AIBooru

Discussion for + Proposing solutions for has:metadata modeltags:0

Posted under General

I have been at this for a few days now, reducing the amount of these posts by about two pages... only for it to slowly be filled up again by bad tagging jobs and sometimes poor metadata.

So really from what I've heard in #aibooru on Discord from Lyren, iodoff had a bot that would automate part of this job by using the model hashes of metadata to find models (which is what iodoff's script still does), but this bot clearly hasn't been in service for a bit. Perhaps there were issues with this bot, but clearly all it leads to is for people like me who care about this problem to be forced to manually look at post metadata and figure out what models are used.

What I've noticed is always one of these things, ranked in order of what I believe to be easiest to automate tagging for:
1. Model hash leads to a CivitAI page or there is a valid HuggingFace page with the model in question, leads to a model that already has a tag, thus leading me to presume bad tagging job
2. Model hash leads CivitAI page or there is a valid HuggingFace page with the model in question, but the model in question doesn't have already have a tag
3. Model hash doesn't tell me anything, but the model name suggests it is a merge of two models (I tag both models I can see in the cases of these posts)
4. Model hash doesn't tell me anything, can't find the model anywhere, leading me to presume some kind of private model (I have created the sosos tag for instance)
5. Bad metadata, see post #61806 for an example of this

I believe it would be best to at least automate 1 and 2, and leave manual intervention open only in cases of 3, 4, and 5. I would like this thread to simply discuss methods as to how to actually automate those two cases specifically, but to also see if this could be a discussion thread for posts that fall under has:metadata modeltags:0 as well.

Thankless work. As someone who went back and did the same thing with the model_request tag, I'd appreciate it, too. Big issues are:

1. Huggingface and Civitai have really awful options for searching by hash (or in the case of HF, searching at all). Shorthash in early models also makes the problems worse. Would be a good idea to compile a master list of hashes + models.

2. Metadata parsing has a lot that could be improved but the nature of SD as a developing tech, lack of standardization /w metadata (comfyui metadata is extremely cancerous to read through) really make it vary on a case-by-case basis which makes it extremely hard for implementing automated metadata tags, because there's going to be a lot of edge cases.

But an automated bot would cut down a lot on easier ones, yeah.

Lopi999:

4. Model hash doesn't tell me anything, can't find the model anywhere, leading me to presume some kind of private model (I have created the sosos tag for instance)
5. Bad metadata, see post #61806 for an example of this

Point 4, You can also tag undefined if a post is applicable.
Point 5, is there a tag for bad metadata already? If not, I can see one being useful for sifting things out like that.

BTW if you're still interested in this, one quick hack I found to help identify civitai resources is to just upload the image if it has metadata. Civitai has awful tools for directly searching for model hashes but their resource identification system automatically does it if it's in the metadata.

Dogshit website design strikes again.

1