VilStyle

This LoRA was specifically trained for my friend Vil, aiming to replicate a highly consistent and adorable chibi anime art style they frequently enjoy. It ports this beloved aesthetic to the SDXL ecosystem, offering flexible and cost-effective generation for their personal use.

This is a pure style LoRA (text_encoder_lr = 0), meaning it primarily influences the visual output of the UNet and does not introduce new character concepts or specific elements beyond its artistic flair. It is designed to be highly versatile, applying its distinct style to a wide range of subjects and compositions you can describe.

--- Usage Instructions ---

Activation: To activate the style, the LoRA, please utilise the following triggerwords: chibi and vilstyle

Recommended Settings:

• LoRA Weight: Optimal performance is typically found at 1.0. Experiment to achieve your desired blend with the base model.

• Model: Illustrious-XL (May also work on NoobAI as it is essentially the same architecture)

• Sampling:

  • Sampler: Euler A CFG ++

  • Steps: 20 Steps

  • CFG Scale: 1.4-1.7

--- Style Description ---

The distinct style of this LoRA is characterized by:

• Chibi Aesthetic: Characters are depicted in an irresistibly cute, super-deformed (chibi) style with simplified anatomy, often featuring small bodies and proportionally large heads.

• Expressive Eyes: Large, bright, and highly expressive eyes are a hallmark, conveying a wide range of emotions with minimal detail. They often feature prominent highlights.

• Clean Line Art: Bold, clean, and consistent line art defines the characters and objects, giving images a polished and illustrative feel.

• Vibrant & Soft Colors: A delightful blend of vibrant, often pastel, color palettes with soft shading, contributing to the overall cheerful and gentle mood.

• Simplified Details: Features like hair and clothing are stylized and simplified, focusing on essential forms while retaining character recognition.

• Kawaii Appeal: The overall impression is one of overwhelming cuteness and charm, making it ideal for adorable character illustrations and emotive portraits.

--- Training Details ---

This LoRA was trained on a curated dataset of images to capture the desired artistic blend, using the Kohya SS GUI with the following parameters:

• Engine: kohya

• Resolution: 1024x1024 (with enable_bucket for optimal resizing)

• Network Rank (Dimension): 32

• Network Alpha: 32 (matching dim for optimal weight distribution)

• Optimizer: AdamW8Bit

• UNet LR: 5e-4 (5×10−4)

• Text Encoder LR: 0 (ensuring pure style training)

• LR Scheduler: cosine_with_restarts (3 cycles)

• Advanced: min_snr_gamma = 20 was enabled to enhance training stability and improve output quality.

• Batch Size: 4

• Max Training Epochs: 30

• Keep Tokens: 1 (suggests initial trigger word if used)

• Shuffle Caption: True (aids in regularization)