Pony/ILLUSTRIOUS dictionary! - full tag catalogue
詳細
ファイルをダウンロード
このバージョンについて
モデル説明
This is a huge project and a work in progress. I am publishing this as is because I think the tool I am makin can be useful to the community even in its current state, and I intend to update it as I move forward.
PONY|ILLUSTRIOUS FULL TAG CATALOGUE
What is this thing here? This is a parse of all tags from Danbooru with encounter frequency over 500 (i.e. there were more than 500 images with any given tag of Danbooru at the time of the parse). The aim of the project is to sort, test and prune all the tags into individual thematic lists.
This servesthe folowing purposes:
Avoid nonexistant tags in prompts. Most people try to guess tags when generating, resulting in messy prompts and tags in there that don't work. Pony-based models understand only Danboory tags, and ILLUSTRIOUS requires tags to be input EXACTLY as they are on booru, or they won't work most of the time. With the catalogue here you can check the correct spelling of tags and search if a particular tag you're thinking of exists. This will help you avoid inputting dead tags into prompt.
Check your options and find inspiration. There are thousands of tags, and while you can check them on Danbooru itself, they are sorted in weird categories and impossible to check at a glance. With the sorted lists here you can see in one place all the options you have for a particular aspect of the image you're generating that a Pony-based model can understand. Open "artstyle.txt" and see all the artstyles you can choose from on one page. Open "pose.txt" and see all the poses Pony knows. Essentially this is a dictionary of words you can use to communicate with a model, structured to allow you to look up the aspects of the picture you need.
Randomise certain aspects of your image. These files are Wildcards, so you can yse them as such. Want to experiment with the same scene in different locations? Use the location wildcard. Want to try different outfits for the main character? use the clothes wildcard etc. Take one of each and get a competely random picture :)))
Okay, with the general purpose explained, here's what's inside exactly and how it works.
This is a work in progress. As of today all tags with Danbooru frequency over 5 100 have been processed and sorted into the wildcard files. Current progress is 2 602 out of 14 328 tags processed.
The file alltags.txt is ALL the danbooru tags with frequency over 500. All the tags that have a comma and number after them are yet unprocessed. All the tags that have nothing after them have been processed and feature in category lists.
What I mean by "processed":
Tag has been tested by actually generating an image with it. If it's in the list, it works for sure.
Tag has been assigned to one or more categories and included in one or more of the category wildcards. Some tags might seem like they are one thing, while in fact they're another. Finding how each tag works is one of the goals of this project. For example: you might think "armchair" is an item that you could put into an image, in fact it is not - actually tagging armchair creates a scene where armchair takes up most of the image and the main character is in the armchair. Therefore "armchair" tag has been classed as "location" not as "item".
Tag has been reinforced/supported when needed. Sometimes (rarely) tags only generate the thing they're supposed to represent only when used with other tags or require a supporting tag. For example "ffm threesome" tag is very inconsistent or doesn't work at all if used by itself. It does work as intended if used like "ffm threesome, 1boy, 2girls, hetero".
In the alltags.txt some tags have been pruned: tags that dont work, metatags that are meaningless for generation (like "wrong artist" or "requires comments" or whatever), individual character names, individual anime series names, individual games/franchise names. Franchise names I actually know have been moved to franchise.txt, the ones I dont know - pruned completely.
Right, so this is it for the generalities. Now for what's actually inside:
alltags.txt - all the booru tags, including the unprocessed ones with frequency over 500 minus the pruned tags in the processed part.
accessory.txt - small things that appear on the person like jewelry and such.
anatomy.txt - tags that indicate a certain piece of anatomy should be present in the picture and also modificatio of that anatomy (small breasts, large breasts etc.)
armor.txt - pieces of armor
artstyle.txt - tags that impact the overall style of the image
background.txt - tags that create solid color or simple backgrounds
bodymod.txt - tags that modify the subject body, but still keep it human and are at smaller/less impactful on the overall composition compared to "anatomy.txt". Things like color of nails, tattoos, teeth go here. Except eyes, which have their own list.
clothes.txt - currently a bit of a mess because how many clothes tags there are. Right now includes all kinds of clothes, except legewear and underwear/sexy clothing, which have their own lists. There are A LOT of clothes tags, so in the final release the list will be subdivided by type of clothing.
clothesmod.txt - tags that modify clothing somehow. If it can apply to different clothing it's here (like "frills"). If it forces a specific type of clothing to appear, it's in the clothing.txt
color.txt - just colors. nothing fancy.
composition.txt - tags that govern image composition, without definig the main subject. Thinhs like "lineup", "full body", "magazine cover" go here. These tags are typically not mutually exclusive unlike most other lists.
