Unlocking the Capabilities of Masked Generative Models for Image Synthesis via Self-Guidance

Masked generative models (MGMs) have shown impressive generative ability while providing an order of magnitude efficient sampling steps compared to continuous diffusion models. However, MGMs still underperform in image synthesis compared to recent well-developed continuous diffusion models with similar size in terms of quality and diversity of generated samples. A key factor in the performance of continuous diffusion models stems from the guidance methods, which enhance the sample quality at the expense of diversity. In this paper, we extend these guidance methods to generalized guidance formulation for MGMs and propose a self-guidance sampling method, which leads to better generation quality. The proposed approach leverages an auxiliary task for semantic smoothing in vector-quantized token space, analogous to the Gaussian blur in continuous pixel space. Equipped with the parameter-efficient fine-tuning method and high-temperature sampling, MGMs with the proposed self-guidance achieve a superior quality-diversity trade-off, outperforming existing sampling methods in MGMs with more efficient training and sampling costs. Extensive experiments with the various sampling hyperparameters confirm the effectiveness of the proposed self-guidance.

Paper

References (64)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer cCTZ7/10 · confidence 3/52024-07-01

Summary

This work proposes a self-guidance method for masked generative models to improve the diversity and quality of class conditional image generation. The main challenge is to design a semantically meaningful smoothing for the discrete VQ token space such that coarse scale information can be extracted. The authors propose an auxiliary task – error token correction to minimize fine-scale details and utilize TOAST for efficient model fine-tuning.

Strengths

- The research problem of introducing self-guidance in discrete space is well-motivated. The background section is well-written, and the discussion of related work is thorough. I find the analogy to guidance of continuous diffusion model very helpful. - The figures are well-made, such as the qualitative visualization is Fig.1. - The quantitative metrics from the experiments also demonstrate the effectiveness of the method. The ablation studies on different hyperparameters are informative.

Weaknesses

- From the qualitative visualizations alone (Fig.4), the proposed method does not seem have too different outputs compared with MaskGIT (the middle column). I wonder if the guidance benefit is stronger on other resolutions’ results, like 128 x 128 or 64 x 64. - Maybe some figures could be added to accompany section 3.2 to better illustrate the process of the auxiliary task.

Questions

- What is $\bf{m}_i$ in equation (7)? Could you provide more intuitions or visualizations of the claims in line 192- 195. Why does this auxiliary task naturally minimizes fine-scale details? - What is the range of the experimented temperature values in Fig. 3? From Fig.3, it seems that the proposed sampling method has very different performance based on the choice of temperature. The authors also mention that they choose different temperatures based on the resolution and sampling steps. I wonder why is the proposed method so sensitive to temperature? - Have the authors tested the model on the 128x128 resolution ImageNet benchmark? What does the FID/IS curve look like? - What happens with strongly guided samples? For instance, for classifier-free guidance in diffusion models, strongly guided samples exhibit saturated colors. I wonder if there is any artifact related to this proposed method when the choice of guidance scale or sampling temperature is high?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors already mention some directions for future work, such as demonstrating the guidance on text-conditioned image generation.

Reviewer Lc5P7/10 · confidence 3/52024-07-10

Summary

This paper explores the use of masked generative models for image generation. While these methods are typically efficient, they sometimes fall short in terms of quality compared to diffusion models. To reduce this gap, this paper proposes a self-guidance algorithm, which is further improved with semantic smoothing. These ideas are shown to be effective for improved image generation using MGMs. The method is compared to a wide variety of baselines on class-conditioned generation, on common metrics, and against a set of baselines.

Strengths

- This paper tackles an important problem in image generation, which is the use of Masked Generative Models for efficient image generations. This family of models have shown to be more efficient than diffusion models, but have so far struggled to catch-up with the quality they can achieve. This paper improves the quality of these models, to make them competitive in quality while preserving their benefits in terms of efficiency. - At the core fo the method lies a self-guidance method, which I believe is a sound and well motivated idea, which can be potentially included in generative models for other applications. - The ideas in the paper are simple in a good way, and they are not particularly tied to a specific application. I believe they could be generalizable to many other problems. - I believe the related work is well analysed, as far as I am familiar with previous work on this topic. - Evaluation is extensive, many different hyperparameters are evaluated with respect to their impact on generation quality.

