Question about codec-style patch selection implementation

[Ray-ui]

Hi authors, thank you for the great work on OneVision-Encoder.

I have a question about the “codec-style patch selection” mentioned in the paper/website.
From the codebase, I see that the model forward supports sparse inputs via patch_positions,
but I could not find an explicit implementation of the codec-style patch selection pipeline
(e.g., how patches are selected from raw videos).

Could you please clarify:

1. Is the codec-style patch selection implemented anywhere in this repo, or are users expected
   to provide patch_positions externally?
2. If it is external, do you plan to release the preprocessing/selection code (or a reference implementation)?
3. What signals are used to estimate patch importance during training (e.g., motion, residuals, saliency)?

I am trying to better understand and possibly reproduce the codec-style pipeline.
Thanks for any clarification!

[anxiangsir]

We've recently improved the documentation for this part and have implemented codec-style patch selection. We've also released the preprocessing/selection code here:

https://github.com/EvolvingLMMs-Lab/OneVision-Encoder/tree/main/llava_next#training-data-conversion-pipeline

If you have any further questions, @YunyaoYan will be happy to help!

[Luodian]

looks like @anxiangsir covered this - the codec patch selection pipeline is documented under llava_next/ and the preprocessing code (stage1 energy extraction + stage2 patch packing) is released. closing this out, reopen if you have follow-up questions.