Update 11/13 (multiple characters):

Because I've been asked by a few people:

you can prompt multiple characters:

char count, character names, position of the characters, quality 

or even combine multiple characters into one:

1girl/1boy, character names (one with cosplay keyword), quality, physical attributes to preserve 

Update 11/3 (from Illustrious paper):

Complicated desired outputs = Complex prompts with mix of natural language and tags

Complex prompt structure and order:

1. char count (1girl, 2girls), characters' names (if an existing character like hatsune miku)

2. natural language describing the output, with periods separating sentences

3. list of tags

4. quality tags at the very end (usually just "masterpiece")

Simple Prompt Example:

1girl, hatsune miku, angel, masterpiece, general

Result: Basic, but creative interpretation

Complex/Upsampled Prompt (using TIPO):

1girl, hatsune miku. 

An illustration of a girl with long white hair and wings. she is wearing a school uniform with a red bow on her head and a pair of headphones on her ears. the wings are spread out behind her, creating a sense of movement and energy. the overall style of the illustration is anime-inspired. 

solo, skirt, feathered wings, necktie, smile, very long hair, collared shirt, long hair, headset, blue eyes, aqua necktie, looking at viewer, black footwear, black skirt, twintails, grey shirt, bare shoulders, detached sleeves, full body, zettai ryouiki, closed mouth, miniskirt, sleeveless, boots, thighhighs, shirt, standing, wing collar, aqua hair, sleeveless shirt, pleated skirt, angel wings, absurdly long hair, wings, black thighhighs, 

masterpiece, general.

Result: More detailed, consistent results with specific elements (wings, uniform details, etc.) being more precisely rendered

with this negative prompt:

worst quality, comic, multiple views, bad quality, low quality, lowres, displeasing, very displeasing, bad anatomy, bad hands, scan artifacts, monochrome, greyscale, twitter username, jpeg artifacts, 2koma, 4koma, guro, extra digits, fewer digits, jaggy lines, unclear 

UPDATE 11/2: potential helper at the beginning (or maybe end!) of your prompt!

(masterwork, x, y, z, award-winning, masterpiece, best quality, hyper-detailed, 8k uhd::1.4), 

x = image type, i.e. sketch, photo, portrait, manga page, etc.

y = character name(s)

z = artist (if specified)

(followed by character count (1girl, 2girls), character description, location, action, other details)

Example:

(masterwork, portrait, princess midna, fan no hitori, award-winning, masterpiece, best quality, hyper-detailed, 8k uhd::1.4), 1girl, large breasts, blue eyes, skinny, slim, sexy, gaunt, smile, blue textured bodysuit, cleavage, fur trim, outdoors, castle, looking at viewer, anime coloring, shiny skin,Cinematic Light,

Negative prompt: lowres, worst quality, bad quality, bad anatomy, sketch, jpeg artifacts, signature, watermark, artist name, old, oldest

Steps: 26, baseModel: SDXL, quantity: 4, width: 832, height: 1216, Seed: 2455922073, draft: false, nsfw: true, workflow: txt2img, Clip skip: 2, CFG scale: 6, Sampler: Euler a, fluxMode: undefined

Okay so, Illustrious XL.

Been messing with it a ton lately and thought I'd share some stuff I've learned.

Overview

My testing has revealed several key differences from other model types like Pony:

• Text generation is usually cleaner

• backgrounds show better coherence

• concept recognition = surprisingly robust

• significant reduction in watermarks

• importantly, negative prompts tend to behave better

  • though, as a trade-off, they are much more necessary

• additionally, I can confirm, as others have found, both style and character training seem to produce more reliable results

On a subjective note, I've found that I just tend to like the images produced more versus something like Pony. Also, I find that it's just easier to work with and to get better images quickly.

Optimal Settings

Based on both testing and official documentation:

• CFG Range: 4.5-7.5 (sweet spot around 5.5)

• Recommended Sampler: Euler A

• Steps: 20+ (24 recommended)

Working Prompt Structure

As I mentioned above, you definitely need to massage the prompt a bit more than most other models, at least at this point in time (Illustrious XL v0.1). As a result, you need to have quality tags in your positive prompt and a pretty extensive negative prompt to get it to work well.

Prompt Structure

Core Structure

1. Character count (1girl, 2girls, etc.)

2. Character names (if any)

3. Quality tags

4. Physical features & clothing

5. Pose & anatomical details

6. Environment/background

7. Additional quality/style tags

Positive prompt tags (put at end or beginning)

From the paper, their actual example quality tags are much simpler:

• "masterpiece"

• "general" (as an optional rating tag for safe for work images)

• "absurdres"

• "newest"

Technically these are all you need to produce a good image, according to the Illustrious paper. I've also corroborated this in my own testing.

However, I've found that these prompt tags can produce fairly consistent results when used in combination with those tags above:

perfect quality, best quality, absolutely eye-catching, 

and with more realistic/detailed imagery:

perfect quality, best quality, absolutely eye-catching, ambient occlusion, raytracing, 

• ambient occlusion/raytracing help with 2.5d/semi-real style especially

Negative prompt

After a lot of experimentation, I found this works the best, on average, pretty much every time:

lowres, (bad), bad anatomy, bad hands, extra digits, multiple views,fewer, extra, missing, text, error, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early,chromatic aberration, signature,artistic error, username, scan

or as a shorter alternative:

lowres, worst quality, bad quality, bad anatomy, sketch, jpeg artifacts, signature, watermark, artist name, old, oldest

Full prompt examples/gen data

Here's a couple examples to show how 90% of my prompts look:

• Black Cat / Felicia Hardy (Spiderman) Lora example use:

1girl, felici4, anatomically correct, proper proportions,
long white hair, large breasts, domino mask, black lips, athletic build,
well-defined standing pose, dynamic pose,
detailed urban environment, night scene, city lights,
glowing blonde hair, blue eyes, looking at viewer, shining skin,
from below angle, professional lighting,
masterpiece, best quality, absurdres, newest

-or-

perfect quality, high quality, masterpiece, absolutely eye-catching, ambient occlusion, raytracing, felici4, 1girl, long hair, white hair, large breasts, (mask), domino mask, blue eyes, black lips, skinny, glowing blonde hair, rainbow inner hair, looking at viewer, shining glossy skin, perfect huge breast, (goosebumps:1.1), solo, excessive sweat, ((from below)) 

lowres, (bad), bad anatomy, bad hands, extra digits, multiple views,fewer, extra, missing, error, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early,chromatic aberration, signature,artistic error, username, scan

Steps: 24, baseModel: SDXL, quantity: 4, width: 832, height: 1216, Seed: 1934634232, draft: false, nsfw: true, workflow: txt2img, Clip skip: 2, CFG scale: 5.5, Sampler: Euler a, 

• Ashley Graham (Resident Evil 4 Remake) Lora example use:

1girl, ashley_grah4m, anatomically correct, proper proportions,
neon lighting, glowing blonde hair, rainbow inner hair, blue eyes,
well-defined pose, standing pose, balanced pose,
detailed environment, professional lighting, clear composition,
looking at viewer, seductive expression,
masterpiece, best quality, absurdres, newest 

-or-

masterpiece, best quality, ashley_grah4m, 1girl, neon, glowing blonde hair, rainbow inner hair, blue eyes,looking at viewer, seductive

lowres, (bad), bad anatomy, bad hands, extra digits, multiple views,fewer, extra, missing, text, error, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early,chromatic aberration, signature,artistic error, username, scan

Steps: 24, baseModel: SDXL, quantity: 4, width: 832, height: 1216, Seed: 1115576560, draft: false, nsfw: true, workflow: txt2img, Clip skip: 2, CFG scale: 5.5, Sampler: Euler a, 

Creating Effective Backgrounds

Based on paper findings and testing, backgrounds require specific attention:

Structure for Background Prompts:

1. Environment Base:

detailed environment, [location type], clear composition

1. Architectural Elements:

[material types], [structural elements], [decorative elements]

1. Lighting and Atmosphere:

[time of day], [lighting type], [atmosphere effects]

Example Background Combinations:

Indoor Scenes:

luxurious room, detailed architecture, marble floor, ornate furniture,
crystal chandeliers, tall windows, decorative columns,
warm ambient lighting, soft shadows, volumetric lighting

Outdoor Urban:

detailed cityscape, modern architecture, glass buildings,
city streets, urban details, store fronts,
night scene, neon lighting, street lamps, ambient occlusion

Natural Settings:

detailed landscape, rolling hills, dense forest,
rocky outcrops, flowing water, detailed foliage,
golden hour lighting, atmospheric haze, dynamic clouds

Full Example:

1girl, sprThja, anatomically correct, proper proportions, full body, toned figure, large breasts, black hair, long flowing hair, rabbit ears, dark alluring eyes, confident smile, standing pose, well-defined pose, balanced pose, looking at viewer, purple leotard, deep cleavage, purple leggings, purple gloves, yellow cape, detached collar, luxurious detailed room, marble floor, ornate furniture, detailed architecture, night scene, professional lighting, warm ambient lighting, soft shadows, clear composition, masterpiece, best quality, absurdres, newest 

Background Best Practices:

1. Start with broad environment definition

2. Add specific architectural or natural elements

3. Include material descriptions

4. Define lighting and atmosphere

5. Maintain consistency with character lighting

6. Use environmental quality tags

Tips for Better Backgrounds:

• Add depth indicators (foreground, midground, background)

• Include atmospheric effects

• Specify clear lighting sources

• Use architectural details for indoor scenes

• Add environmental context

• Keep perspective consistent with character pose

Common Background Issues:

• Inconsistent lighting between character and background

• Perspective mismatches

• Lack of detail in middle ground

• Poor integration with character

• Missing environmental context

Solutions:

1. Use consistent lighting descriptors

2. Add specific perspective tags

3. Include depth and distance markers

4. Specify material and texture details

5. Use architectural or natural anchoring elements

Known Issues

• Multiple competing style tags tend to produce inconsistent results

• Needs specific prompting to work well

Best Practices

• Use those prompts from above!

• Start with minimal parameters

• Pay particular attention to lighting descriptors

• Monitor CFG impact on output quality

My Own Training Examples

If you want to give them a try, here's some of my own models that I've trained on various Illustrious merges as well as Illustrious itself:

• CicaLust: A style implementation focusing on anime/shiny skin, much like my CicaStyle style Lora for Pony

• Ashley Graham (Resident Evil 4 Remake)

• Kallen Kozuki / Kallen Stadtfeld (Code Geass)

• Black Cat / Felicia Hardy (Spiderman)

Additional help re: Angles/Lighting

angles, use/mix these after quality tags at beginning of prompt:

from above,
from below,
close-up,
portrait,
POV,
birds-eye,
wide shot,
isometric,
(+ view, depending)

lighting, after angle tags, beginning or very end of prompt:

Cinematic Light,
Hollywood Lighting,
Backlighting,
Rim lighting,
Soft lighting,
harsh lighting,
Dramatic light,
film-style contrast,
soft shadows,
harsh shadows,

(you may have to fiddle with the position and wording of those, but they should mostly work as is)

Conclusion

Since this is a very new model, expect ongoing discoveries about optimal usage patterns.

Feel free to ask questions or share your own findings and experiments!