Object-Centric Slot Diffusion

The recent success of transformer-based image generative models in object-centric learning highlights the importance of powerful image generators for handling complex scenes. However, despite the high expressiveness of diffusion models in image generation, their integration into object-centric learning remains largely unexplored in this domain. In this paper, we explore the feasibility and potential of integrating diffusion models into object-centric learning and investigate the pros and cons of this approach. We introduce Latent Slot Diffusion (LSD), a novel model that serves dual purposes: it is the first object-centric learning model to replace conventional slot decoders with a latent diffusion model conditioned on object slots, and it is also the first unsupervised compositional conditional diffusion model that operates without the need for supervised annotations like text. Through experiments on various object-centric tasks, including the first application of the FFHQ dataset in this field, we demonstrate that LSD significantly outperforms state-of-the-art transformer-based decoders, particularly in more complex scenes, and exhibits superior unsupervised compositional generation quality. In addition, we conduct a preliminary investigation into the integration of pre-trained diffusion models in LSD and demonstrate its effectiveness in real-world image segmentation and generation. Project page is available at https://latentslotdiffusion.github.io

Paper

Similar papers

Peer review

Reviewer QNuS7/10 · confidence 5/52023-06-08

Summary

This paper proposes Latent Slot Diffusion (LSD), an object-centric learning framework combining the Slot Attention module and a Latent Diffusion Model (LDM) based slot decoder. The model is trained in an auto-encoding manner, where the loss is the slot-conditioned denoising loss in the LDM-based decoder. Extensive experiments demonstrate the effectiveness of LSD in 1) unsupervised scene decomposition 2) object property prediction 3) image generation/editing, surpassing previous SOTA SLATE which uses a auto-regressive Transformer-based slot decoder.

Strengths

- The paper is well-written and easy to follow. The figures are of high quality - The use of LDM as the slot decoder is very intuitive and reasonable, considering the recent trends in object-centric models (CNN decoder --> dVAE+Transformer decoder --> KL-VAE+LDM decoder). The results verify this design choice - The experimental results cover lots of different tasks, and show very strong performance. The improvement in generation capacity is very impressive compared to baselines

Weaknesses

I don't see any big issue with the paper. One might consider LSD as "simply replacing the slot decoder with an LDM", but I think this modification is reasonable, and works pretty well across datasets and tasks. SLATE/STEVE also just replace the CNN decoder in Slot Attention with Transformer decoders, and they have been widely used due to their strong unsupervised segmentation performance. I believe LSD's improvement in generation quality will also facilitate future research in object-centric generative models. That being said, I think the paper should have more discussion of its limitations. See the `Questions` and `Limitations` sections below.

Questions

1. Regarding the pre-trained VAE model: - I appreciate the authors' efforts in having SLATE+ with the same VAE for a fair comparison. Have you tried training a VAE from scratch, and comparing this LSD variant with SLATE (which also trains dVAE from scratch)? 2. Regarding the LSD + Stable Diffusion (SD) experiments in the Appendix (**not affecting paper decision**): - This is an interesting experiment, especially the results on the real-world COCO dataset. Do you have any numerical results to compare with the SOTA object-centric model DINOSAUR [1]? - Can the authors show some generation or reconstruction results? I am curious because I believe even LSD cannot generate realistic unconstrained real-world images (FFHQ images are constrained as they only capture human faces). Will using a pre-trained SD decoder help here? [1] Seitzer, Maximilian, et al. "Bridging the gap to real-world object-centric learning." ICLR. 2023. 3. Minor question (**not affecting paper decision**): a [concurrent work](https://arxiv.org/abs/2305.11281) you might want to cite, but I understand that paper comes out after the submission deadline.

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

4 excellent

Contribution

3 good

Limitations

The authors discuss limitations in Section 6. However, there are more limitations of this work (and the entire object-centric learning field): - LSD is still unable to generate realistic unconstrained real-world images - LSD suffers from the part-whole ambiguity issue, as can be seen from the COCO examples I'd like to see the authors discuss these points in the paper.

Reviewer ZjQw7/10 · confidence 4/52023-07-01

Summary

Previous work has shown that transformer-based image generative models can be trained for object-centric learning which can handle complex scenes. That is, given unstructured observation, transformer-based models learn to find latent compositional structures to bind relevant features. This work makes an attempt to explore the feasibility and potential of integrating diffusion models into object-centric learning. A major contribution of this work is a novel Latent Slot Diffusion (LSD) model which combines conventional slot decoders with a conditional latent diffusion model, conditioned on object-centric slots provided by Slot Attention. The authors have shown the effectiveness of the proposed model by evaluating several object-centric tasks including unsupervised object segmentation, downstream property prediction, compositional generation, and image editing.

Strengths

- **Scope and relevance**: Considering the growing interest in diffusion models, this paper is exceptionally timely as it broadens their application towards object-centric learning. - **Significance of contributions**: This paper presents a novel diffusion model for object-centric learning without the requirement of supervised annotations. It answers the raised question about how diffusion-based generative modeling can benefit object-centric learning properly. - **Experimental results**: The experiments in this paper are enough to prove the effectiveness of the proposed model. - **Clarity**: The main body of the paper is written very well.

Weaknesses

- **Limited technical contribution**: The proposed slot-contained diffusion is a trivial extension of existing text-to-image generation in Latent Diffusion Models (LDM) [59], which replaces text inputs with latent slots. - **Unclear model details**: To my understanding, the whole model is trained for image generation objectives. How to apply such a generative model to unsupervised object segmentation is not clear. - **Unclear training details**: if I am not mistaken, the object-centric encoder is trained independently rather than jointly with the latent slot diffusion decoder. I would suggest the authors clarify this. Thus, control experiments might be needed to study the effect of training strategies. - **Some implementation details are missing**: Without providing the source code, some important hyperparameters are not presented to justify the reproducibility of this work, eg., the number of $K$ for k-means cluster, noise schedule $\alpha$, and steps $T$.

Questions

In addition to the above weaknesses, here are two more questions: - In line 231, given that the object segmentation masks are derived from the attention masks of Slot Attention, why using diffusion-based generative modeling can help unsupervised object segmentation? - The LSD model has been evaluated across multiple tasks and outperforms existing state-of-the-art methods. However, it raises the question: to what extent does the enhanced performance truly stem from the diffusion-based generative modeling rather than the improved image encoder? It's an intriguing point to ponder. Overall, this paper is a good effect. I will raise my rating if the authors can address my concerns.

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

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

The authors did not adequately address the limitations of this work. A significant constraint of this study is the requirement for two distinct visual encoders: a pre-trained image auto-encoder and an object-centric encoder. This dual requirement results in a relatively heavy model for inference. This leads to the question: Is there any potential to share certain components between these two encoders to enhance efficiency?

Reviewer TTEN7/10 · confidence 3/52023-07-03

Summary

The paper studies the diffusion models into object-centric learning. The authors introduce the concept of latent slot diffusion (LSD) that can replace slot decoders conditioned on object slots. The model can work in an unsupervised compositional mode without requiring annotations such as text. Their experiments show that LSD performs better in complex scenes comparing with other methods that do unsupervised compositional generation.

Strengths

The proposed method LSD can be viewed as a model substituting conventional slot-decoders with a conditional latent diffusion model, where the conditioning is done via the object slot attentions. It can also be viewed as a unsupervised conditional compositional diffusion based generation. The ablation experiments are helping in explaining the crux.

Weaknesses

Object segmentation using LDS is yet to be perfected, leading to suboptimal downstream applications.

Questions

What are potential ways to improve segmentation using LSD?

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

3 good

Contribution

3 good

Limitations

As a generative model, the method needs to consider the privacy and impacts of image manipulation.

Reviewer uRVS8/10 · confidence 5/52023-07-04

Summary

This work introduces a methodology for using diffusion-based models to obtain object-centric representations. This is done using slot representations, obtained from the Slot Attention module applied to the original image, as a conditioning variable for a latent diffusion model. The diffusion model and the Slot Attention module are trained end-to-end. The proposed model is called Latent Slot Diffusion (LSD) and is evaluated on unsupervised object segmentation, downstream property prediction, compositional image generation, and image editing. The datasets used for evaluation are ClevrTex, MOVi-E, MOVi-C and, for the first time in the field, FFHQ. The performance of LSD is evaluated quantitatively on the first 4 tasks and is compared against SLATE and SLATE+. The proposed model outperforms both baselines in the tasks according to most metrics.

Strengths

**Originality**. The paper proposes a novel way to combine two existing models: slot attention and latent diffusion. The presentation clearly shows what are the novel elements. **Quality**. The method shows a successful way to leverage diffusion in object-centric learning. Especially the results on property prediction show that it has concrete benefits for the representation learning itself, which might be useful for several different tasks, not only the ones shown in the paper. **Clarity**. The submission neatly shows all the experiments that were carried out and the description of the underlying method is clear. **Significance**. The work is a necessary exploration of leveraging the generative power of the diffusion approach in the object-centric setting, highlighting the complexities related to having too-strong of a decoder and providing clear examples of its strengths, especially by using a very complex dataset (FFHQ).

Weaknesses

**Quality**. The lack of error bars (i.e. standard deviation) in the analysis makes the quantitative analysis weaker. Additionally, it would be interesting to further explore the problems with FG-ARI, as it is currently the standard metric used in the field, and although I agree with the statements, this is only based on experience and intuition, and not a proper scientific analysis, which has not been carried out yet. Lack of comparison with traditional models such as the improved Slot Attention architecture proposed in [1] makes the performance of the model harder to evaluate. Are the good results obtained primarily due to the improved architecture (e.g. larger encoders, better decoder), or is there something fundamentally good about using diffusion (e.g. the iterative improvement) that results in better object representations? **Clarity** The description of the method used for unsupervised object segmentation is lacking. There is only a reference to the use of the attention masks from slot attention, it would be much better to refer directly to the nice mathematical notation used before in the text. *References* 1) Biza, Ondrej, et al. "Invariant Slot Attention: Object Discovery with Slot-Centric Reference Frames." arXiv preprint arXiv:2302.04973 (2023).

Questions

As object segmentation is performed using the masks from the slot attention module, it is clear that it is not possible to obtain masks with the full resolution of the images in many cases. Could this be an area of improvement for future work or is it possible to already try using some unsupervised super-resolution technique to get higher-quality masks? Could be worth considering object segmentation as a visualization of what the model is doing instead of a separate task? Do you have any insights into how the model would perform if slot attention is applied to the latent representation of the latent diffusion model instead of the original image? This would help clarify the choices made during the development of the method.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

The authors have sufficiently addressed the limitations of the proposed model, as well as the broader impact of their approach. However, this can be improved further, by being more explicit about direct effect of ethnic bias, ageism or other forms of mis-representation that the model could lead to, which is standard practice for modern diffusion model, considering the extreme high-quality of the generated images. Direct manipulation of certain characteristics enables these effects in a much easier way.

Reviewer QNuS2023-08-10

Rebuttal Acknowledgment

I thank the authors for the response. I maintain my original rating of Accept after reading reviews from other reviewers.

Reviewer ZjQw2023-08-11

Rebuttal Acknowledgment

I thank the reviewers for their detailed clarification, which has released my concerns about their work. Therefore, I raise my rating to accept.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC