[GH-ISSUE #24] Potential Bug Report: "do_normalize" setting in image preprocessor config #17

Closed
opened 2026-10-26 21:10:06 +00:00 by bsakaguchi · 1 comment
Owner

Originally created by @Junction4Nako on GitHub (Oct 30, 2024).
Original GitHub issue: https://github.com/deepseek-ai/Janus/issues/24

Thanks for the great work!

I find that the current setting for "do_normalize" in the image preprocessor config is set to False, which does not align with the original setting in SigLIP (https://huggingface.co/google/siglip-large-patch16-384/blob/main/preprocessor_config.json)

I conduct a simple test by inputting a red image into Janus with two settings:

  1. when do_normalize=False, the output is "pink" which aligns with the image without normalize (by converting the un-normalized feature map back to the PIL image)
  2. when do_normalize=True, the output is "red", which is correct.

Therefore, I think there might be a bug in the current config.

Could you please check it? Thank you and look forward to your reply.

Originally created by @Junction4Nako on GitHub (Oct 30, 2024). Original GitHub issue: https://github.com/deepseek-ai/Janus/issues/24 Thanks for the great work! I find that the current setting for "do_normalize" in the [image preprocessor config](https://huggingface.co/deepseek-ai/Janus-1.3B/blob/main/preprocessor_config.json) is set to False, which does not align with the original setting in SigLIP (https://huggingface.co/google/siglip-large-patch16-384/blob/main/preprocessor_config.json) I conduct a simple test by inputting a red image into Janus with two settings: 1. when do_normalize=False, the output is "pink" which aligns with the image without normalize (by converting the un-normalized feature map back to the PIL image) 2. when do_normalize=True, the output is "red", which is correct. Therefore, I think there might be a bug in the current config. Could you please check it? Thank you and look forward to your reply.
Author
Owner

@charlesCXK commented on GitHub (Nov 14, 2024):

@Junction4Nako
Sorry for the late reply. We have reviewed the corresponding files and found that do_normalize should indeed be set to true. We've already updated the relevant files (huggingface.co/deepseek-ai/Janus-1.3B@98b20d076c). Thank you for your kind reminder!

<!-- gh-comment-id:2477160308 --> @charlesCXK commented on GitHub (Nov 14, 2024): @Junction4Nako Sorry for the late reply. We have reviewed the corresponding files and found that `do_normalize` should indeed be set to true. We've already updated the relevant files (https://huggingface.co/deepseek-ai/Janus-1.3B/commit/98b20d076c1fd1e12a33e802b7a2eee88ec6364a). Thank you for your kind reminder!
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/Janus#17
No description provided.