Untied Bow Tie - SDXL

※USE IT AT 1 WEIGHT OR IT'S LESS LIKELY TO WORK (OVER 1 CAN CAUSE MORE ARTIFACTS, BUT IF IT DOESN'T WORK TRY ADD 0.1-0.2)

※TRIGGER WORD IS "UNTIBOTI" (short for untied bow tie) AND YOU MUST INCLUDE WORDS LIKE "SPLIT LOOSE BLACK TIE" IN YOUR PROMPT BUT DO NOT USE THE WORD "BOW TIE" IN YOUR POSITIVE PROMPT

※REPEAT, DO NOT USE THE WORD "BOW TIE" IN YOUR POSITIVE PROMPT, OTHERWISE THE GENERATED IMAGES WILL INCLUDE "TIED BOW TIE", YOU CAN WRITE "BOW TIE" IN YOUR NEGATIVE PROMPT

※IT HAS VISIBLY LOWER SUCCESS RATE WHEN GENERATING WOMEN, OR SHIRT COLORS OTHER THAN WHITE

※ALWAYS GENERATE >8 IMAGE COUNT IN ONE BATCH, 60% oF tHe TiMe, It WoRkS eVeRy TiMe

RECOMMENDED SETTINGS:

Checkpoint: RealismEngineXL / JuggernautXL / NewReality / RealVis

Aspect Ratio: 13:19 / 7:9 / 1:1

Interface: Fooocus + Default Styles

Sampler: Euler a (30+ steps) (STRONGLY RECOMMEND) / DPM++ 2M SDE Karras or 3M SDE KARRAS (MUST USE 60+ steps) (HAIR MAY LOOK WEIRD)

LoRA weight: 1-1.2 (over 1 may cause apparent artifacts)

Template Prompt:

Positive: "untiboti, (bodytype)(race)(sex), (shot size), (environment), unbuttoned (color) shirt, split loose black tie, (other appearance description)"

Negative: "(bow tie:1.4), single tie, tuxedo"

—————————MORE EXPLANATION OF MY SETTINGS—————————

Ⅰ Checkpoint:

Realism Engine XL V3 works really well, and it's one of my favorite SDXL checkpoints so you can use that, JuggernautXL V8 also works fine, but Realism Engine is the best;

Ⅱ Aspect Ratio:

I trained the lora on three ratios, 1:1, 7:9, and 13:19. Because 7:9 is the closest to 3:4 and 13:19 is the closest to 2:3 (my most frequently used ratio for photo editing), and square is the default. Other ratios like 4:7 also works, however if you stretch out to 1:2, the untied bow tie will also be stretched, I DO NOT RECOMMEND RATIO OVER 1.75;

Ⅲ Interface:

Personally I only use Fooocus as my main SD UI, I use ComfyUI to do minor stuff but mostly just use Fooocus to generate pics and inpaint. I recommend checking all the default style options (V2, Sharp, Enhance) because it doesn't really affect how this lora performs;

Ⅳ Flexibility:

Overall this model is very flexible, it can do men and women, any race, any environment, however, due to the limitation of my dataset, it has a very high tendency to generate white shirt and black tie, and it also has a higher tendency to generate men, and to generate tuxedo jacket (if you don't put tuxedo in negative prompt). Shirts with colors other than white might be a little bit harder to make, but it is still possible if you generate multiple batches.

Ⅴ Sampler:

Because this model is mildly overfitted, the hair of the person may look a bit strange if you're not using euler a as your sampler. If you still need the texture of other samplers, like 2M SDE Karras, I suggest you use the lora to make the body and background first, and then inpaint the hair without using the lora. Sorry about that.

———————MORE RAMBLING IF YOU DON'T MIND READING———————

I trained this LoRA because it's extremely hard to make AI (including SDXL, DALLE3, Midjourney) generate this specific look of untied bow tie. Basically, as long as your positive prompt include "bow tie", it'll only generate tied bow tie, and even if you use alternate word, like undone tie, loose tie, it won't work. However, untied bow tie is actually a widely acknowledged, classic sexy look. So I decided to train one myself. It took me a huge amount of time to prepare the dataset, and then do the testing. This is my first public model, hence I would really appreciate you guys' feedback! Thanks! :D