IT TOOK A REALLY F****** LONG TIME, SO AT LEAST GIVE IT A LIKE, WRITE A COMMENT AND SAVE THE ARTICLE!
If you plan to write a paper based on this article or use its data, please be sure to mention the author and leave a link to this publication.
PREFACE
Let's start with the fact that the NoobAI-XL model is based on another model, and that one, in turn, is based on another model, and that one on another... well, you get the idea. Let's take a look at the hopefully complete list of models that make up NoobAI-XL.
NoobAI-XL ⬅️ Illustrious-XL ⬅️ Kohaku-XL beta ⬅️ NekoRayXL ⬅️ CounterfeitXL + AIO-Anime + SDXL 0.9 ⬅️ SDXL 1.0 (where '⬅️' denotes 'based on')
As we can see, the list is quite impressive, and there's a lot mixed in there, they EVEN MANAGED to mix in SDXL 0.9. Based on recent research, we already know that it's not the quantity of data that matters, but its quality, so I would say this is not very good. What follows from this? Build your model based on the base one. But it is what it is...
I have sufficiently researched both the models described above and their datasets, at least those that are available. And I can say that I have compiled from all the data an almost mega, judging by its size, dataset of negative tags. By the way I reviewed all the booru used for the right tags.
EACH of these tags has a basis, so complaints about the quantity will not be accepted.
Believe me, I've seen enough 500+ MB json files ⬇️
TAGS! TAGS! TAGS!
For your convenience, I've categorized the entire list using AI:
1. Art Generation and Quality
Art Generation Method: ai-generated, ai-assisted, stable diffusion, nai diffusion
General Quality Assessments: worst quality, worst aesthetic, bad quality, normal quality, average quality, low quality
Subjective Quality Descriptors: very displeasing, displeasing, ugly, worst
Technical Quality Issues: lowres, jpeg artifacts, compression artifacts, blurry, artistic error, bad proportions, bad perspective, aliasing, jaggy lines, scan artifacts
2. Art Style and Medium
Style Descriptors: monochrome, sketch, concept art, flat color, flat colors, simple shading, abstract
Medium/Tools: traditional media \(artwork\), microsoft paint \(artwork\), ms paint \(medium\)
Artistic Composition: simple background, asymmetrical
3. Image Content and Composition
Anatomical and Structural Issues: bad anatomy, bad hands, bad feet, disfigured, deformed, extra digits, fewer digits, missing fingers
Censorship: censored, bar censor, mosaic censoring
Completeness: unfinished, missing, extra, fewer, bad, hyper, error
Framing and Borders: out of frame, cropped, letterboxed, framed, border, panel skew
Multiple Images/Views: multiple views, sequence, comic, 2koma, 4koma, multiple images, turnaround, collage
Presence of Text/UI Elements: speech bubble
4. Image Manipulation and Resolution
Resizing and Scaling: resized, downscaled, source larger
Compression: lossy-lossless
5. Metadata and Provenance
Artist Information: unknown artist, banned artist, artist request, artist name, doesnotexist
Branding and Copyright: signature, username, logo, watermark, copyright name, copyright symbol
Temporal Information: oldest, old, early
6. Miscellaneous and Unclear Tags
Ambiguous or Subjective: what, off-topic, tagme, unclear
Noise/Artifacts: adversarial noise
Image Types: photo, icon, 3d
The complete list I use in generations:
ai-generated, ai-assisted, stable diffusion, nai diffusion, worst quality, worst aesthetic, bad quality, normal quality, average quality, oldest, old, early, very displeasing, displeasing, adversarial noise, unknown artist, banned artist, what, off-topic, artist request, text, artist name, signature, username, logo, watermark, copyright name, copyright symbol, resized, downscaled, source larger, low quality, lowres, jpeg artifacts, compression artifacts, blurry, artistic error, bad anatomy, bad hands, bad feet, disfigured, deformed, extra digits, fewer digits, missing fingers, censored, bar censor, mosaic censoring, missing, extra, fewer, bad, hyper, error, ugly, worst, tagme, unfinished, bad proportions, bad perspective, aliasing, simple background, asymmetrical, monochrome, sketch, concept art, flat color, flat colors, simple shading, jaggy lines, traditional media \(artwork\), microsoft paint \(artwork\), ms paint \(medium\), unclear, photo, icon, multiple views, sequence, comic, 2koma, 4koma, multiple images, turnaround, collage, panel skew, letterboxed, framed, border, speech bubble, 3d, lossy-lossless, scan artifacts, out of frame, cropped, [abstract], [doesnotexist], IMPORTANT NOTE!
I use a prompt like this:
artist:, {prompt},
masterpiece, best quality, good quality, newest, very awa, absurdres, highres, mandatory - artist:, masterpiece, best quality, good quality, newest, very awa
desirable - absurdres, highres
Adding in the beginning artist:, even without specifying a particular artist, improves the overall quality of the generation.
I understand if you want something shorter, such a promt of only official tags I will test too.
So to your attention small and medium negative prompt:
S:
worst quality, worst aesthetic, bad quality, normal quality, average quality, oldest, old, early, very displeasing, displeasing,M:
ai-generated, worst quality, worst aesthetic, bad quality, normal quality, average quality, oldest, old, early, very displeasing, displeasing, adversarial noise, what, off-topic, text, artist name, signature, username, logo, watermark, copyright name, copyright symbol, low quality, lowres, jpeg artifacts, compression artifacts, blurry, artistic error, bad anatomy, bad hands, bad feet, disfigured, deformed, extra digits, fewer digits, missing fingers, censored, unfinished, bad proportions, bad perspective, monochrome, sketch, concept art, unclear, 2koma, 4koma, letterboxed, speech bubble, cropped, [doesnotexist], XL:
ai-generated, ai-assisted, stable diffusion, nai diffusion, worst quality, worst aesthetic, bad quality, normal quality, average quality, oldest, old, early, very displeasing, displeasing, adversarial noise, unknown artist, banned artist, what, off-topic, artist request, text, artist name, signature, username, logo, watermark, copyright name, copyright symbol, resized, downscaled, source larger, low quality, lowres, jpeg artifacts, compression artifacts, blurry, artistic error, bad anatomy, bad hands, bad feet, disfigured, deformed, extra digits, fewer digits, missing fingers, censored, bar censor, mosaic censoring, missing, extra, fewer, bad, hyper, error, ugly, worst, tagme, unfinished, bad proportions, bad perspective, aliasing, simple background, asymmetrical, monochrome, sketch, concept art, flat color, flat colors, simple shading, jaggy lines, traditional media \(artwork\), microsoft paint \(artwork\), ms paint \(medium\), unclear, photo, icon, multiple views, sequence, comic, 2koma, 4koma, multiple images, turnaround, collage, panel skew, letterboxed, framed, border, speech bubble, 3d, lossy-lossless, scan artifacts, out of frame, cropped, [abstract], [doesnotexist],LET'S GO TESTING!
For testing will be used software written by me, if you want to use it, write in comments I will upload the code on github.
We start with a short and simple prompt and end with a long and complex one.
1girl, solo, S:
M:
XL:
Test:
2girls, (back-to-back:1.4), makima \(chainsaw man\), yoru \(chainsaw man\),
rella, (nanaken nana:0.95), u u zan, banishment, stu dts, (miv4t:0.95), (wlop:0.9), (suzumi \(ccroquette\):1.1), rei \(sanbonzakura\), torino aqua, (Morikura En:1.05),
alternate costume, touhou, cheongsam,
from above, (upper body:1.2),
holding sheath, hand on hilt, legs apart, lunge, unsheathed, ready to draw, dynamic angle, straight-on, looking at viewer,
old withered vines and ancient trees, a murder of crows descending into darkness, a small bridge over gently flowing water, quaint houses nestled by the river, S:
M:
XL:
Test:
1girl, fantasy, ethereal, floating hair, (iridescent), glowing, (aurora), galaxy background, floating, celestial, crystal clear, magical girl, heterochromia, detailed eyes, multicolored hair, flowing dress, constellation, shooting stars, sparkle, chromatic aberration, lens flare, depth of field, detailed lighting, starry background, bioluminescence, flower petals, translucent, angel wings, long flowing hair, gradient hair, pastel colors, hair ornament, jewelry, tiara, bare shoulders, detailed skin, strapless dress, cinematic lighting, dark background, glowing particles, butterflies, glitter, bubbles, rainbow, cherry blossoms, fog, volumetric lighting, caustics, reflective, gems, crystal, prism, detailed facial features, gold trim, gossamer, wind effect, light rays, luminous, glass, transparent, intricate details, necklace, bracelet, earrings, feathers, dramatic angle, dramatic lighting, bloom, glow, soft focus, fireflies, water droplets, venetian glass, crystal prism, rainbow reflection, pearl, silver accents, S:
M:
XL:
Test:
You can draw your own conclusion and write it in the comments.