My first attempt at a Krea2 merge, based on Kroma 0.3.1 turbo with additional furry concept training, and LoRAs extracted from Willy's Furry Mix and Wulver. Good style flexibility, and excellent LoRA compatibility. LoRAs should be trained on Krea2 base.
Description
nvfp4 version
FAQ
Comments (8)
Did the homework for the black images
qwen3vl_4b_int8_convrot
qwen_image_vae
你好，我在使用这个检查点时，只会生成彩色噪点图片。其他Krea2检查点都没有这个问题，已经排查了很多问题，无法解决。
Does this also bring booru tag understanding into the model?
Some e6 concepts are here, but this is a natural language model so tags do not work very well.
@someyoungguy69 How does natural language work when you have several details to specify?
Is it as simple as preceding tags separated by comma with "she has" or "he has"?
e.g. "she has light fur, digitigrade legs, athletic body, claws, and a vertical scar on her eye."
Would that be correct for natural language? Hybrid models confuse me with whats actually the correct formatting for prompts or if its also a hybrid of both types.
@CarnalDiffusion To get a good idea of how to prompt this, take a look at the sample images. Describe everything you want to see, describe the setting, describe the character and their clothing or accessories. Describe the background, and describe the level of detail and style. Use full sentences. Experiment with it.
This is honestly an awesome model. I don't care about furry things, but it just responds so much better than other models to the prompts. I gets poses correct a lot more than most other models and it actually gets pretty good realism too when prompted correctly. And cherry on top is that is works great with the character lora's I already trained on base Krea2. This is my go to atm.
bf16 please :(