PLEASE READ THE DESCRIPTION FOR UNDERSTANDING HOW MODEL WORKS

If you still don't wanna read, here is a good starting point:

Positive prompt:score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up,

Settings, how it works

CFG/STEPS

Basically for ever point of CFG add 5 steps, starting from 4/20 up to 11/55 (don't be scared of 10 cfg, it does amazing and follows more the prompt)

This is not write in stone, finding a good balance is the key, as a tip, if your prompt is short i would use a large cfg like 10 and few steps, like 30. Also i would start from 7cfg 35 steps

CFG works as intenteded, as lower it is, more freedom, but at higher it will follow more the prompts and if you prompt well this means more quality.

Steps: If your promts are large, the model needs food, that means more steps, buut if yout prompts are short, it needs less steps or it will start to overpaint things.

Positive promts:

None: it will follow more the prompt but i may loose pony quality

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up: What i use the most

Score_9: you can put any score, it will make different results, keep in mind putting a lower score doesn't mean bad quality, just different images

Positives that alters the quality, the model in reality is quite versatile:

rating_safe, rating_questionable, rating_explicit works like pony

realistic, irl, photorealistic, extremly deitaled, detail, hyperrealistic it will boost the realism

List of prompts that can change style:

3d

render

doujinshi

outline

lineart

digital art

disney art

ghibli

realistic

anime

cartoon

painting

2k, 4k,8k, 24k

I am sure there is more, but i have to discover

NEGATIVE PROMTS

all the list above will change the style, but 1 that changes a lot is 3D, if you put 3d in negative it will make more anime like, combine with outline positive or others to achive different styles and results.

Also don't need any negative, avoid putting long negatives especially with quality tags or will change the results a lot, use the negative for what is for.

SAMPLERS

2m karras is good, but not always the best, DDPM does also incredible results, 3M karras does images more warm but needs more steps (like 40+), SGM Uniform works also well, Heun can give lot of quality but alo needs more steps.

RESOLUTION

Is better and faster put higher resolutions without high res, than lower resolutions and high res

But these are the best resolutions for me:

512*768+highres x2, 800*1200, 1024*1024, 1024*1280, 1024*1364, 1024*1424, 1280*1280

The model as based in pony, works better with tags, it likes medium lenght prompts like 250 tokens, still you can use descriptions cos is merged with SDXL, but i recomend tags.

Stye loras may not work as intended, some words may have more weight than others, adjust it.

Finally, to create these models is lot of hours drinking coffe that i give for free, if you wanna support me, or want me to continue, consider buy me one!: Kofi