WRTN_ERIE_STYLE V1

(TEST Version) ERIE (WRTN) STYLE LoRA

LoRA FEATURES

1. Learned Style: This LoRA is trained on the art style of ERIE, the mascot character for WRTN's character chat service, 'CRACK'.
2. Dual Functionality (Style vs. Chibi): This LoRA was trained with the dual goal of recreating the ERIE character and serving as a general art style LoRA. However, in its current version, using it purely as a style LoRA (without the 'chibi' tag) often yields anatomically superior and more stable results.
3. Test Version: This is a test version. Some tag contamination has been observed when using the 'chibi' tag.

TRAINING & COMPATIBILITY Base Model: NoobAI v1.1 EPS Compatible Models: Works well with models based on ilxl, noobai, and most other anime-style checkpoints.

HOW TO USE

Trigger Word: erie (wrtn) (Using just 'erie' might also work)

Recommended Tags: chibi, short (Without these, the output will mainly mimic the simple color palette of the original ERIE art. Use 'short' depending on the context.)

To Generate the ERIE Character: For a high probability of generating ERIE, include: white hair, red eyes, low ponytail

Recommended Weights:

• LoRA Weight: 0.7 ~ 1.5
• 'chibi' Tag Weight: 0.7 ~ 1.5

CURRENT LIMITATIONS (Test Version) (These can be somewhat mitigated by adjusting weights)

1. Anatomical Instability (Chibi): Anatomical errors are more likely to occur, especially when using the 'chibi' tag to generate the character. For better anatomy, consider using it as a style LoRA without the 'chibi' tag.
2. Backgrounds: The chibi style is more likely to break when generating complex backgrounds (i.e., anything other than 'white background' or 'simple background').
3. Tag Complexity: The chibi style may degrade as prompt complexity increases.
4. Line Art: High LoRA weights can result in messy or broken line art. Post-processing, such as using Hires. fix, is recommended to clean up the lines.

REASONS FOR LIMITATIONS (Excuses!)

1. New Techniques: I used unfamiliar settings while experimenting with new training methods.
2. Small Dataset: The model was trained on only 3 images, so it relies heavily on the base checkpoint's knowledge, which limits its own stylistic variation.
3. Image Quality: The original ERIE images were mostly low-resolution. The process of upscaling and preparing them for training may have introduced some unwanted noise patterns.