Hi,
First of all, thank you for your awesome work on the flexble tokenizer.
I have a question regarding some discrepancies I noticed in the experimental results.
Specifically, the gFID scores reported in Table 9 (Appendix E, page 29) appear to be better than those in Table 1 (page 9) of the main paper—for example, gFID of 1.71 in Table 9 versus 1.86 in Table 1 for 32 tokens.
Similar discrepancies seem to appear between Table 9 and Figures 7, 9, and 24 as well.
Could you clarify the reason for this difference?
I suspect it may be due to different inference parameters (e.g., the number of denoising steps in the decoder), as I did not find any mention of whether the optimal inference settings described in Appendix G were applied to the results in Table 9.
Any clarification would be greatly appreciated.
Thank you!
Hi,
First of all, thank you for your awesome work on the flexble tokenizer.
I have a question regarding some discrepancies I noticed in the experimental results.
Specifically, the gFID scores reported in Table 9 (Appendix E, page 29) appear to be better than those in Table 1 (page 9) of the main paper—for example, gFID of 1.71 in Table 9 versus 1.86 in Table 1 for 32 tokens.
Similar discrepancies seem to appear between Table 9 and Figures 7, 9, and 24 as well.
Could you clarify the reason for this difference?
I suspect it may be due to different inference parameters (e.g., the number of denoising steps in the decoder), as I did not find any mention of whether the optimal inference settings described in Appendix G were applied to the results in Table 9.
Any clarification would be greatly appreciated.
Thank you!