Interpreting the Weight Space of Customized Diffusion Models

We investigate the space of weights spanned by a large collection of customized diffusion models. We populate this space by creating a dataset of over 60,000 models, each of which is a base model fine-tuned to insert a different person's visual identity. We model the underlying manifold of these weights as a subspace, which we term weights2weights. We demonstrate three immediate applications of this space that result in new diffusion models -- sampling, editing, and inversion. First, sampling a set of weights from this space results in a new model encoding a novel identity. Next, we find linear directions in this space corresponding to semantic edits of the identity (e.g., adding a beard), resulting in a new model with the original identity edited. Finally, we show that inverting a single image into this space encodes a realistic identity into a model, even if the input image is out of distribution (e.g., a painting). We further find that these linear properties of the diffusion model weight space extend to other visual concepts. Our results indicate that the weight space of fine-tuned diffusion models can behave as an interpretable meta-latent space producing new models.

Paper

Similar papers

Peer review

Reviewer 215A6/10 · confidence 5/52024-07-02

Summary

The paper proposes to learn a manifold over customized diffusion model weights as a subspace for interpretable downstream applications such as sampling, editing, and inversion.  At the technological level, the authors collect the weights of over 60,000 models as the dataset and fine-tune those weights via LoRA to constitute the desired editing space. Experiments are conducted on CelebA dataset.

Strengths

- The paper is generally well organized and easy to follow, with reasonable motivation and intuitions. - The idea of using weights as the dataset to constitute an editing space is interesting. - Among the experimented application scenarios, while editing and inverting have been widely studied in the field (see Weaknesses), the exploration of the OOD projection seems more interesting to me.

Weaknesses

- The paper lacks a comprehensive comparison with prior literature, resulting in inaccurate claims on “Alas, in multi-step generative models, like diffusion models, such a linear latent space is yet to be found”. In [a], it has been explicitly and theoretically proven such interpretable and linear latent space can be found in frozen DMs, without the need to fine-tune the model parameters, neither the DreamBooth tuning nor LoRA-based. Recent studies [2,3] have extended these findings for customization tasks and latent diffusion models. These methods have covered the scenarios including latent sampling, editing, and inverting, and thus should be discussed. - From the technical point of view, the authors put much more emphasis on the empirical investigations on the interpretability of latent spaces in DMs, but neglect the dimension of diffusion time steps from a more fundamental perspective. It would strengthen the paper if the authors could elaborate on the impact of time steps on the construction of your w2w space because multiple existing works [a,b,d] suggest the time step of latent spaces in DMs could carry important information on the distribution learning process and thus influence the downstream tasks.  - From the experiments and evaluation perspective, it would be helpful to perform some user studies as subjective evaluations as many other generative application works do. In addition, these attribute directions and their properties have been demonstrated other than identity context, it may strengthen the paper if the authors could extend their application scenarios to a wider context. --- [a] Boundary Guided Learning Free Semantic Control with Diffusion Models, NeurIPS 2023. [b] Diffusion in Diffusion: Cyclic One-Way Diffusion for Text-Vision-Conditioned Generation, ICLR 2024. [c] Continuous, Subject-Specific Attribute Control in T2I Models by Identifying Semantic Directions. 2024 [d] Diffusionclip: Text-guided diffusion models for robust image manipulation. CVPR 2022.

Questions

Please see my comments in the Weaknesses section.  I have mixed feelings about this paper. On the one hand, the paper risks over-claiming on some high-level motivations as pointed out in the Weaknesses section, lacking a clear positioning concerning several important related works. The methodological level design is also not entirely new and intriguing. On the other hand, I acknowledge that using model weights as datasets is interesting, and the efforts of “scaling up” to over 60.000 models seem to be one of the biggest contributions to me.

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

The limitations and broader societal impact are discussed in the main paper.

Reviewer wMSF5/10 · confidence 3/52024-07-08

Summary

The paper investigates the weight space spanned by a large collection of customized diffusion models, proposing a novel subspace termed weights2weights (w2w). The study populates this space with a dataset of over 60,000 models, each fine-tuned to encode a different person's visual identity. The paper demonstrates three main applications of this space: sampling novel identities, editing existing identities, and inverting a single image to reconstruct realistic identities, showing that these weight spaces can serve as a highly interpretable latent identity space.

Strengths

1. The concept of modeling a manifold over customized diffusion model weights is novel and shows potential for broad applications in generative modeling. 2. The PCA and LoRA are used to construct the w2w space. Furthermore, the PCA can be used to edit the image attributes. The method is novel and effective. 3. The paper includes extensive experiments that demonstrate the effectiveness of the proposed w2w space in generating, editing, and inverting identities.

