Introduction
Krea2 Cat Tower is Krea 2 Turbo's fine-tuned model.
This model was trained with the goal of enhancing anime style.
Please refer to the official Krea 2 Turbo Huggingface repo about Krea 2.
How to Get Started with the Model
You need to use the webui that supports Krea 2 and INT8 ConvRot.
ComfyUI: v0.27.0 or later (v0.28.0 or later is required to use the INT4 ConvRot version.)
Forge Neo: 2.27 or later
Recommended Setting
It adheres to the official settings of Krea 2 Turbo.
Resolution: 1024-2048px
Steps: 8
CFG scale: 1
Sampler: euler, er_sde
Scheduler: simple, normal
Prompting guidelines
Please refer to the Krea 2 Prompting guidelines.
Unless you specify a particular style in the prompt, the default style will be anime-style.
License
Description
INT4 (W4A4) ConvRot quantized version.
If you have sufficient VRAM, I recommend the INT8 ConvRot version.
Current issues
Sometimes, sweat droplets may appear on the character's face even if you haven't specified it in the prompt.
FAQ
Comments (41)
5090 does get loud, don't it! 😁
@bkidderz My RTX 5090 makes a noise that sounds like coil whine in addition to the fan noise. Coil whine noise does not occur unless the system is under a heavy load, such as during AI inference or training. I sent it in for repair, but it was returned because the agency couldn't reproduce the problem😔
@nuko_masshigura тут главное вовремя надеть наушники, моя же 5090 щелкать любит при нагреве)
@nuko_masshigura 因为电感啸叫是正常现象
Does this model has a strong visual style? For instance, if i use a LoRA for a character from a franchise with a very specific art style, is there a risk that this model will alter it? I ask because i am looking for a model that doesn't significantly change the style of the character LoRA I'm using.
@antarek To enhance its anime-style rendering, this model has been trained on a specific art style. In my personal opinion, in the case of Illustrious, the art style of the base model influenced the art style of the character LoRA. However, in the case of Krea2 and Anima, I felt that if the character LoRA had a certain art style, LoRA's art style would have a stronger effect than the base model's art style.
@husong198912 No. The file size of INT8 ConvRot is roughly the same as that of FP8 Scaled, yet INT8 ConvRot is generally considered superior in terms of both accuracy and processing speed. Therefore, there are no plans to create an FP8 Scaled version.
good
This is the best version of Krea2 for creating anime-style images that I've ever used.
I like this version of Krea 2 – it has a very pleasant style. It's a pity there's no full‑steps version, and for 8 steps some prompts are too complex to execute, but even so, the model can still produce interesting results
It feels like the girls this model generates are heavily tuned toward nuko_masshigura's preferred facial style. I love that kind of aesthetic as well, so I'm really happy with it. Simply put, the girls are insanely cute.
Do I need extra NSFW LORA to Unlock NSFW? or this model is already unlock
@Hisa_eromiko Regarding v1.0, while some training on NSFW content was performed, its impact was toned down because it adversely affected the generation of SFW content; consequently, even when used in conjunction with a LoRA, it does not yield satisfactory results.
Are you trained this with Turbo directly? Or Firstly on Raw?
@kisoubourei501 I merged models trained using the raw model into the turbo model.
I found the model is not good at generating NSFW content... Sad
CIVIT community in a nutshell xD
will you release a fp8 version?
@demonkingoftyranny As I mentioned to someone else previously, I have no plans to create an FP8 version because INT8 ConvRot is generally considered superior in terms of both accuracy and processing speed.
@nuko_masshigura oh really? well then nevermind, i thought fp8 would be better quality.
@demonkingoftyranny fp8 is about 2.5 times slower than int8 from what i noticed, and int8 seems to be also more accurate than fp8 so yes.
I really wish all new models will have int8 versions and stop using fp8 (fp8 was of gguf q6 accuracy so it was always quite bad)
@AltairTheArc I cannot confirm that, the int8 convrot takes on my 1024x1536 1st pass (lets just look at the 1st pass for now as an example), 1.2 it / s, that are 0.83333... s / it, while fp8 scaled takes 1.12 s / it, so fp8 takes 1.344 times the time int8 takes. Maybe you have an older card which does not support fp8 / does not have fp8 hardware acceleration? for ada lovelace and blackwell fp8 should be fast. And from what i have read fp8 should have no visible quality degradation compared to fp16.
@demonkingoftyranny its 3060 so maybe, but even if speed gains are not as substantial to you, there is also the quality in question, and int8 convrot preserves just more quality than fp8, i dont want to lie here* but if i recal int8 convrot was toe to toe with gguf q8 while fp8 was only comparable to gguf q6 (gguf q8 is comparable to fp16)
*(im not sure about int8 convrot accuracy rating, i'm sure of fp8 rating)
For me 1024x1024 on fp8 is 6.5s per it while on int8 it's 2.9s per it.
@AltairTheArc well yes i believe you guys, i only asked for fp8 because i thought it would be better but it was lack of my knowledge. Being still a bit faster and better quality is great.
@demonkingoftyranny AMD graphics cards cannot use int8convrot quantised models because the ROCm on Windows lacks important operators.
@husong198912 and why do you tell me that? My condolences i guess?
@demonkingoftyranny it seems I replied to the wrong friend.In conclusion, the FP8 model is necessary.
@husong198912 Hardly, from what i remember from benchmarks, top AMD card was hardly preforming better than crummy GTX1650 Super (which was utterly ridiculous) - and that's when it even was preforming, because in a couple of tests it would just not run them half of the times (errors), as well as it would not run some AI models (whatsoever).
AMD top was on the bottom of the charts in all tests, pretty much, competing with nividias GTX 900, GTX 1000 and 1600 series.
Keeping FP8 (inferior format by accuracy ratting) just so AMD could crawl when Nividia can sprint would be a lunacy.
@husong198912 I don't own AMD cards, so I can't verify this myself, but wouldn't it be possible to use ComfyUI-INT8-Fast-ROCM? That said, if there is demand for it, I am considering providing an FP8 version. However, I'm currently busy preparing for the next version, so the release will likely be in the next version.
@AltairTheArc In reality, AMD graphics cards aren't actually that bad for AI image generation; the main issue is that the ROCm software stack on Windows is incomplete. Features like llama.cpp and Triton require manual compilation and source code modifications—processes that can sometimes fail, often without a clear reason.
@nuko_masshigura You are far too generous.and sincere thanks for your efforts.
AMD graphics cards require manual source code modifications to get Triton working properly; sometimes it succeeds, and sometimes it fails.
Amazing Checkpoint, looks great.
Do this merge teach the model more anime characters?
@kiheromasterki849 This model was not trained with a focus on adding anime characters.
It looks great at first glance, but the eye symmetry, which is really great in Krea 2, seems degraded. Perhaps a lot of AI-generated pictures in the dataset? Just a wild guess.
AMD graphics cards using the int8convrot model will cause comfyui to crash.The reason is
Triton on Windows with ROCm can't find <hip/hip_runtime.h> when compiling its kernel launcher, because the PyTorch ROCm installation doesn’t include the full ROCm SDK headers.
You can Edit the driver.py to enable Triton to find and load the AMD HIP dynamic link library (amdhip64.dll) at runtime.
You can find driver.pyfrom"X:\Comfyui\.venv\Lib\site-packages\triton\backends"
You can use this command to check if the operation was successful.
python -c "import triton; print('Triton backend OK')"
If this result is ”Triton backend OK“, it means you are successful.
question: which model is your favorite right now krea2, anima, chenkin, or illustrious?
@mosukoai It's Krea2. While it has drawbacks such as requiring high PC performance, censorship, and the Danbooru tags not working properly, the high quality of the output images and the responsiveness of the prompts are the best of any model I've used so far. I have some complaints about version 1.0 of this model, and I'm currently working to resolve them. I'd also like to improve the Anima model, but right now I'm focusing on Krea2.
Krea2 is vastly superior over all other models, it is also better than ZImage, Flux.2 Klein 9B, Qwen Image 2512. It is just insane how good Krea2 is and i am always amazed that this is a local model that i can run on my PC, it is definitvely the future. It is so much better than anima and illustrious, they do not stand any chance.
I'm surprised that there aren't that many anime checkpoints on civitai for Krea 2