I'm looking for the least biased for a large training set on water parks and water slides where there is a lot of content models struggle to deal with.
No, not in some creepy way. I actually want to eventually try training a LoRA on O'Neill cylinders but I don't have the experience, and have a limited dataset. Water slides are a similar set of issues with model internal thinking and comprehension. Presently, all images of water slides start off as stairs, models do not understand gravity as a propulsion method and water as a lubricant on a slide. The water falling is literally a waterfall and falling/slipping/sliding are associated.
I've made several trained attempts at LoRAs already with limited success and around 200 images but there are too many mixed concepts and difficulties.
Now I'm trying to make a much larger training set. I have a well tuned Qwen 2.5 VL 7B doing captions, but want to compare the results with tags training in various models. I need something that is not going to freak out because a normal-clothed not-at-all-sexualized kid is in some images or especially ignore someone in the middle of the air on a launch slide or making a thrill-ride facial expression.
My dataset is a bunch of YT type videos and I am just using FFMPG to do periodic screenshots. The videos are not at all NSFW and most are years old, so they pass moderation. I only focused on getting as many images as possible that include sections of people sliding with the camera on their face. My goal is to capture the true cultural diversity and unique cultural norms of a water park similar to how models know beaches have different cultural norms with internal thinking and alignment.
So, do you all know what is the most recent and reliable tags generating model where alignment does not get in the way?