To reverse engineer the model, you need to download it to: runtime/models/MiniCPM-V-4_5
https://modelscope.ai/models/OpenBMB/MiniCPM-V-4_5/files
You need to extract caption_python.7z to the runtime directory. This is the Python environment. Due to the large number of subfiles, it can only be uploaded as a compressed package. If you don't want to download it or feel it's risky, you can download Codex and have it download a new environment for you.
Detailed tutorial:
https://youtu.be/h27Sedb_v08
Features:
-
Edits videos to the frame rate/resolution needed for training.
-
Includes cropping functionality to remove unwanted subtitles/black borders.
-
Offers faster frame range selection.
-
Records timeline, cropping box, and cue words for each data point, allowing for easy secondary editing without the need for manual adjustments and proofreading required by traditional editing tools.
-
Includes a cue word derivation function. Requires 16GB of VRAM for local operation. Low VRAM mode can be enabled in the settings if VRAM is insufficient.
-
English language can be enabled in the settings.
-
Supports batch conversion of frame rate/resolution for existing datasets.