The goal was to create a "turbo" variant with increased seed variance while maintaining strong prompt adherence.
Recommended settings: 16–25 steps, CFG 1.0 to 1.4 max
See the gallery grids for comparisons across samplers, steps, CFG, and seeds.
Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.
Description
Native FP8 (float8_e4m3fn)
FAQ
Comments (33)
thanks for this, it's quite good and int8 is fast!
will this be released in regular FP8? your regular center model is fantastic, i'm getting great NSFW results but seed variety is a definite need.
yes done ;)
Is this a Raw+Turbo merge? I'm currently using Raw + the Turbo LoRA at ~0.5 weight. Is this essentially the same thing?
raw plus turbo plus loras, plus a bit of my turbo model merge, just tested it on one prompt same seed sure not same thing 🤷♂️
1.0 is a lot better than default krea, can't wait to see how you keep improving this.
Whats the best quality, FP8 or INT8?
Excellent model! Currently the best Krea-2 hands down. A quick question: can you make a Convrot INT8 quant? It really bumps quality compared to the regular INT8 while retaining the same speed advantage. Very close to FP16 in my experience.
Thanks! I checked the tensor keys: it already uses Comfy native INT8 quantization with comfy_quant metadata and weight_scale tensors per layer, so it should be the new convrot-style INT8 rather than plain INT8. It needs very recent ComfyUI/comfy-kitchen support.
krea 2 center semiraw is what i hoped z-image turbo would be when it launched lol it's perfect for everything i need. this model unlocks the nsfw/raunchier prompts without the need for the filter bypass methods.
What do you suggest for a workflow if you don't mind me asking? I am just starting to use Krea2.
my gens all have a standard workflow embedded, you don't need anything special to run krea 2.
Thank you! Great work!
Ok, show of hands, who is running this with the 13gb of VRAM it recommends? And how?
I'm trying to run it and it insists that it needs 23gb.
Aha, never mind, it wasn't this model that was eating the extra VRAM, it was the Qwen text encoder, which is almost 6GB for some reason. I loaded that with the CPU and now it works. Does anyone know where I might find a smaller Qwen text encoder? 6gb just for that seems excessive.
qwen3vl_4b_fp8_scaled is 5 GB in size, maybe try that one. Also, I have 6 GB VRAM on my RTX 2060 and I can run the checkpoint with qwen3vl_4b_bf16 (8.6 GB in size) and 4 different LoRAs with ComfyUI's dynamic VRAM optimization. I've also got 16 GB of system RAM and a fairly fast SSD. Definitely give it a try if your GPU supports Comfy's dynamic VRAM loading.
I'm running the fp8 variant and it is an astonishingly good model. It dramatically improves NSFW generations, adheres to my prompts without any issues, and goes well with the LoRAs I've tried it with so far.
INT8 + Sage attention is pretty crazy, it's the fastest SOTA model out there by far.
Apologies for even asking for an FP8 of this, i stopped being lazy and updated my comfyui for INT8 support and it's working beautifully. 40 seconds for a 1.8MP image, pretty crazy.
No worries int8 support is pretty new in comfyui ;)
This is now my favorite model. Turbo was too fast and lacked creativity, base was too slow, this is a perfect mix. It understands so many concepts and can do pretty much every art style I threw at it. It even understands centaurs, which is a gaping whole in most models! Does great at rendering text, fast, efficient. Thanks for making this!
I added a bunch of images where I put it through its paces.
It plays very nicely with loras too, which was the final confirmation i needed to fully leave behind a few sdxl models i still used AND Anima for realism with fictional characters.
I just switched over to the int8 model after reading about it a bit and WOW it's almost twice as fast... Guess I'll need to get on int8 convrot support in ggufy and see what else can benefit from it
@ferretduck yes depends on the card from what I heard but especially 30series gets like 2x. I think on 50 series its barely 25-30%
This model produces excellent result BUT this isn't a "turbo" one at all if forced to use from 16 steps up to 25 steps.
yes that is why I called it semi raw
Incredible model, thank you for sharing this. Do you have any plans for a GGUF version or maybe release a BF16 version so we could convert it to GGUF ?
I will was waiting for real gguf support to drop down want custom nodes and did not check if city96 has updated
@TiwazM I'm using Neo, gguf is fully supported. Fyi I converted it from int8 to q8 with ggufy (thanks to Claude who pointed that software). Very easy.
First of all, thank you so much for making this incredibly good model. I really asked a lot of questions. This is the first time I’ve left a comment saying the model is good. Everything is perfect.
Would it be better if the semiraw were upgraded to the bf16 20gb model?
Best Krea2 model I checked out so far! Thank you for this. Is there info, on what loras were merged?
The best custom checkpoint at the moment, thank you! SemiRaw approach has great potential