Weaknesses

1. The construction of w2w space needs over 6000 weights and the number of the model can further improve the ID score. The computing resources should be provided. And is there a contained solution to balance computing resources and performance? 2. The paper uses several evaluation metrics (e.g., identity preservation, LPIPS, CLIP score), some metrics like CLIP score can describe the quality of image generation, but might not fully evaluate the nuanced attributes related to human faces. 3. In the weights manifold, new models can be sampled to find directions that separate identity properties. However, are the first m principal components all identity attributes, and do they correspond one-to-one?

Questions

Please refer to the weaknesses above.

Rating

5

Confidence

3

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors have discussed the limitations of this work.

Reviewer rzPz7/10 · confidence 5/52024-07-11

Summary

This paper constructs a dataset that contains LoRA models trained on images of 60,000 different people. A weight manifold is determined based on the parameters of these 60,000 models using PCA. Sampling, editing, and inversion can be performed on this manifold, and the rationality of this manifold is demonstrated.

Strengths

1. This paper constructs a dataset that contains LoRA models trained on images of 60,000 different people. 2. This paper demonstrates the existence of a LoRA model manifold that encapsulates identity information and enables the generation of new identities by sampling from this manifold. 3. The paper also presents methods of editing and inversion, expanding the scope of manifold applications.

Weaknesses

1. There are some recent works present the way of generating identity preserved images given only one reference image. It is better to provide the comparison between this paper and these works. 2. LoRA training often affects the performance of the base model to some extent, such as small generation variance. More results to examine this effect will be better.

Questions

Refer to the Weaknesses 1 and 2.

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes, limitations are discussed in this paper.

Reviewer aBHC5/10 · confidence 4/52024-07-12

Summary

This paper explores the latent space of weights in customized diffusion models, introducing the weights2weights (w2w) space, a subspace encoding different human identities. By fine-tuning over 65,000 models, each representing a distinct human identity, the authors model this weight space using low-rank adaptation (LoRA) and principal component analysis (PCA). The key contributions include demonstrating three applications of the w2w space: Sampling: Generating novel identities by sampling from the weight space; Editing: Performing semantic edits, such as adding a beard, by traversing linear directions in the weight space; Inversion: Reconstructing a realistic identity from a single image, even if it is out-of-distribution. The paper validates the w2w space’s expressiveness through quantitative evaluations and qualitative observations.

Strengths

1) The w2w space is interesting because it shifts the focus from traditional pixel or feature spaces to the model’s weight space, allowing for new forms of manipulation and analysis. 2) The paper is well-organized and clearly written, with detailed explanations of the proposed methods and their theoretical underpinnings.

Weaknesses

1) The overall idea of this paper is more similar to [1], the significant difference I see so far is that [1] does PCA on token embedding and w2w does PCA on Lora in Unet. I suggest the authors could discuss more about the differences between this paper and [1]. Also, I have some methodological doubts: A. does w2w have a similar phenomenon to mean person in [1]? (see Figure 2 in [1]) B. Have the authors studied the effect of PCA on the generation of results? (see section 4.3 in [1]) [1] Inserting Anybody in Diffusion Models via Celeb Basis 2) While this paper provides new ideas and insights into personalized generation models, I am concerned about his real-world application, what are the advantages of w2w compared to existing personalized generation efforts such as IP-adapter, InstabntID, custom diffusion, Mix-of-show, etc.? Faster generation or smaller training parameters? The experiments also don't give a comparison with existing personalization generation efforts, however, I think this is important to note

Questions

1) I'm curious if style information can be decoupled via w2w space in addition to attributes. I would suggest that the authors could add the results of style editing to their paper. 2) How does w2w perform on multi-concept generation?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The paper discussed the limitations of this work.

Reviewer aBHC2024-08-13

post rebuttal

Thanks to the authors for their efforts during the rebuttal. The author's response solved part of my problem. But I still have a lot of concerns that haven't been clarified. 1. The biggest concern is that collecting a large number of weights to build the w2wk space is far more complex than collecting a large number of images. In addition, in terms of the number of parameters, Celeb-Basis only has 1024 learning parameters. 2. And for question 1, what guarantees that information like cartoon drawing is not encoded in the principal components? For example, the weight information represents a Buzz Lightyear. Wouldn't Pixar style and Buzz Lightyear images appear at the same time? Then, why is the new image created based on the weight information not affected by the Pixar style? 3. For question 2, perhaps my question is ambiguous and I am sorry for this, I am actually concerned with the generation of multi-subject, whether the identity information of each person can be maintained when generating multiple characters at the same time (I only saw two case shows in Figure 6 and 7 in the paper).

