AIBooru

tag clarifications

Posted under Tags

I noticed some of my posts were tagged with animefull_(model) at first I started removing it because I assumed it was a mistake because I used novelai in them and the model name I had in the commentary was nai-aminefull-final-pruned. Now that I think about it novelai has multiple models so maybe noting that mine were done with the anime one makes sense? Which tag should be used?

Another thing is the group of tags upscaled, resized, img2img, inpainted and first-party_edit I understand that some of them come from danbooru but I never knew which ones should I use with my posts. I rarely post images straight out of stable diffusion but my changes are usually very small (mostly just fixing fingers and color correction). Does this qualify as inpainting or first party edit? Sometimes I do upscaling and pass the image to img2img without changing the prompt, get a bunch of results and then combine them into one image. So I guess I should tag the with img2img and what else? upscaled? inpainted? first party edit?

antlers_anon said:

I noticed some of my posts were tagged with animefull_(model) at first I started removing it because I assumed it was a mistake because I used novelai in them and the model name I had in the commentary was nai-aminefull-final-pruned. Now that I think about it novelai has multiple models so maybe noting that mine were done with the anime one makes sense? Which tag should be used?

We don't have variations of models, such as all the waifu diffusion versions is just simply waifu diffusion without the version number at the name. So maybe animefull (model) should just be nuked.

antlers_anon said:

Another thing is the group of tags upscaled, resized, img2img, inpainted and first-party_edit I understand that some of them come from danbooru but I never knew which ones should I use with my posts. I rarely post images straight out of stable diffusion but my changes are usually very small (mostly just fixing fingers and color correction). Does this qualify as inpainting or first party edit? Sometimes I do upscaling and pass the image to img2img without changing the prompt, get a bunch of results and then combine them into one image. So I guess I should tag the with img2img and what else? upscaled? inpainted? first party edit?

Any method should be tagged, even if the edit is small, that's at least how I've proceeded with the tagging. There hasn't been a discussion about this ever though, which indeed makes it very vague how the tagging with these tags is supposed to be done and maybe I've been doing it wrong. They really need proper wikis, such as upscaled straight out says it should be avoided, which conflicts with the upload rules' statement "Upscaled images are allowed, only if using latest AI." /// Any extra thoughts on when these tags are supposed to be used are appreciated.

Lyren said:

We don't have variations of models, such as all the waifu diffusion versions is just simply waifu diffusion without the version number at the name. So maybe animefull (model) should just be nuked.

They could also be aliased too, if desired. I feel hesitant about either idea though, because while having multiple model tags on a post isn't ideal, it's done that way because merged model is a thing. Also, given how many specific models exist across the place, I don't know if it's a good idea to simplify them to the main model they're a variation of.

Lyren said:

Any method should be tagged, even if the edit is small, that's at least how I've proceeded with the tagging. There hasn't been a discussion about this ever though, which indeed makes it very vague how the tagging with these tags is supposed to be done and maybe I've been doing it wrong. They really need proper wikis, such as upscaled straight out says it should be avoided, which conflicts with the upload rules' statement "Upscaled images are allowed, only if using latest AI." /// Any extra thoughts on when these tags are supposed to be used are appreciated.

I agree the wikis need updating badly. There was a point, I believe where upscales were disallowed entirely, but they became okay to upload, only if they were upscaled using something that wasn't quite old (like Waifu2x as I had been given as an example.)

I was always under the assumption that if any edits were made, they should be tagged too but the only one I'm not 100% sure about is the inpainted tag or what qualifies a post to be tagged with it.

Edit: I updated the img2img, inpainted, and upscaled, wikis. If any changes are needed, then I or someone else can get them in, but I added some basic info for the first two, and added the same clarification in the upload rules to the last one.

Updated

ANJU said:

They could also be aliased too, if desired. I feel hesitant about either idea though, because while having multiple model tags on a post isn't ideal, it's done that way because merged model is a thing. Also, given how many specific models exist across the place, I don't know if it's a good idea to simplify them to the main model they're a variation of.

Hmm, if we decide to keep variations of models, implication would make more sense than alias. The biggest problem with this I can see is for non-first-party uploaders tagging, something like #novelai doesn't really tell which variation was used. I don't know if model hash helps with identifying which one was used, I don't know much about model hashes myself.

ANJU said:

Edit: I updated the img2img, inpainted, and upscaled, wikis. If any changes are needed, then I or someone else can get them in, but I added some basic info for the first two, and added the same clarification in the upload rules to the last one.

In inpainted wiki you mention that inpainted images should probably have first-party edit as well. I always thought first-party edit is only for manual handwork like Photoshop editing.

Lyren said:

Hmm, if we decide to keep variations of models, implication would make more sense than alias. The biggest problem with this I can see is for non-first-party uploaders tagging, something like #novelai doesn't really tell which variation was used. I don't know if model hash helps with identifying which one was used, I don't know much about model hashes myself.

I suppose that would make more sense. If everything was aliased then we'd have to go back to the days of making comments to add info. I don't know about model hashes either, so.. I can't help there.

Lyren said:

In inpainted wiki you mention that inpainted images should probably have first-party edit as well. I always thought first-party edit is only for manual handwork like Photoshop editing.

Yeah, I had that messed up in my head. I took that line out.

Lyren said:

Hmm, if we decide to keep variations of models, implication would make more sense than alias. The biggest problem with this I can see is for non-first-party uploaders tagging, something like #novelai doesn't really tell which variation was used. I don't know if model hash helps with identifying which one was used, I don't know much about model hashes myself

Hashes don't always fully identify a model but this mostly applies to merged models. I have a few mixes that are vastly different but share the same hash. Afaik each novel AI variation has a different hash but I don't use it anymore so I don't know which one is which.

ANJU said:

Edit: I updated the img2img, inpainted, and upscaled, wikis. If any changes are needed, then I or someone else can get them in, but I added some basic info for the first two, and added the same clarification in the upload rules to the last one.

Thanks, it's much more clear now. I'd add that the img2img tag should not be used when the process was used only for inpainting or upscaling but I'm not sure if that's even the way we want it.
In case of first party edits, does this apply to fixing mistakes and adding elements using inpainting? I think it should be used only for cases where some clearly non-AI elements were added, like text or stitching multiple images to make a comic panels.

antlers_anon said:

Thanks, it's much more clear now. I'd add that the img2img tag should not be used when the process was used only for inpainting or upscaling but I'm not sure if that's even the way we want it.

In case of first party edits, does this apply to fixing mistakes and adding elements using inpainting? I think it should be used only for cases where some clearly non-AI elements were added, like text or stitching multiple images to make a comic panels.

Img2img honestly should be its own thing. It's a generation method combining the use of a "base" image and using prompts to modify that base, in whatever desired ways. It's not always using something like post #251, and you can do small fixes/edits, but it's still using a base to help making another image.

I wouldn't consider it a first-party edit either, because I think that tag should be reserved for outside changes made. Anything from adding text/dialogue, to adding a story, stitching images together, making a comic/manga, or removing/changing/adding elements from the image in general would be fair game to call it that, for me.

I fully admit that I don't know where the inpainted tag fits into this. I looked up some info and added whatever basic info I could extrapolate from what I did find, so I know it could be better.

1