Style Comparisons for Krea2! (And other tips and tricks!)

Hey hey! So, here is a style comparison for Krea2(and a few other tips and tricks)! Yay!

Base prompt(Image 1):

• A lone young woman standing on a rocky cliff, holding a sword at her side. Wind moves her long hair and coat. A vast mountain landscape stretches into the distance beneath a cloudy sky. She looks toward the horizon, seen from a three-quarter view. Clear silhouette, natural pose, simple composition, strong depth and atmosphere.

Now, to these, I appended: Image 2: In Anime Key Visual style, generate: {Base prompt} Image 3: In Watercolor Anime Illustration style, generate: {Base prompt} Image 4: In Berserk Manga style, generate: {Base prompt} Image 5: In Crayon Shin Chan style, generate: {Base prompt} Image 6: In adventure time style, generate: {Base prompt} Image 7: In marble sculpture style, generate: {Base prompt} Image 8: In lego set style, generate: {Base prompt} Image 9: In Watercolor Anime Illustration style, generate: {Base prompt} Image 10: In Minecraft style, generate: {Base prompt} Image 11: In funko pop style, generate: {Base prompt}

These were the styles I tested. You can find the styles that the base model knows here: library

Anyways, my settings: Resolution: 1280x2048 Sampler: whatever Moedayo devs set it to Cfg: whatever Moedayo devs set it to (shows 3.5) Steps: whatever Moedayo devs set it to. (Shows: 25) Seed: 1548958856 No LoRAs. I recommend the Turbo checkpoint, specifically this one on moescape as I believe it's the BF16 variant(the highest quality variant.)

The last image is how I set the resolution. On moescape, Krea2 can do upto 2048 pixels each side, for a maximum of 2048 x 2048. However, for local users, it can do upto 6megapixels, natively, after which it starts hallucinating. After 8MP it becomes kinda unusable. However, to go past that, you can use Krea Enhance & Upscale, and you can even go past 10k pixels on a single side. I know, crazy right?

Additional info: Krea2 is a 12B parameter model, compared to Anima(2B and a 2.9B variant now exists.) this makes it much larger, and much better(imo) Than many other checkpoints. I even like it better than GPT2.5, seedream and many other Image models. Ofcourse, taste is subjective.

It was trained in a way that makes it ignores many NSFW prompts, however, you can use this LoRA(spell):

Krea2 TextFusion Refusal-Reduction LoRA

I additionally like this spell too: DETAILCORE FOREVER 🔍 BY STX

Additionally, time for some tips and tricks; For even superior prompt adherence, use Krea2 Raw with 52 steps. (Ik, crazy right?) However, keep in mind, Raw version can only do 1MP images!!! Then, you can use it as Image 2 Image with Krea2 Turbo with higher resolution ^^ Use only NL. Tags don't work as well!! Krea2 can take style image as reference as well! ^^ Upto 10 images locally iirc! Be descriptive, however, try to keep prompt in 2000 words. It knows many languages! Same prompt in different languages can have different effects If I am correct! So, have fun! This means, you can type prompts in your native language, or in a language you are learning(maybe a good way to better learn a new language?) You can do text! In multiple languages! For my more advanced friends, I was told its text capabilities are in between Anima, and Ideogram.

Other people have also tested Krea2 like: StevenK. [Their tests are on the official moescape discord I believe.]

TheCuteInYellow [Whom I believe have a very large gallery of tests, testing which material cam be prompted on armor.]

Special Credits to: PPX [Who gave me the beans to test it out.]

Sorry for my poor English friends! It's not my first or second language! ✌️