Question about reproducing Janus Pro evaluation results (request for eval config / code)

[tls0523]

Hello, thank you for releasing this benchmark and the leaderboard — it is very helpful for the community.
I am currently trying to reproduce the Janus Pro results reported on the leaderboard. However, I am running into difficulties reproducing the evaluation scores and generation quality.
Specifically:
I am using the official Janus Pro inference code released by the authors.
With this setup, I am unable to reproduce the reported leaderboard metrics.
The generated images also appear noticeably worse than the examples or results implied by the leaderboard scores.
Because of this, I suspect there may be additional evaluation code, inference configuration, or generation settings (e.g., sampling strategy, guidance scale, number of steps, resolution, post-processing, etc.) that are not fully specified in the public inference scripts.
I would like to ask:
Is the Janus Pro evaluation code used for this benchmark available?
If not, would it be possible to share the exact inference and evaluation configuration used to produce the leaderboard results?

[A113N-W3I]

Hello, I’ve located the inference code that was used with JanusPro at that time.

https://ai.feishu.cn/docx/ODM5dhl2nocWKMxe9NxchqINnLf?from=from_copylink

Based on my inspection, it seems that no particular or special settings were used, but I’d appreciate your confirmation.