Question about underlay effectiveness of Layoutprompter

[theKinsley]

Thank you for such a great work. Recently, I delved into the paper and the code provided for the content-aware layout generation task, and it appears that Layoutprompter handles the underlay element in a manner quite consistent with the treatment of other elements. Furthermore, within the Ranker module, the overlap of the underlay with other elements is actually subject to a penalty. This raises some surprise regarding the high Underlay effectiveness reported in the paper.

Considering that GPT-3 text-davinci-003 is no longer supported, I have chosen to use Llama 3 as the LLM component for running Layoutprompter. The outcomes revealed a notably low underlay effectiveness, particularly the strict effectiveness, which is nearly 0. I am wondering if there might be specialized Content-aware Ranker code that has not been released to the community. If such code is available, I would be very keen to learn about it.

Your insights on this matter would be greatly appreciated.

[kaoshizhou]

Thanks for your interest in our work. Currently, we use unified Ranker module for different layout generation tasks. Actually, we have used the language command to guide LLM generating underlay elements as the background of other elements. And we empirically found that GPT-3 text-davinci-003 could well comply this.

Given your experimental results on LLaMa3, we speculate that this might be due to the capability gap of different LLMs (i.e., GPT-3 could be better at such task). And we highly recommend that you could try other available LLMs to reproduce the results.

Finally, I personally agree that developing tailored Ranker module for content-aware layout generation could be useful and worth a try (e.g., encourage high overlap between underlay and other elements). And it promises to improve the two Und metrics.

Hope this helps you and feel free to ask if you have other questions :)