Thank you. We hope our responses adequately address the concerns regarding the novelty of our work.
Before the advent of diffusion models and the widespread use of transformers, structural data generation mainly relied on CNN-based models such as VAE and GAN. However, these methods often operated on latent spaces, potentially leading to the loss of the data's structural features. In early diffusion image model development, U-Net was commonly used as the denoising backbone instead of Transformers, requiring the mapping of data to a continuous latent space or the application of techniques to manage aspects of discrete data. Following the introduction of Transformer-based diffusion (DiT), a series of studies have been inspired by it. However, they still adhere to DiT's original design, mapping inputs into a continuous latent space. We propose a transformer-based diffusion model dedicated to general layout generation. This model operates directly on the space of geometric items, improving explainability and facilitating easier adaptation by other models.
**Q1: The core proposition of the paper, which involves the utilization of the input coordinate space for layout design generation through continuous diffusion models, is not entirely novel. Similar approaches have been discussed in prior works such as [1, 2].**
LEGO [1] mainly focuses on using additional constraints to rearrange the vectorized input data instead of generating vectorized data from scratch. HouseDiffusion [2] uses graph-based layouts instead of bounding boxes, diverging from our method. Besides, it deals with layouts in separate (continuous and discrete) sections, potentially impacting overall structural coherence.
**Q2: Experiments and Comparison: The experiments presented currently lack comprehensiveness. The results in Table 1 do not facilitate a fair comparison between the proposed method and existing methods. Although Tables 2 and 3 provide more data points, they restrict their focus to MaxIoU and Alignment scores. Furthermore, the results suggest that the proposed method underperforms compared to the baselines. This underscores the need for a more in-depth analysis and comparison.**
For a fair comparison, we use the data from papers that provide a clear methodology for computing metrics. Play and UniLayout are not open-sourced. The authors of Play responded to us with details. Because implementation details lead to significantly different results, for example, adopting Play's approach using 1024 samples to compute scores results in an FID score of 1.73 for conditional generation (conditioned on category), which is better than the number presented in the LayoutDM paper. However, using only 512 samples, the score becomes 4.21, which is worse than LayoutDM. While our model underperforms the baselines in some circumstances, our method generally performs better overall.
**Q3: Considering the large batch sizes which are used in the experiments (10k and 6k) compared to the conventional batch sizes up to 2048, adding a table on the effect of the batch size on the final result can be very insightful.**
We have tried to train the model using batch sizes of 10k, 5k, 3k, 1k. The outcomes indicate that all configurations yield similar FID scores within the range of ± 0.2, which can be attributed to the randomness of the diffusion sampling process.
**Q4: I couldn't find any direct comparison on pros and cons of the Dolfin and Dolfin-AR. It is better if you also add both versions to other tables as well.**
For unconditional generation, we provide results for Doldin and Dolfin-AR. Dolfin outperforms in metrics like the FID score, while Dolfin-AR excels in the alignment score. Regarding conditional generation, since Dolfin-AR does not support it, we only present results for Dolfin.