Let's unlock true NSFW KREA-tivity!
This finetuned checkpoint for Krea 2 takes the aesthetics to the next level, while unlocking the NSFW capability by this great model by default!
Recommended Settings including the Turbo LoRA:
Steps: 12
CFG: 1.0
Sampler: euler_ancestral
Scheduler: Beta
FP8 Download
For some reason, when you upload multiple formats (BF16 + FP8) on one model, which they now support, they hide the FP8 under the BF16 model group in the side menu...
Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.
This model may only be merged or re-used with this licence intact.
Description
FAQ
Comments (8)
No int8???
I dunno how to convert to that.
Could this be used in AI-toolkit to train LoRAs on via the Krea 2 (raw) model architecture.
I think so, yeah. You just need to point to this model instead of the original model. It's a bit of a mess to do without splitting it up into parts, but it's doable.
Would you be able to start providing Int8 Convrot maybe? c:
In comfyUI you can use the single node Star Ultimate Model Converter to convert any BF16 Krea 2 models into Int8, Int8_ConvRot, nvfp4, mxfp8, fp8, fp16 or fp32. Turbo models can also be converted into Int8_ConvRot, so they become even faster at per step generation without quality loss.
I used it several times, and it works great.
https://github.com/Starnodes2024/comfyui-starnodes-modelconverter
Can't PR your nodes and push but still:
Bug: node fails to load with comfy-kitchen ≥ 0.2.x (TensorCoreConvRotW4A4Layout import error)
Environment: ComfyUI portable (Windows), embedded Python 3.13, torch 2.10.0+cu130, comfy-kitchen 0.2.16, RTX 5090 (SM120).
Symptom: On import, both star_ultimate_converter.py and star_model_converter_pro.py raise:
ImportError: cannot import name 'TensorCoreConvRotW4A4Layout' from 'comfy_kitchen.tensor'
comfy-kitchen 0.2.16 ships TensorCoreMXFP8Layout, TensorCoreNVFP4Layout, TensorWiseINT8Layout, TensorCoreFP8Layout, TensorCoreAWQW4A16Layout, TensorCoreSVDQuantW4A4Layout — but not TensorCoreConvRotW4A4Layout.
Root cause: All four layouts are imported on one line, so the single missing name fails the whole from comfy_kitchen.tensor import .... Control falls into except ImportError, which also loses ck, ck_registry, and the three layouts that do exist — so every deep-quant format breaks (fp8, nvfp4, mxfp8, int8, int8_convrot), and the nodes effectively don't register. Only int4_convrot actually needs the missing symbol. requirements.txt pins comfy-kitchen>=0.1.0, so current installs hit this.
(Secondary: the fallback print("⚠️ ...") can throw UnicodeEncodeError on a cp1252 Windows console and mask the real error. Harmless if the launcher forces UTF-8, but worth dropping the emoji.)
Fix: Make only the convrot-w4a4 symbol optional, in both files:
from comfy_kitchen.tensor import TensorCoreMXFP8Layout, TensorCoreNVFP4Layout, TensorWiseINT8Layout
try:
from comfy_kitchen.tensor import TensorCoreConvRotW4A4Layout
except ImportError:
TensorCoreConvRotW4A4Layout = None # int4_convrot unavailable on this comfy-kitchen
And guard the int4_convrot branch so it raises a clear message instead of AttributeError-ing on None.quantize(...):
elif target_format == "int4_convrot":
if TensorCoreConvRotW4A4Layout is None:
raise RuntimeError("int4_convrot requires TensorCoreConvRotW4A4Layout, not available in this comfy-kitchen build. Update comfy-kitchen or use another format (e.g. int8_convrot).")
layout = TensorCoreConvRotW4A4Layout
Verified: With this patch all 4 nodes register on comfy-kitchen 0.2.16, and int8_convrot runs end-to-end (TensorWiseINT8Layout.quantize(x, per_channel=True, convrot=True, convrot_groupsize=256) succeeds). Only int4_convrot stays unavailable until comfy-kitchen re-exposes TensorCoreConvRotW4A4Layout.
What turbo lora do you recommend for 12 steps?