🟣 INFO Notice

This article was last seriously updated in the last quarter of 2024, right before the release of NAI XL v-pred v1.0 (also Laxhar Lab's "last" version). Additionally, I'm on a long hiatus from image generation models due to my personal dissatisfaction of my findings. Today (whatever date it is right now), some information may be useless or outright wrong. Make of that what you will about the contents of this page.

🟣 INFO Terminology

  • INAI: IllustriousXL/NoobAI-XL abbreviation. Caveat: I mostly use NoobAI-XL now, so some techniques may not translate to IllustriousXL. I just assume it works.

🟣 INFO Multiples of 64

64 128 192 256 320 384 448 512 576 640
704 768 832 896 960 1024 1088 1152 1216 1280
1344 1408 1472 1536 1600 1664 1728 1792 1856 1920

🟣 INFO [WIP] Preset prompts and negatives

INAI and its finetunes

Prompt: original, cover image, album cover, textless version, newest, best quality, very aesthetic, absurdres, masterpiece

Negative: (speech bubble, sketch, lowres, text, cropped, worst quality, low quality, normal quality, adversarial noise, artifacts, jpeg artifacts, gif artifacts, bad anatomy, bad hands, fewer digits, bad proportions, signature, watermark, username, blurry, revision, comic, manga, 4koma, multiple views, chart, unfinished, chromatic aberration, artistic error, scan, abstract, artist name, censored, resized, ai-generated, ai-assisted:1.3)

Pony DiffusionXL and its finetunes

Prompt: score_9, score_8_up, score_7_up, rating_questionable, source_anime

Negative: score_6, score_5, score_4, source_pony, source_cartoon, censor, watermark, blurry, lowres


🟣 INFO [WIP] Samplers and Schedulers on INAI

These resources are useful:

Terminology

  • Ancestral: Ancestral Sampling is a discrete, simpler method focused on generating diversity in outputs by injecting noise in reverse steps. Adds controlled randomness at each step of the reverse diffusion process, mimicking a generative "ancestral" process where randomness represents uncertainty or diversity in the model's predictions. It is tied to discrete timesteps and operates within the standard diffusion framework.
  • SDE: Stands for Stochastic Differential Equation, a DE where at least one term is a stochastic process. Theoretically rich, continuous-time approach in creating randomness by using the SDE to model the forward and reverse diffusion processes as continuous-time dynamics.

Some samplers are dropped entirely because of low confidence of its capabilities. I have separated samplers into the following separate classes:

  • Deterministic Samplers:
    • Euler
    • Heun
    • Heun++2
    • LMS
    • PLMS
    • DPM Adaptive
    • DPM2
    • DPM++ 2M
    • UniPC BH2
    • iPNDM V
    • DEIS
    • DDIM
    • DDPM (deterministic, but different)
  • Higher order needs much higher steps and lower CFG. Stochastic Samplers:
    • DPM++ SDE, DPM++ SDE GPU
    • DPM++ 2M SDE, DPM++ 2M SDE GPU
    • DPM++ 3M SDE, DPM++ 3M SDE GPU
  • Ancestral Samplers:
    • Euler A
    • DPM2 A
    • DPM++ 2S A
  • Dropped Samplers:

These are the schedulers that are tested. Typings are made from a combination of the observations of the results and how it works:

  • Karras (Type 1)
  • Exponential (Type 1)
  • Beta (Type 1)
  • Simple (Type 2)
  • Normal (Type 2)
  • DDIM Uniform (Type 2)
  • SGM Uniform (Type 2)
  • Align Your Steps 10, AYS (Type 3)
  • Align Your Steps Generate in Iterative Steps, AYS GITS (Type 3)

🟣 INFO Interesting prompts

Color

  • limited palette sets an ambiguous multi-color theme.
  • color contrast sets an ambiguous contrasting multi-color theme to emphasize objects/things. relates to complementary colors.
  • analogous colors are colors close to each other.
  • monochrome sets an ambiguous one-color theme.
  • spot color to color in parts of the image to emphasize objects/things.
  • partially colored to color in parts of the image incompletely/unfinished.
  • high contrast
  • greyscale for black, white and grey colors.
  • COLOR theme where COLOR is any of these: white, grey, black, red, orange, brown, yellow, green, aqua, blue, purple, pink
  • colorful for emphasis on colors.
  • muted color darker low saturation colors.
  • pale color lighter low contrast colors.
  • pastel colors very light colors.
  • dark is a softer alternative to black theme.
  • light is an ambiguous alternative to add lighting. See instead the specific section later below.

Styles, art techniques and types

Prints

Composition, perspective

Lighting, lights, particles

Formats and Compositions: Subimages

Formats and Compositions: User Interface, Diagrams, Graphs

Others

🟣 INFO Undesirable prompts

Metadata

These are metadata tags on danbooru indicating unwanted or undesirable images:

Text, Symbols

Text, logo, symbols that reside in the artwork:

Formats and Compositions: Subimages

These artwork formats/compositions I have low confidence in when generating or training. This category involves "subimages" or "self-similar subimages" that reside in the artwork:

Formats and Compositions: User Interface

These artwork formats/compositions I have low confidence in when generating or training. This category involves interface elements that reside in the artwork:

Formats and Compositions: Graphs, Charts, Sheets

These artwork formats/compositions I have low confidence in when generating or training. This category involves images containing graphs, charts, sheets and other arranged information:

Formats and Compositions

These artwork formats/compositions I dislike or low confidence in when generating or training:


🟣 INFO [WIP] Artist names

Currently, most artists are tested using the workflow below (please open in comfyui) using noobai-XL-Vpred-0.65s-cyberfix:

οΏ½klzzwxh:0011οΏ½

Terminology

Each entry has the following information:

  1. artist tag as taught to INAI.
  2. Username linked to the danbooru artist page
  3. ~123 artworks count on danbooru
  4. 2 integer likeness score between 1 and 3. Higher scores correlate to closer style emulation to the original artist. Lower scores correlate to either ineffective emulation or adaptation of the original artist.

Here are various tags :

  • [Multi-formats]: This artist has undesirable or unsuitable artwork formats for teaching models their style.
  • [Exploratory]: This artist does not produce stylistically consistent artworks.
  • [Biased]: This artist biases heavily in making artworks around a (very) small number of concepts reducing effectiveness on different kinds of art.

UNKNOWN

Not (thoroughly) tested artists:

βŽ—
βœ“
1
2
3
4
5
6
7
8
names aren't even all correct. these are lifted from civitai users' prompts.
(asteroid ill:1.2) (yun ling:1.2) (kan liu \(666k\):1.1) (greenteaa:1.1)
(tidsean:0.9),(misyune:1.1),wanke,(rei \(sanbonzakura\):0.9),suihei sen
tekito midori <lora:mqgire:.3
furu00tck
nyalia
mengxuanliart
iiiichimaru 03 

GOOD

INAI succeeded or made acceptable results in learning the style of these artists:

nyantcha Nyan β˜… ~2200
example

toosaka asagi ι ε‚γ‚γ•γŽ ~1900
example

tomose shunsaku γƒˆγƒ’γ‚»γ‚·γƒ₯ンァク ~1700
example

kaamin \(mariarose753\) γ‚«γƒΌγƒŸγƒ³ ~1600
example

fujima takuya θ—€ηœŸζ‹“ε“‰ ~1600
example

yd \(orange maru\) YD ~1600
example

k-suwabe ケースワベ ~1500
example

agahari をガハγƒͺ ~1500
example

dishwasher1910 D.I.S.H ~1300
example

modare γƒ’θͺ° ~820
example

pottsness pottsness ~590
example

missile228 MISSILE228 ~530
example

araneesama Araneesama ~500
example

gsusart GSUS ~470
example

dino \(dinoartforame\) DIno ~470
example

remsrar REMSRAR ~460
example

koh \(minagi kou\) ζ΅·ε‡ͺコウ【KOH】 ~440
example

liduke Liduke(ζ—₯子) ~370
example

wlop wlop ~360
example

wangxiii BW-ring ~340
example

turisasu ツγƒͺγ‚΅γ‚Ή ~310 [unsuitable images]
example

reoen γ‚ŒγŠγˆγ‚“ ~310
example

hara kenshi はらけんし ~300
example

ciloranko CiloRanko ~200
example

z3zz4 Z3zz4 ~170 [unsuitable images]
example

namiki itsuki 双木 樹 ~140
example

BAD

INAI failed or made unacceptable results in learning the styles of these artists:

fune \(fune93ojj\) ちね ~110 [unsuitable images]
example

eita 789 γ‚‘γ„γŸ ~90
example

eboda-x ι’ θŒ„ζ΅‡ζ°΄ε·₯ ~90 [unsuitable images]
example

yuyumu 鱼俞木 ~67 [unsuitable images]
![example](mylink){30%:30%}

doo58455 Doodoo🐾 ~45
![example](mylink){30%:30%}


🟣 INFO Character Design Prompts

Each tag/prompt is sorted from the character's top to bottom. For example, hair details go first while lower body details go last.

Hatsune Miku

  • V2 Design, default design:
    vocaloid, hatsune miku, 1girl, aqua hair, hair between eyes, hair ornament, headset, twintails, very long hair, aqua eyes, aqua nails, aqua necktie, bare shoulders, black sleeves, detached sleeves, grey shirt, necktie, number tattoo, tattoo, shirt, sleeveless, small breasts, skirt, boots, thigh boots, thighhighs
  • Append Design:
    vocaloid, vocaloid append, hatsune miku (append), 1girl, blue hair, very long hair, long hair, hair ornament, grey eyes, grin, bare shoulders, high collar, sleeveless, sleeveless shirt, white shirt, alternate costume, black gloves, gloves, elbow gloves, bridal gauntlets, anklet, jewelry, blue nails, center opening, clothing cutout, navel, stomach, thigh cutout, thigh gap, black pants, no shoes

Kasane Teto

  • Synth V Design:
    synthesizer v, utau, kasane teto \(sv\), 1girl, ahoge, red hair, drill hair, twin drills, hair ribbon, hair bow, white bow, white ribbon, parted bangs, hair between eyes, hair ornament, shoulder boards, shoulder belt, black belt, flat chest, red trim, striped clothes, striped jacket, grey jacket, long sleeves, sleeves past elbows, petticoat, black sleeve cuffs, double-breasted, buttons, navel, grey skirt, striped skirt, layered skirt, skirt set, red belt, kneehighs, red socks, knee boots, lace-up boots, black footwear, cross-laced footwear

🟣 INFO [WIP] Scoring AI generated images

1. Use an aesthetic scorer model

There are many aesthetic scorer models that aims to quickly determine whether an image is considered more "aesthetic" than another by reducing all of it into a single floating point number (the aesthetic score, ascore). Usually it ranges between 0 to 1, or 0 to 10. The current method used in INAI is waifu-scorer v3.

2. [WIP] Use an image noise estimation/measure

Find a technique that estimates image noise.

3. Manually measure properties of an image.

Depending on what kind of property you are searching for and differences of the image are, this method may not be effective way to determine an image's quality. Additionally, it will be hard to compare multiple similar images. I suggest you create a small score range for each thing you measure, like an integer between 1-3.

The following lists examples of things to look out when comparing images.

  • Defects and quality of the anatomy, objects, things and background.
  • Defects and quality of lines, gradients, patterns, colors, brush strokes.
  • How things are composed.
  • Presence of image noise. Use filters in an image editor like GIMP, Krita, Photopea or Photoshop to find them.
  • Adherence to prompt. This includes artist styles, objects, poses, actions, etc.
  • Flexibility in steps, CFG, samplers, schedulers, etc.
  • Personal preferences.

Bottom page

Edit Report
Pub: 14 Nov 2024 05:20 UTC
Edit: 23 Feb 2025 13:36 UTC
Views: 135