Introduction
Krea2 Cat Tower is is Krea 2 fine-tuned model.
This model was trained with the goal of enhancing anime style.
Hugging Face page is here:
https://huggingface.co/nukomasshigura/Krea2-Cat-Tower
How to Get Started with the Model
In order to perform inference, the following files are required in addition to this model.
Text Encoders (use one of them)
VAE
Turbo LoRA
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.)
Refer here for how to place files in ComfyUI.Forge Neo: 2.27 or later
Quantized and BF16 models
Only the INT8 ConvRot model is available on Civitai.
If you need other quantization models such as INT4 ConvRot, FP8 or NVFP4, please download them from the Hugging Face page.
v2.0 BF16 model are also available on the Hugging Face page.
All Quantized models and BF16 model are available on Civitai.
"int8" (INT8 ConvRot) is recommend for inference.
Recommended Setting
It adheres to the official settings of Krea 2 Turbo.
v2.0 is Raw model, so I recommend using Turbo LoRA for inference.
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.
I recommend constructing your prompt using the following structure.
Main Subject + Pose/Action
→ Appearance, Clothing & Details
→ Props & Materials
→ Composition, Framing & Perspective
→ Environment & Background
→ Lighting, Color Palette & Mood
→ Overall Aesthetic & Medium
Unless you specify a particular style in the prompt, the default style will be anime-style.
Training
Trained using a dataset of images generated with Cat Tower and a dataset of images not generated by AI.
100% Natural-language, no booru-tags
Limitations
Sometimes, sweat droplets may appear on the character's face even if you haven't specified it in the prompt.
License
Description
This is a Raw model. I recommend using Turbo LoRA for inference.
1) Datasets have been refined
The captioning method was changed from JoyCaption to Qwen 3.8/3.6 or Gemma 4. Additionally, all captions for the main dataset were manually checked and corrected.
Replaced some training images and added a small number of new training images.
2) Changed the fine-tuning method from LoRA merging to full fine-tuning using Fizgig.
3) Changed the size of training buckets from 1.0 MP to 1.5 MP.
Current issues
Sometimes, sweat droplets may appear on the character's face even if you haven't specified it in the prompt.