Hi, thanks for releasing LLaVA-OneVision-2.
I’m trying to understand the codec-aligned video preprocessing described in the README / technical report. I checked the current main repo, and it seems the released code mainly follows a decoded-frame path: decoding videos, sampling RGB frames, patchifying them, and building patch_positions.
However, I couldn’t find the codec-specific preprocessing code, such as extracting motion/residual information, parsing I/P/B frames, constructing I/P canvases, or selecting sparse P-frame patches.
Could you please point me to the relevant code for the codec preprocessing pipeline? Is it in another branch, an external dependency, or not released yet?
Thanks!
Hi, thanks for releasing LLaVA-OneVision-2.
I’m trying to understand the codec-aligned video preprocessing described in the README / technical report. I checked the current main repo, and it seems the released code mainly follows a decoded-frame path: decoding videos, sampling RGB frames, patchifying them, and building patch_positions.
However, I couldn’t find the codec-specific preprocessing code, such as extracting motion/residual information, parsing I/P/B frames, constructing I/P canvases, or selecting sparse P-frame patches.
Could you please point me to the relevant code for the codec preprocessing pipeline? Is it in another branch, an external dependency, or not released yet?
Thanks!