Authorsrebuttal2024-08-13

We appreciate your response and are happy to clarify further. **“Biggest concern is the collecting a large number of weights…”** We agree that collecting a large number of model weights is more complex than a large number of images. However, the bigger picture of this work is the “idea of using weights as the dataset” (215A) in order to demonstrate "the existence of a LoRA model manifold” (rzPz). As mentioned in the global response, we introduce this idea of a **meta**-latent space **creating new models**, with “broad applications in generative modeling” (wMSF), which we believe is beyond just the personalization application. We would like to note that although creating this space requires a notable amount of compute, using the space for the variety of applications is lightweight. We discuss these practical details below. **“...in terms of the number of parameters, Celeb-Basis only has 1024 learning parameters.”** While Celeb-Basis uses 1024 coefficients to define an identity in token space, it requires optimizing an MLP at test-time with 525,312 parameters and backpropagation through the text-encoder, leading to \~3 minutes and \~21GB VRAM on a single NVIDIA A100. In contrast, we directly optimize 10,000 coefficients (\~2% of Celeb-Basis's parameters) without text-encoder backpropagation, reducing optimization time to \~1 minute and VRAM usage to \~8GB. Furthermore, our studies (Table 1 and 2 in the rebuttal PDF) show *w2w* offers better identity preservation without overfitting. We would also like to note that Celeb-Basis works in token embedding space, which is already fairly low dimensional. In their case, they use 1024 dimensions to span the original 1536 dimensional space (768 dimensions for first name + 768 for last name), resulting in ~66% of the original dimensionality. In contrast, we compress a much higher dimensional LoRA space (100,000 dimensions) to 10,000 dimensions (10% original dimensionality). The fact that even LoRAs, which are by design low rank, have significantly lower rank semantic subspaces is quite intriguing, potentially inspiring further exploration of interpretable and controllable subspaces in model weights. **“...what guarantees that information like cartoon drawing is not encoded in the principal components?”** When you say, “the weight information represents a Buzz Lightyear,” the weights of the base Stable Diffusion model encodes prior concepts such as “Buzz Lightyear,” “Pixar,” etc, but not the LoRA weights added on top. During Dreambooth fine-tuning, we train LoRAs on realistic human identities, so the only new inserted concept is the identity. We apply PCA to the LoRAs, not the base model, and since there’s no style variation among the encoded identities, no principal components encode style. However, if “style-editing” means generating images of the identity in different styles with prompts, that’s possible because the base model retains prior concepts. For example, in Fig. 7 and Fig. 8, we generate a person as a statue or painting using prompts. Furthermore, in rebuttal PDF Fig. 8, we show that a separately trained style LoRA can be merged with an identity model from *w2w* space. **“... whether the identity information of each person can be maintained when generating multiple characters at the same time…”** Thank you for the clarification. Identities from *w2w* space can be generated alongside existing characters or concepts in the base diffusion model (e.g., the celebrity examples in Figures 6 and 7 of the paper). However, two *w2w* models can't be naively merged by adding weights, as they share the same subspace and will interfere, leading to identity interpolation. Separate LoRAs can be trained to encode identities and then be merged with *w2w*, but they cannot live in the same subspace. This can be enforced using something like [1]. On the other hand, a LoRA encoding another concept such as style can be seamlessly merged as seen in rebuttal PDF Fig. 8. We will discuss this as a limitation in the paper. **Summary** We hope we addressed your main concerns. Overall, we have shown from an efficiency and performance standpoint, that *w2w* performs stronger than Celeb-Basis as well as other baselines. More importantly, as you mention in your original review, our work provides “new ideas and insights into personalized generation models.” We believe that the merits of our work should not be subsumed by Celeb-Basis just because of the personalization application. If that were the case, then works like Dreambooth can be thought of as just Textual Inversion but optimizing model weights instead of token embeddings. But works like Dreambooth provide further insights beyond the application. [1] “Orthogonal Adaptation for Modular Customization of Diffusion Models.” Po et al. 2024

Reviewer aBHC2024-08-14

Thank you for clarifying the doubts. Most of my concerns have been addressed. I will update my rating. However, I still think the use of weights does introduce a large amount of difficulty in preprocessing. Also how to merge multiple LoRa is a big challenge now, and the token-based approach can easily generate multiple roles.

Authorsrebuttal2024-08-10

We thank all reviewers for their time and valuable comments on our work. We are grateful that this paper has overall been well-received by the reviewers. As the end of the discussion period is approaching, we kindly ask all reviewers to acknowledge if our rebuttal has addressed their concerns and give us an opportunity to address any further follow-up. Thank you again for your participation.

Authorsrebuttal2024-08-11

Thank you for your further comments. We are happy to hear that we sufficiently answered some of your concerns, and we appreciate you raising your score. We agree with you that the weight space and noise space are not completely independent. This can make for some interesting future work.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC