Video + Screenshot pipelines

[TomLucidor]

Can this be used to extract character data (for whole body, including backside clothing)? If it can match things, it should be easy to match OOD data (models that are not explicitly trained on certain characters) by showing 5-50 examples and see if there are similar items in video frame for extraction purposes.

Borrowed ideas from the following as example

• https://github.com/cyber-meow/anime_screenshot_pipeline
• https://github.com/deepghs/cyberharem

The core issue would be having AniSeg to handle the backside of characters and not necessary having the full context, but if a user can tag certain backsides to as corresponding to specific characters, then the automation pipeline to use that in full capacity to automate the rest (the more user tags the data, the more information is generated for the pipeline to discern between characters AND identifying "not a character")

[NON906]

In this repository, images are converted into vectors using "clip-ViT-B-32", and the distance between the vectors is used for sort and filtering.
For tags, I use "wd14-tagger-standalone".
It depends on whether they are similar and if they can be tagged.