Weaknesses

- The paper is at times hard to follow, lacking clarity in some sections (particularly 5). Images and figures are sometimes small and hard to see. - I do not believe implementation details are enough for reproducibility. Code will not be provided, which will make this paper hard to reproduce and it will reduce its potential impact and room for future comparisons. - Given that a particular benefit of the method of this paper is its computational efficiency in comparison to diffusion-based approaches, I believe a detailed analysis of these should be included. - I believe the paper would benefit from showing results on text-to-image generation.

Questions

- How would this method generalize to text-to-image methods or image-to-image translation problems? - How tied is this method to VQGAN? Could other encoders be used? If so, what would be their impact?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

4

Limitations

The paper is fine in this regard. It could, however, provide more details on how the limitations could be addressed in the future.

Reviewer qSSj6/10 · confidence 4/52024-07-12

Summary

The paper focuses on enhancing Masked Generative Models (MGMs) for image synthesis. The authors identified several reasons for the underperformance of MGMs: 1) lack of sequential dependencies, 2) multi-modality problem, 3) compounding decoding errors, 4) limitations of low-temperature sampling and 5) ineffective guidance techniques. To address these issues, the authors present novel techniques and methodologies to improve their performance. The main contributions of the paper are as follows: 1) Generalized Guidance Formulation. 2) Self-Guidance Sampling Method. 3) Parameter-Efficient Fine-Tuning.

Strengths

- Originality The paper presents a highly original approach by extending guidance methods from continuous diffusion models to MGMs. In particular, the use of self-guidance sampling method and an auxiliary task for semantic smoothing in the VQ token space are novel approach that address specific challenges in MGMs. - Quality The paper demonstrates a high level of technical quality. The authors have thoroughly explained why their methods can enhance performance and have supported their claims with extensive experiments and ablation studies on various variables. - Clarity The paper is clearly written and well-structured. The authors provide detailed context relative to prior work, facilitating a clear understanding of the background and contributions of the proposed approach. The detailed explanations of the equations and methodologies, accompanied by ample examples, significantly enhance the overall readability and accessibility of the paper. - Significance The contributions of the paper are significant for the advancement of the MGMs field. By addressing key challenges in MGMs, the proposed methods offer substantial improvements in the quality and diversity of generated samples.

Weaknesses

Particularly due to inadequate visual results, some aspects are difficult to comprehend. Section 3.3 asserts that TOAST is the most suitable model for the study's task, but lacks sufficient explanation on why TOAST is the only suitable model. It would be more convincing if the paper showed some visual results from using other models, which would provide sufficient justification. Explanations for the choice of hyperparameters are insufficient. While the paper demonstrates trade-off relationships in hyperparameter settings across some datasets in removal experiments, the sampling time steps do not exhibit such trade-offs. It is crucial to provide adequate explanations for why these settings differ from those maintained for the ImageNet 512×512 dataset.

Questions

Most hyperparameters in Figure 5 show a trade-off relationship between FID and IS scores. However, increasing the number of sampling time steps improves performance in both metrics. So, could you explain why sampling time steps were set to 12 and 18 in each experiment? Even though other methods are known to use hundreds of sampling time steps, 50 seems quite small in comparison.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors did not fully address the limitations of the paper but did mention the need for attention to AI ethics due to the rapid growth of generative models, touching on the societal impact. To improve the paper, we suggest the following: 1) Although ImageNet is a comprehensive dataset, it is insufficient to prove generalizability. Therefore, conducting experiments on other datasets used in existing MGMs, such as MSCOCO or Conceptual Captions, will make the paper more complete. 2) Include more visual results. To clarify the reasons for selecting specific models or hyperparameters used in the paper, it is much more effective to show them through visual results.

Reviewer Lc5P2024-08-11

Post rebuttal discussion

Dear reviewers and authors, After the authors rebuttal, my concerns regarding the method and its reproducibility, the visual results and other minor concerns, have been adequately addressed. I am therefore leaning towards increasing my score. Could the other reviewers provide feedback?

Reviewer cCTZ2024-08-12

thank you

Thank you to the authors for their detailed response and the valuable insights provided on the intuition behind the proposed method. Based on this additional information, I have decided to increase my rating to 7: Accept.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC