This LoRA will allow you to inference on the Krea2 Turbo model in the style of a reference image. It was trained with AI Toolkit using the experimental Krea2 reference image trainer with the turbo training adapter. It was trained on thousands of hand curated style pairs I have been building over the years. It was trained to handle 1-2 reference images.
For inference in Comfy UI you will need this custom node. ComfyUI-Krea2-Ostris-Edit Use the template in that repo with this lora for the effect. No trigger word is needed just image and prompt.
Description
FAQ
Comments (5)
I get much better results when I downsize the style reference image to 0.25 megapixels instead of 1MP. Also, I get much better results with the style lora strength set 0.5. I'm using the workflow from the github.
For the lora strength, at 1.0 is breaks the image. Fractured, blurry, blobby, plasticy, low contrast. At 0.4, the style transfer is still good, and the image is more crisp, but there are bad grid pattern artifacts.
great advice
Ostris, you are the 2nd person to answer my prayers. You are the goat, and you will be blessed.
I can't reproduce your results with this LoRA and your workflow in the github repo. How dependent is it on the text encoder? I see you have a different named one than I have saved in that workflow.
+1