Main activation tags /// Main outfit (top to bottom)
A chubby man with glasses /// He is wearing a blue polo shirt, dark pants and sneakersStyle control:
Anime screenshot. - anime TV show style [put the prompt at the end]
Recommended samplers:
- [Krea 2 Turbo] Euler Ancestral Scheduler: Simple CFG: 1.0 Steps: 8 Lora strength: 1.0
Prompt tips (green - positive, red - negative):
- He is wearing a white shirt, red polo shirt, purple polo shirt, grey newsboy cap, white and green striped shirt over a blue undershirt, brown coat, cowboy hat, red scarf- additional canon clothes
- He has facial stubble and messy hair - messy version
Last image in the gallery can be imported into ComfyUI with a basic workflow:
No custom nodes required.
Description
Seiichi Kinoshita from SHIROBAKO trained on Krea 2 with Musubi Tuner.
This is a beta/proof of concept release to see how Krea 2 handles stylized character Lora training. Also because I haven't seen many pure character Loras published on Civit yet.
The Lora can probably be much further improved with better captions/training settings. For now I used the default training settings recommended by Kohya in the Musubi Tuner documentation, however, I think there are probably better settings/hyperparameters you can use to improve generalization and character likeness.
I also have no experience with Krea 2 inference. I literally trained the Lora before I even had a functioning Krea 2 workflow in Comfy lol. For now I am using mostly the official Comfy template with Turbo fp8 model.