SPAE: Semantic Pyramid AutoEncoder for Multimodal Generation with Frozen LLMs

In this work, we introduce Semantic Pyramid AutoEncoder (SPAE) for enabling frozen LLMs to perform both understanding and generation tasks involving non-linguistic modalities such as images or videos. SPAE converts between raw pixels and interpretable lexical tokens (or words) extracted from the LLM's vocabulary. The resulting tokens capture both the semantic meaning and the fine-grained details needed for visual reconstruction, effectively translating the visual content into a language comprehensible to the LLM, and empowering it to perform a wide array of multimodal tasks. Our approach is validated through in-context learning experiments with frozen PaLM 2 and GPT 3.5 on a diverse set of image understanding and generation tasks. Our method marks the first successful attempt to enable a frozen LLM to generate image content while surpassing state-of-the-art performance in image understanding tasks, under the same setting, by over 25%.

Paper

Similar papers

Peer review

Reviewer 1ht37/10 · confidence 4/52023-07-05

Summary

This paper introduces the Semantic Pyramid AutoEncoder (SPAE) for enabling frozen Large Language Models (LLMs) to perform understanding and generation tasks involving non-linguistic modalities such as images or videos. SPAE converts raw pixels to lexical tokens extracted from the LLM's vocabulary, allowing the LLM to perform multimodal tasks. The proposed method is evaluated using in-context learning experiments with frozen PaLM 2 and GPT-3.5, demonstrating success in image understanding and generation tasks. The main contributions include the first successful method using a frozen language model for image content generation, a new SPAE tokenizer producing interpretable representations, a new progressive prompting method for long cross-modal sequences, and evaluation on visual understanding and generation tasks.

Strengths

1. Compared to LQAE, the method presented in this paper demonstrates superior performance in image reconstruction. 2. By interacting with large language models such as GPT-3.5 and PaLM-2, this paper effectively utilizes the in-context learning capabilities of these models to generate corresponding images, offering an intriguing research direction.

Weaknesses

1. This paper's experiments are limited to 128x128 resolution, significantly lower than the mainstream 256x256 clarity. 2. To achieve comparable performance with VQGAN, a larger latent space (# tokens) is necessary. 3. Is it possible to attain the same performance as VQGAN + frozen codebook + semantic guidance and an expanded latent space (# tokens)? The need for a pyramid structure comes into play here. Whether the pyramid structure is necessary requires experimental support. 4. The VAE proposed in this paper employs perceptual loss and utilizes a VGG network with supervised pre-training on ImageNet during the training process. In contrast, other methods like LQAE do not use perceptual loss. A strategy similar to ViT-VQGAN should be adopted here, using a version without perceptual loss for training on the comprehension task. 5. There is a lack of model variants in the ablation experiments on the understanding tasks. 6. In lines 203-204, it is mentioned that "We train with a batch size of 256 for 450k steps, which takes 1.4k TPUv3-hours." This training cost is larger than that of both VQGAN and RQ-VAE. It would be interesting to investigate the impact of the training cost on the performance of the VAE proposed in this paper.

Questions

Figure 8 in the appendix is intriguing, but the experiments are only conducted on the MNIST dataset. It would be compelling to see the experimental results on datasets like MS-COCO.

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

2 fair

Contribution

4 excellent

Limitations

addressed

Reviewer 2BCt6/10 · confidence 4/52023-07-05

Summary

This paper proposes Semantic Pyramid AutoEncoder (SPAE) which enables frozen LLMs to perform both understanding and generation tasks involving non-linguistic modalities such as images or videos through in-context learning. The authors first introduce SPAE, a pyramid tokenizer that produces interpretable representations of semantic concepts and fine-grained details. In addition, the authors also propose a new progressive prompting method that facilitates the in-context generation of long cross-modal sequences. The proposed method outperforms SoTA in few-shot image classification accuracy by a large margin.

Strengths

- The proposed approach does not require further tuning of LLM. It is thus very portable. - The paper is well-written and the approach, motivations, and experiments are well-derived and clear. - The paper presents an interesting algorithm to leverage LLMs to address non-linguistic tasks, such as how to combine pretrained models and how well it would perform, even without tuning LLM and its codebook. - Impressive few-shot results on image classification.

Weaknesses

Although this paper has shown how a frozen LLM can generate image content or realize visual understanding tasks. The tasks the authors choose to evaluate in this paper are still limited (only few-shot image classification on ImageNet). Can the authors provide more quantitative experiments to validate the SPAE’s generalization to various tasks (e.g., caption, VQA…)? Besides, there are only qualitative examples of image-to-image generation, which only prove that SPAE “can” generate image content, not that it can generate well.

Questions

please refer to the weakness part.

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

The authors have described the limitations of this work.

Reviewer mRhs5/10 · confidence 4/52023-07-07

Summary

This paper presents a novel multi-modal generation method named SPAE which leverages semantic autoencoder and in-context denoising for semantic reconstruction. The SPAE converts raw pixels into interpretable lexical tokens from the LLM's vocabulary, capturing both semantic and fine-grained visual details. The paper conducts in-context learning experiments using frozen PaLM-2 and GPT 3.5 on a diverse set of image understanding and generation tasks for showing the effectiveness.

