dualreason 
Has anyone ever used AI/ML to predict the score of an image based on its tags (on Derpibooru or other image-hosting sites)?
Context:
I’ve pursued more than a few commissions over the years and it has me wondering how much of a score is derived from an artist’s skill/style and how much is predetermined by the content (ex. an animation is bound to get more favs than a static counterpart, main characters score better than OCs, etc).
I’ve pursued more than a few commissions over the years and it has me wondering how much of a score is derived from an artist’s skill/style and how much is predetermined by the content (ex. an animation is bound to get more favs than a static counterpart, main characters score better than OCs, etc).
I’ve run some high-level stats and can see some overall trends in the tags (ex. NSFW scores about 4x better on average than SFW, to put some numbers on it), though I find myself running smaller and more complicated sub-group analyses trying to estimate how well a certain set of tags I plan to use on future commissions will rate (while acknowledging the unique set of tags I plan to use has no exact one-to-one precedent). It takes a lot of effort and compute time to put this together and at this stage I’d rather just hand off the data to a machine learning model to do all that sub-group consideration and extrapolation/correlation automatically (I’m using the Derpibooru tag+scoring daily data dump as an input, many GB to iterate over).
So my question is: has anyone done this before? If there’s already a published model or approach for estimating an artwork’s score based on the tags, I’d rather save some time and use publicly available resources/approaches/tools (I would like to avoid re-inventing the wheel).
My hope is that this could help level-set some future commissions, ex. acknowledging that an artwork with more characters will be more work to make, but also that it’s likely to score 20% worse than a past commission because a majority are OCs. Or estimate a range of likelihood of performance (I have 80% confidence the score after 3 days will be within 200-300 favs). Or answer general questions like maybe 60% of a score can be predicted from the non-artist tags and 40% is driven just by which artist made it. Or estimate the effect of adding/removing small tags (ex. “floppy ears” or “fluffy”) will have on an image score (estimate how impactful/important details may be).
Note: I’m already familiar with high-level programming languages like Python and I have a introductory-level understanding of machine leaning (ex. a model that works on tagging data I expect will be structured and trained differently than Stable Diffusion for working with imagery), but haven’t set up any serious models myself from scratch yet.
Any inputs/thoughts/guidance on where to start with this project is deeply appreciated, thank you
