CivArchive
    ← All articles
    Published June 24, 2026by liutyi

    KREA 2 Turbo

    2.9K views23 reactions7 comments on CivitAI0 collected
    comparative study

    Memes

    Let's start review of KREA 2 Turbo with memes.

    hoho meme

    or Woman yelling at a cat type of meme generated by KREA 2 Turbo:

    What is so special about it? this part:

    It is exact image used as meme template. Not something like scene recreated from prompt. The models know the meme. Like Nano Banana 1/2 or ChatGPT Image 1.5/2, grok do. Other models will create something based on text prompt. but these are same faces same place.

    Three Spidermen Pointing Meme

    Generated by KREA2 Turbo:

    you may ask "So what?", Emm again... it is exactly the meme. Not another version drawn using model than unaware of it. Like see spider-man-triple-meme-template and spiderman-pointing-at-spiderman-template the original meme series templates are:

    so model know this meme too. Impressive.

    Now when I share what impressed me the most for the first day of use, we can proceed with Intro and all the usual article things. Should we?

    Intro

    KREA 2 Turbo and KREA 2 Raw is open (not very, but still... will check license later) models from KREA. It is available (since Jun 22, 22026) as download for local generation on Huggingface. Raw version is supposed to be used for training mostly and Turbo for generations.

    CivitAi official KREA 2

    Those are available for generation onsite. Were published on May 28, 2026:

    Added month later

    Unofficial checkpoints

    Used in SD.Next are

    Some unofficial models on CivitAi are

    Model was released with LORAs from KREA on huggingface

    https://huggingface.co/collections/krea/krea-2-loras

    Image Quality

    I got some list of prompts that executed on every new model, so I can compare it to previous. It is three sets for now. I would post links to the prompts and full results in the article end. Let's see some images from those test sets.

    test v1 2048px

    Photorealistic

    Taxi window portrait

    Hand details

    Oslo Opera House (it is correct image, most of the models do not know it)

    model also know Obama, Arnold Schwarzenegger in Terminator 2 (will not post here since civitAi not happy with celebrity images so..) You can check it on test v1 2048px page.

    Abstract Art

    abstract art

    gold and sepia

    city theme abstract art

    should I say it is nice? or it is obvious?

    Art styles

    ok with abstract art but other art styles are not that good. Not only ASCII art not known by the model. Compare KREA 2:

    with Nano Banana 2 (I know it is SOTA and GOAT, but anyway)

    or with ERNIE Image Turbo

    I would give "less than expected" for paint styles.

    But it will be LORAs someday... right?

    I did also some photo and illustrative Styles test (click the link to see all).

    The results are not bad. Its able synthwave, pixel art, polaroid and other. Not produce expected result on papier-mache art or cast paper art style. So not all, but still not bad. some examples:

    funko pop art style.

    or caricature style

    synthwave style.

    Minimalist abstract art.

    returning to my general-purpose tests...

    test v2 1024 samples

    photorealistic

    hand and key - good.

    hands under the glasses - ok, it is also good

    broken mirror - and this one also good

    Illustrative

    sociotype (far from ChatGPT 2.0 illustrative quality, but still not bad)

    braking the paper wall (it is ok)

    gangster duck (I like it)

    solar system (again not GPT level, but not bad)

    test v3 1024px and 2048px

    this one done in both resolution already so we can compare output of 1024px ws 2048 px

    photorealistic

    2048px

    1024px

    fairy

    2048px

    1024px

    surreal 2048px

    complex concept

    Compare with Medium and Large

    KREA 2 Turbo

    pareidolia effect (looks like not able), but at least not put plain faces over the glass.

    BTW, I have got this also generated with commercial variants

    KREA 2 Medium

    KREA 2 Large

    I do not know.. ok. let see another

    Turbo

    Medium

    Large

    I do not know. like it is different for sure. but the examples aren't justified the cost of Large, but it might be wrong examples. And Large is recommended for photorealistic. So, it might need another test for those commercial versions to see what is really going on there.

    See Test v3 results of Turbo 1024px vs Turbo 2048px vs Medium vs Large

    Resources needed

    For local generations I got Intel iGPUs on Intel core ultra 9 185H and 285H with 64 GB and 128 GB of RAM. Both are able to generate with KREA 2. It consumes about 46 GB of "unified" (v)RAM in bfloat16. Just to compare with similar models Z Image Turbo 6B (Total parameters: 10.2B), FLUX.2 Klein 9B (Total parameters: 17.3B), ERNIE Image Turbo 8B (Total parameters: 15.3B), KREA 2 Turbo 13B (Total parameters: 17.1B)

    Generation time on Intel core ultra 9 285H iGPU

    • KREA 2 Turbo (8 step 128x128) generation time is 19 seconds

    • KREA 2 Turbo (8 step 256x256) generation time is 20 seconds

    • KREA 2 Turbo (8 step 512x512) generation time is 38 seconds

    • KREA 2 Turbo (8 step 1024x1024) generation time is 154 seconds

    • KREA 2 Turbo (8 step 2048x2048) generation time is 738 seconds

    • ERNIE Image Turbo (8 step 1024x1024) generation time is 95 seconds

    • Z Image Turbo (9 step 1024x1024) generation time is 85 seconds

    • FLUX.2 Klein 9B (4 step 1024x1024) generation time is 37 seconds

    On real GPUs it will be faster, but the relative speed should remain the same.

    Resolution range

    1024px and 2048px is definitely work, lets check some other.

    128px

    256px

    512px

    1024px

    that is good.

    Other

    Quick compare to ERNIE Image Turbo.

    Will generate the illustrations from article https://civitai.red/articles/31633/ernie-image-turbo with KREA2.

    Woman subway portrait (not defaulting to Asian)

    Chinese prompt generated by ERNIE Built-In prompt enhancer (Asian, that is ok there)

    same prompt translated back to English (Caucasian, and that is what you expect), Like you expect wall tiles and station name actually on a station not a subway car.

    Fingers. Same two Seeds. Seed 20260619 and Seed 21

    both OK.

    1920x1200 image is ok it should be negative space on a border and so on.

    but some gradients and "stains" are on a background.

    1600x400.. Emm ERNIE is kind of better there

    Will check if 12 steps will somehow help with that. not much actually

    Lighthouse (still a bit better on ERNIE but not by far)

    Artifacts check images

    there are some patchy/stains/noise... hmm

    omg, on ribbon it is takes its extreme.

    License

    https://www.krea.ai/krea-2-licensing you can read all that, but I just asked Claude to compare with Apache 2.0 (ERNIE Image Turbo) license and get:

    Apache 2.0 is a permissive, business-friendly, irrevocable open-source license with virtually no strings attached. Krea 2's license is much more restrictive — it blocks serious commercial players (>$1M revenue), can be revoked at will, requires content policing on your end, and puts all legal risk squarely on you. It's closer in spirit to a "community/non-commercial" license than true open source.

    Issues

    So let us summarize the problems:

    1. License is not an Apache 2.0, not critical for personal use, but anyway...

    2. Some patchy-stains-kind-of-artifacts on dark colors.

    3. Seems it is not good with a lot of text. Some text is ok.

    4. Most emotions filtered. Only neutral and smile remain.

    5. Need LORAs for real Art style uses.

    Links

    test v1

    https://wiki.liutyi.info/display/AI/KREA+2+Turbo+2048px

    compare to

    test v2

    https://wiki.liutyi.info/display/AI/KREA+2+Turbo+test+v2

    compare to

    test v3

    compare to

    All prompts can be found at

    Ai Art test series