Strengths

● Semantic Pyramid AutoEncoder (SPAE) is introduced for enabling frozen LLM to perform both understanding and generation tasks involving non-linguistic modalities, such as images or videos. ● Different from previous work, SPAE produces interpretable representations of semantic concepts and fine-grained details in the form of multilingual linguistic tokens by leveraging frozen LLM. ● The proposed method outperforms the best-published few-shot image classification accuracy by an absolute 25% under the same in-context setting.

Weaknesses

1. The 5-shot evaluation on ImageNet classification is somewhat unfair. It is not clear about the performance with GPT-3.5 under this setting. The paper only reports the results of PaLM-2. 2. The hyperparameters in the paper are quite tricky (e.g. threshold for each layer, the weight of the loss). Besides, the effect of dynamic weight used in the final objective is not ablated. 3. Since SPAE is only trained on ImageNet, it is not unknown whether this method scales well on large-scale datasets (e.g. Laion-400M, Laion-2B, COYO-700M, etc).

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

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 paper demonstrates the generation under the low-resolution setting. The higher resolution (e.g. 1024x1024, 2048x2048) remains unexplored. Can you provide some comparisons with diffusion, such as high resolution and complex scenes?

Reviewer ekNF6/10 · confidence 3/52023-07-07

Summary

The paper introduces SPAE, a method that aligns visual representation with a fixed LLM representation. SPAE effectively captures semantics and visual fine-grained textures for a range of cross-modality tasks. The paper showcases good and solid results in few-shot learning and reconstruction/generation tasks.

Strengths

- The paper is well written and well motivated. - The proposed idea is interesting, cross-modality with codebook is not a novel idea, but using the fixed pre-trained LLM as dictionary is an interesting idea and seems working well. - The results are solid and convincing.

Weaknesses

- The quantitive comparisons are not fair comparisons; as the language model sued in SPAE is much bigger and trained on much more data than the baselines. One interesting baseline is to replace the heavy LLM to smaller LMs and discuss the impact. - Lack of comparisons to other works on generation tasks. It would be good to include comparison to other generative models with same prompt.

Questions

- Is it possible to perform zeroshot tasks with the visual encoder and LLM, it would be a good example to compare SPAE with other contrastive based models. - How critical is the LLM used in this paper, can we switch LLM with other language models and how will that impact the performance?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

None

Authorsrebuttal2023-08-15

A Kind Reminder on Reviewer-Author Discussions

Dear Reviewers, Thanks again for your effort and valuable feedback during the review process. We have made our responses and hope they address your concerns. Please let us know if you have further questions after reading our rebuttal. We hope to address all the potential issues during the discussion period. Best, Authors

Reviewer ekNF2023-08-16

Thanks for the rebuttal

The work is very interesting, and the rebuttal addressed most of my concerns.

Authorsrebuttal2023-08-16

Thank you for the comments

Thank you for the comments. We appreciate that you would consider increasing the rating as all concerns are addressed.

Reviewer mRhs2023-08-17

I appreciate the authors' response. Their rebuttal addressed part of my concerns. This work is quite interesting, but whether it can generalize to scaling and high resolution still needs to be validated. Therefore, I still choose to maintain my previous score.

Authorsrebuttal2023-08-17

Request for Elaboration: Thanks for the Feedback – Kindly Seek Clarification on Reviewer's Comment

We are thankful for the reviewer’s feedback and assessment that our “work is quite interesting”. We would appreciate if the reviewer could elaborate on the comment “whether it can generalize to scaling and high resolution still needs to be validated”. We have presented the results of **256x256 images using SPAE trained on ImageNet21K (~13 million images)** (see Table R2.3 in the rebuttal) **[Scaling]** We have demonstrated that training SPAE on 13 million images can be completed within a matter of 1-2 days . The results may already provide substantial validation of scalability given that training on larger datasets such as Laion-400M would merely entail increasing the training time by a factor of 30 or deploying additional GPU/TPU devices. **[High resolution]** We would like to provide more clarifications. In the text-to-image models (e.g., Parti, MUSE, Latent Diffusion), the native image generation resolution has typically been 256x256. To obtain high-resolution, such as 1024x1024, it would require separate super-resolution models. In Fig. 17 of [the rebuttal pdf](https://openreview.net/attachment?id=sIoDsBBe1G&name=pdf), we have presented the results of 256x256 images which can be used as inputs to the super-resolution model. Therefore, our results may validate that our method can scale to high resolution images as incorporating super-resolution apparently falls beyond the scope of our paper We kindly request the reviewer to carefully assess whether we have addressed the concern raised regarding scaling and high resolution. Please let us know for additional questions. Thank you

Reviewer 2BCt2023-08-18

Thank you for the authors' responses. They have addressed my comments. I therefore keep my initial positive rating.

Authorsrebuttal2023-08-18

Thank you.

Thank you for the comments. As we have addressed your concerns, we appreciate it if you consider raising your score.

Reviewer 1ht32023-08-19

Re: Rebuttal by Authors

Overall, the reviewer is satisfied with the author's response, particularly the experimental results presented in Table R4.1 and Table R4.2. The author is requested to include these elements in the final version of the manuscript. The reviewer has increased the rating.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC