LS Vividus XL版本v3.0 (ID: 855905)

Use "by <artist>" for styles
Add artists at the beginning for stronger effect

IMPORTANT: If you find the model lacking in some aspect, swap out the CLIP; UNet is flexible
Eg: for stronger artist styles/text, use Tofu CLIP
for stronger character knowledge, use Urang Diffusion/holodayo/kivotos CLIP

Previous versions

Experimental anime model merge. Model might fall back to default SDXL anime style on some prompts. I encourage you to experiment and would love to see your gens!
The FP16 fixed VAE is included.

beta2:
Updated merge recipe with 2 new models added. Model is now slightly better at composition and prompt understanding.
Recommendations:
Resolution: Carefully choose an aspect ratio that suits your composition. Mismatched aspect ratio can cause the model to refuse to listen properly to the prompt.
Positive prompt: Add "anime art style" at the beginning. Use a hybrid approach with a description of the composition followed by tags specifying the elements.
Negative prompt: zombie, comic, sketch
Sampler: DPM++ 2M SDE Karras
Steps: 20 or more
CFG: 8 or higher

beta1:
Recommendations:
Resolution: Carefully choose an aspect ratio that suits your composition. Mismatched aspect ratio can cause the model to refuse to listen properly to the prompt.
Positive prompt: Add "anime art style" at the beginning to reinforce anime style if required. Use a hybrid approach with a description of the composition followed by tags specifying the elements. Colour tags can bleed into other tags, use Cutoff/PromptAttention to fix this.
Negative prompt: zombie, comic, sketch
Sampler: DPM++ 2M SDE Karras
Steps: 20 or more
CFG: 8 or higher

描述:

Rebuilt from scratch with new recipe

训练词语: