Gemma 3 and Wan2.1

[DCVirtualCosmos]

With the release of those powerful models, one could start dreaming of a tool to caption videos to train Wan LoRAs. And Gemma 3 seems the perfect tool. It's pretty smart, follow instruction quite well, there are versions of it uncensored already on hugging face, and it can analyze perfectly a sequence of images to describe what is happening in a short video.
So, it would be nice if Taggui could:

• Open .mp4 videos like the images, playing them in bucle when selected in the left selector.
• Allow to caption them like images
• Being able to use a .safetensor or .gguf file directly put into the /models folder for captions.
• A little menu to drag a selected number of frames from the video to be analyzed for the LLM.

I will try to do this myself when I got time, but perhaps you are faster!

[rudolphos]

Bump. Gemma 3n 2B and 4B is more efficient and optimized model that can run on lower-end hardware as well. Currently not yet added to taggui. It's available via transformers library - https://huggingface.co/blog/gemma3n