Version notes
v1.0
First version that is on par/better than the illustrious mix
Model focus
3DCG/Semi realistic face; sharper/mature facial features with subtle skin texture.
Weakness
Noise pattern on image. Recommend using WAN 2.1 fp32 VAE and film grain postprocessing node.
Hands can be funky sometimes.
Anime style hairstyle can be hard to prompt
Recommended settings
(can see workflow via uploaded images)
euler + simple
steps 8-10, cfg 1
standard, simple, stable, works most of the time
downside - lacks texture, uniform texture, softer looking
euler + beta/(beta57-more tex)
steps 8-12, cfg 1
more texture than simple but can cause more artifacts
euler_a + beta57
steps 8-15, cfg 1
even more texture and details but also can get the image wrong
can cause the image to look more illustrative/less realistic
deis_2m / deis_3m + beta57
steps 3-4, cfg 1
similar to euler+beta57 but can have more stable structure while having as much texture
might have more realistic proportions compared to euler
downside - can be soft?
deis_3m is longer but might have better structure
Remarks
Personally use 1024x1536 resolution most of the time. Lower resolutions makes the noise pattern more noticeable when upscaled.
Upscale model recommendation: 2xLiveActionV1_SPAN. Keeps the image similar to the original without extra sharpening or lost of details while 2x resolution. Also fast.
WAN 2.1 fp32 VAE and film grain postprocessing node for noise pattern reduction.
Description
FAQ
Comments (22)
So what's the deal with KREA2? The consensus I seem to be getting is that Illustrious is still more precise/advanced?
Maybe it's still slightly ahead in really explicit nsfw things, but in every other way it has been surpassed by Krea2 tbh
@Asaghon and what exactly are those 'other ways'?
@ODBsGoonerExtravaganza Illustrious is based on SDXL, Krea 2 is a far more advanced model just in every possible way, get a chatbot to do a full list of how much better Krea 2 is.
@ferrrett33 and there you go, another 'in every possible way ' non-answer.
For example I don't see a huge difference between Fubuwa's later illustrious checkpoints (4.2 - onwards) and current KREA one. So KREA2 has supposedly better realism'? Okay. What else? Huh? The examples going round on net show nothing that I couldn't and haven't have already done with Illustrious checkpoints.
'Get a chatbot'. Hillarious. Why do you people even bother typing at all when you know next to nothing yourselves? Consider for a second that neither of you two have anything to show for in terms of your own generations.
better composition, better colors, it understands style concepts, it understands specific objects, prompts can be very specific or very loose and you will gain a good image with very little trial and error. Creating your own style is possible but the ones it knows are insane and resemble their sources, reddit has a whole list of some and FAL on huggingface made loras targeting some more generic and meta style. Krea2 is great and for an open source model it is easily #1, only the paid models are better.
@kitekholin Still not quite sold (yet to see any example that would), but that's more like it. Thanks.
Gonna see how I fare training some character loras this weekend.
@ODBsGoonerExtravaganza why on earth are you so hostile? Krea2 (base model) will give you a better image every single time than any sdxl finetune. The only thing Illustrious has on it is maybe some very explicit content.
I'm comparing out of the box base model with just a bypass lora that generates great images with good prompt understanding, without upscaling or detailer shenanigans to years of sdxl finetunes that need more complicated workflows to get the most out of it. But if your happy with sdxl than keep using it. I used Illustrious until Krea2 came out too.
Just try it?
Edit: Training character Lora's is extremely easy and the likeness is great. Don't even need detailers to get the face right. Just try a LoKR rank 4 or 8 / sigmoid in onetrainer, it's amazing.
@Asaghon uh-huh. ok.
@ODBsGoonerExtravaganza holy fuck dude, see a therapist
Krea2 wins in:
- more creative
- adheres to what you prompt better; in illustrious you just add the tags and roll the roulette for the results
- background detail/structure
- more interesting compositions - can manipulate positions of subjects with prompt
- easier to get micro detail/textures vs illustrious
- lighting easier to manipulate
- one pass is enough while being faster than illustrious second pass+facedetailer.
illustrious wins in:
- better nsfw knowledge - s*x positions prompts work better
- anime style hair/booru tags; krea2 doesn't really know straight hair/blunt bangs/etc very well
in krea2 I was able to create the image closer to what I envisioned vs in illustrious. just the two illustrious wins that is missing though.
@ODBsGoonerExtravaganza we have no idea what you're using it for or how complex your images are. Maybe if you're doing something in a very narrow scope, SDXL or Illustrious nails that for you, in fact a certain checkpoint can nail it much better than others, possibly much better than Krea 2 does that cherry picked thing out of 1000 things it can do. If you do broad image generation like I do including hundreds of very different prompts and using dozens of checkpoints then it's no question that Krea 2 is so much more powerful and coherent. I ran controlled tests, think hundreds of prompts seed-identical tested on hundreds of model across 5 model families. SDXL got like a 10% quality score, Krea 2 gets like a 60% quality score, it's vastly better overall -- which doesn't mean there weren't any images that SDXL just did better, of course there were, except it was the other way around 10 times more often.
@Asaghon note how you're the only one here with his panties actually in a twist - while also being the one who began with sticking a laughing emoji under my question (which I deliberately avoided mentioning up until now).
Do not talk to me about therapists.
@fubuwa thanks as always man. That's a type of response I was looking for.
I've still got a huge backlog of illustrious (and wan 2.2, crucially) batches and character loras to complete so I was curious before diving in.
@ODBsGoonerExtravaganza I made several Anima loras already using krea2 to generate the training images, the results are on my profile. I have posted some Krea2 Images just can't remember which since they are mixed with the Anima images. Also what exactly is the goal? if its gooner content, I am assuming this by you username and profile content, Krea2 and Anima actually make for better nsfw content creation than illustrious, I tested this by creating a difficult concept.
Using an AMD r9700 32gb, I get anima turbo generations in 9-15 seconds and krea2 is about 30 seconds, for Illustrious It takes 15 secs for initial generations, 2nd pass and face detailer take it to 45secs to 110 seconds or more depending on image size. I find krea2 better overall since it take less trial and error than Illustrious and for gooner content Anima might be the better option in my personal opinion and experience.
Is BF16 variant available?
nope probably never will. i don't have hardware to use it also
Really beautiful style, but my computer performance is average. If you have time, would you consider developing an int8 quantized version Thank you so much!
will look into it but no promises
This is very interesting, we talk a lot about photorealism but there are very few checkpoints that actually do it. And this DOES. Many of the images you would really think is a photograph. Incredible job. You can make it even more realistic if you prompt it to a different decade and "amateur photo/photograph"
can someone tell me how to see the workflow? im new to comfy and don't know what teh fuck am i doing, it says that the workflow can be seen in the images but, i don't see anything xd
just drag and drop the image into comfyui

