expression.txt - facial expressions. Super-useful (at least for me) because guessing the name of an expression is super-tough. Now you can just look them up.
eyes.txt - tags that define the eyes. Color and some weird anime stuff, but eye-related tags that forced an obviously non-human subject are not here, but in nonhuman.txt
eyewear.txt - separated from accessories because I love girls in glasses :))
fashion.txt - tags that define a certain stye in attire without specifically stating that attire. A small list so far, but I expect a rework of the clothes-related categories before final release.
footwear.txt - yeah, footwear.
franchise.txt - franchises, series or fictional universes. These tags usually don't generate anything specific but can impact style, character looks and such.
gazedirection.txt - a small list that helps manage where the character is looking. These tags are very dependant on the rest of the prompt, so the list will probably be reworked in the future.
gesture.txt - gestures of the subject. Small hand gestures or minor poses that impact only hands and arms. Tags that involve overall body position go into pose.txt
haircolor.txt - yeah, hair color. Including all the anime weirdness.
hairstyle.txt - hairstyles. For the life of me I can't remember all these names, so this is actually the one I use most as a wildcard :))
headwear.txt - items that go on the head and are not accessories. Things like "scrunchie" won't be here, things like "fedora" will. Some things are ambiguous (like "circlet"), I will sort them out later.
improve.txt - tags I tested and am 100% sure impact picture quality. Doesn't mean you're advised to use all of them not the list is exclusive (because ILL and Pony have their own quality rating systems).
item.txt - stuff, things, items. These are thingies that will appear in the image, but will not take the bulk of space nor impact main subject. This includes things that will tend to appear in the subject's hands, even though tags that actually specify the subject is holding an item are in "situations.txt". May be reworked in the future.
legwear.txt - things that the subject wears on legs. footwear has its own list.
lighting.txt - this is supposed to be tags that regulate the lighting of the picture, but it's currently a mess as some tags force a location, some force items etc. Will definitely be reworked in the future, but stays while I'm focusing on processing new tags.
location.txt - tags that define the location of the scene. By location, I mean what's generally around and behind the subject. So sometimes items or situations will be classed as locations if they dominate the backgrond. "armchair" is one of the examples of this.
male.txt - a small list of tags that will force a male subject. Currently a stump, will be reworked.
misc_unknown.txt - tags that I failed to classify for now because they don't fit into my current structure. As work progresses and structure changes, thes will be assigned somwhere. Hopefully.
nonhuman.txt - tags that will prompt a subject that is obviously not human. This includes furry as well as actual animals, monsters and parts thereof.
nudity.txt - tags that will generate some form of nudity. Partially cross over with other lists, so will need a cleanup or reclassification in the future. Also not exclusive - I didn't really put all the tags that generate nudity there, mostly because Pony models generate nudity even if unprompted, so this is kind of redundant for now.
pattern.txt - color and geometric patterns. These will appear on clothing, but sometimes in other places - floors, walls or some other surface.
pose.txt - tags that define whole-body poses. The list includes sexual positions as well as individul subject poses. Maybe I'll divide the list in the future.
profession.txt - tags that mostly impact clothing, but also general content of the image (like location or activity). Also tags that are too general for clothing and do not specify a particular piece of clothing, but rather a certain type, like uniforms of working outfits.
rating.txt - ILL content rating. Just the four tags there and specific to ILL.
scene.txt - tags that define what's on the image, without actually defining the subject. Like 1girl - the girl could be human, demon, furry, futa, whatever: the "1girl" tag will be in "scene" while the tags that define the subject will be in "subject".
sex.txt - non-exclusive list of tags that will produce sex. Some tags that produce sex are also in "pose" and "situation" so this needs to be reworked.
situation.txt - tags that define action, what's happening in the image. This is a rather broad category, so this list will be split in the future.
specialeffect.txt - tags that produce some kind of extra artistic effect, that is not inherent in the image (or sometimes I jsut didn't know where else to put :)). Speech bubbles, borders, motion lines etc.
subject.txt - tags that define the subject of the image. Only tags that keep the subject human are supposed to go here, but currenty this list is mixed-up with nonhuman (because I introduced that structural division after I processed quite a few tags). So this will be re-sorted before release.
undeerwear.txt - underwear as well as sexy clothing goes here. Includes swimsuits and bikini for now.
viewangle.txt - camera positions and angles.
Whew. That's a long list isn't it :)))
Well, I hope you find this resource useful. I will be regularly updating the resource as I process more and more tags and improve the structure. Right now this is very much a Work in Progress, so please don't junde too harshly - I'm changing the structure on the fly, so things might get mixed up.
Please post comments, suggestions and whatever, any feedback is much appreciated and is a huge motivation for me to keep going with this mammoth of a project.