search

jhc13 / taggui

Tag manager and captioner for image datasets
GNU General Public License v3.0
777 stars 37 forks source link

[Feature Request] Auto-tagging, models choosing from a list of tags #130

Closed Trevor-Z closed 6 months ago

Trevor-Z commented 7 months ago

Would it be feasible to implement a feature where we feed the models a list of tags to choose from in order to auto-tag images?

I can do it in more than one step, generate a detailed caption here, then go elsewhere and ask an LLM to choose from a list of tags the ones that apply to the contents of the caption. But it's quite cumbersome.

I can also simulate some of that functionality here, asking a model if the photo is indoors or outdoors, if a car is black or graphite, but I don't know how many questions I can ask and how many options I can offer in a single prompt and still get a coherent answer.

If this is within the scope of the project, maybe being able to set up macros, sequential prompts with step by step instructions, and having the option to save only the last answer, would get us most of the way there, and also greatly expand the functionality of the app.

eraser851 commented 7 months ago

Try experimenting with listing your tags in the 'Include in Caption' box. Separate|each|tag|with|a|pipe. It seems to pick and choose which ones to use.

jhc13 commented 7 months ago

Would it be feasible to implement a feature where we feed the models a list of tags to choose from in order to auto-tag images?

If this is within the scope of the project, maybe being able to set up macros, sequential prompts with step by step instructions, and having the option to save only the last answer, would get us most of the way there, and also greatly expand the functionality of the app.

These features could be technically possible to implement using structured/guided/constrained generation, and they have been requested and considered in the past, but they are not priority features right now.