race distribution, the qulitative results in `Rebuttal PDF Figure 1` along with the statistic in `Rebuttal PDF Figure 2` already demonstrate the robustness and generalizability against large age, gender, and pose variation of our method. And also as stated in **R5Q1**, we leave more detailed distribution anlysis as the future work to discover more about the bias in celeb basis.

Reviewer MXrx2023-08-17

Response

Thank you for the rebuttal. Regarding points 1 and 2, celebrities represent a minor fraction of the overall population. I would recommend that the authors address this limitation in the finalized version. For point 3, I concur that integrating an encoder-based approach with the celeb-based concepts presented in this paper poses certain challenges, which can be explored in future works. I suggest the authors to add some discussion about this and add quantitative comparisons (image quality / efficiency) with encoder-based methods like ELITE and FastComposer as other reviewers suggest. On the whole, I remain optimistic about the submission and keep my original rating.

Reviewer zJJ72023-08-17

Thank you for the rebuttal. I am generally happy with the author replies, and keep my "acceptance" rating.

Area Chair y5r52023-08-19

Thanks to the authors for your response. Best regards, Your AC

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC