[GH-ISSUE #17] About the preprocessor way in high and low input? #12

New issue

Closed

opened 2026-10-26 21:08:48 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 21:08:48 +00:00

Owner

Originally created by @lucasjinreal on GitHub (Mar 13, 2024).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-VL/issues/17

the SAM using different means compare to Siglip,

but from the code, the VLIChatProcessor are just using 1024 as size, imagenet_mean as input, this is right for SAM, but not right for siglip.

the processed tensors are directly sending to HybridTower without any further process.

Why doing this? Is the correct?

Originally created by @lucasjinreal on GitHub (Mar 13, 2024). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-VL/issues/17 the SAM using different means compare to Siglip, but from the code, the VLIChatProcessor are just using 1024 as size, imagenet_mean as input, this is right for SAM, but not right for siglip. the processed tensors are directly sending to HybridTower without any further process. Why doing this? Is the correct?

bsakaguchi closed this issue

2026-10-26 21:08:48 +00:00

bsakaguchi commented

2026-10-26 21:08:48 +00:00

Author

Owner

@Murasame-sway commented on GitHub (May 4, 2025):

What's your opinion on this issue? I've found that some other works using dual vision encoders still employ different preprocessing
@lucasjinreal

@Murasame-sway commented on GitHub (May 4, 2025): What's your opinion on this issue? I've found that some other works using dual vision encoders still employ different preprocessing @lucasjinreal

bsakaguchi referenced this issue

2026-10-26 21:09:17 +00:00

[PR #12] [MERGED] CLI: enables image drop-in from file explorer #62