This model is an INT8_convrot version built from the official Comfy-Org BF16 checkpoint. This one works with the unmodified ComfyUI model loader and does not require any additional node packs to work. Since there's no need for any "on-the-fly-quantization" the initial load is faster. It's also a slightly smaller file size than the other INT8_convrot aggressive version from HF.
NOTE:
Make sure you have ComfyUI version 0.27.0 or newer installed to have the latest INT8 support/fixes.
I have confirmed this works fine with the (SFW/NSFW) Prompt Adherence Lora as well as the others floating around. (See Suggested Resources below)
It's been pointed out that this version is a bit more aggressive than the Comfy-Org official that just dropped and Silveroxide's versions. Those preserve txtfusion/aux layers @ BF16 as well. I haven't observed any issues with this one in testing, but am linking those here so folks can test/try. Quality on this model is VERY good and I cannot tell them apart for what it is worth.
HF "Official" INT8_convrot posted by Kijai reposted on Civitai here.
Silveroxide's INT8_convrot simple/rowwise
Speed comparison:
(See the infographic)
Supported chipsets:
INT8_convrot (This checkpoint)
Nvidia: RTX 20, 30, 40, 50 series
AMD: In Development / Emulated
BF16
Nvidia: RTX 30, 40, 50 series
AMD: Radeon RX 6000, 7000, 8000 series
FP8
Nvidia: RTX 40, 50 series
AMD: Radeon RX 7000, 8000 series
MXFP8
Nvidia: RTX 50 series
AMD: Not natively supported
NVFP4
Nvidia: RTX 50 series
AMD: Not natively supported
INT8
Nvidia: GTX 10/16 series, RTX 20, 30, 40, 50 series
AMD: Radeon RX 5000, 6000, 7000, 8000 series
License and Outputs
Krea does not claim copyright or other intellectual property rights over content generated by users of this model. Users are solely responsible for their outputs and any subsequent use of those outputs. As with other generative tools, the nature of a user's inputs influences the outputs produced, and prompts may produce images that implicate third-party rights. Users are solely responsible for assessing and addressing those risks. See the Krea 2 Community License for more information.
Links
Product: krea.ai/image/k2
API partner: fal.ai/krea-2
Enterprise: krea.ai/enterprise
Moodboards: krea.ai/moodboards
Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.
Description
Initial V1 release
FAQ
Comments (13)
@daceheg192491 Can You share script that allows conversion? I want to convert different / unc ( Krea2TurboBadmilk) model. Thanks in advance
It's actually slower than FP8 Krea2 models on my 3060 Ti 8gb, 32GB RAM. latest Comfy
Bummer. Make sure you're not running with startup options like --fast or fp16 accumulation enabled. Are you using cuda++ or any other options like that?
@daceheg192491 I typically use the KJ loader and force FP16 accumulation(which definitely speeds up gen), but for this I am trying on the default diffusion model loader. I will try without just basic workflow, no LoRA/enhancements. And I don't have flags like that on my launch.
@daceheg192491 super basic workflow. 8 step euler simple. Gen time for this model is 42 seconds(after first run, then new prompt), 32seconds for FP8 model. No changes have been made between runs besides the prompt(basically just a tiny alteration so can go through textencoder).
@AI_DK Try without FP16 accumulation. Also try using "Patch Sage Attention KJ" node and select the cuda++ option (or auto if that doesn't work). Cuda++ gave me the fastest results.
@daceheg192491 It's without any modifications just basic workflow, no fp16 accumulation. Also I don't have sageattention, could never get it to work.
@AI_DK No but if you have KJ's node installed there's a node called that you can wire into the model and one of the options is cuda++. It made mine faster.
@daceheg192491 I need sageattention for that, it wants the module when I try with that node/setting.
CUDA 13.0 / Python 3.13 ComfyUI install is the key to unlock speed boost. Without it, even an RTX 5090 will go slower than FP8. It can be installed for all 20/+ Nvidia GPUs, but to get a speed boost, the GPU must also have the physical silicon circuits unlocked to calculate whole numbers natively:
RTX 20-Series: Bottlenecked by VRAM.
RTX 30/40-Series: Includes native low-precision hardware. It gets a noticeable speed boost, but is ultimately bottlenecked by older memory speeds.
RTX 50-Series: Features newly redesigned Blackwell Tensor Cores built specifically for low-precision math. Combined with next-gen GDDR7 memory, it processes raw 8-bit math directly, resulting in an instant doubling of generation speeds.
Additionally, the Nvidia GPU driver should be on a modern 2026 driver branch (version 570.xx or higher) to act as the translation layer. Otherwise, the terminal will instantly crash with CUDA initialization error.
Is that actually right about requiring python 3.13? i couldn't find any information confirming that information, you got a source?
@redlucario1735 It requires CUDA 13.0, and the ComfyUI install (with 13.0) comes with python 3.13. If you have an older install, then keep it and do a second clean portable install, to isolate13.0/3.13 within that environment.
on rtx 3060ti 8gb 58seconds 1040px wide images, amazing perfect resolutions with other 3 Loras. 10/10 i'm amazed!