Lamb (Kindred)

This LoRA is dedicated to capturing the appearance, style, and essence of Lamb, the graceful hunter and one half of Kindred from League of Legends.

Unlike a pure style LoRA, this model was trained on both the UNet (visuals) and the Text Encoder (concepts). This allows it to understand the core attributes of Lamb as a character, providing greater consistency in generating her signature features, from her mask to her ethereal presence. It is well-suited for creating detailed character portraits, action scenes, and artistic illustrations.

--- Usage Instructions ---

Activation: To generate Lamb, use the primary trigger word lamb (kindred). Reinforce the character's appearance with descriptive tags for best results.

• Primary Trigger: lamb (kindred)

• Essential Tags: white hair, lamb mask, wolf mask, animal ears, hooves, glowing, holding bow

• Negative Prompts: Consider adding human ears if you find the model generates both sets.

Recommended Settings:

• LoRA Weight: 0.7 to 1.0. A weight of 0.8 is a good starting point.

• Model: Any SDXL-based model. Results will vary depending on the base model's style (e.g., anime, semi-realistic).

• Sampling:

  • Sampler: Euler Ancestral CFG ++

  • Steps: 20

  • CFG Scale: 1.4

--- Character & Style Description ---

This LoRA aims to reproduce Lamb's iconic design with high fidelity:

• Signature Masks: Reliably generates both the lamb mask she wears and the wolf mask that represents her partner, Wolf.

• Ethereal Presence: Captures her otherworldly and ghostly nature, often complemented by a soft glow.

• Core Anatomy: Consistently renders her white hair, lamb-like ears, and hooved feet.

• Weapon of Choice: The LoRA is proficient at generating Lamb with her spirit bow in various poses.

The overall artistic style is illustrative and clean, blending well with popular anime-style SDXL models to create vibrant and high-quality images of the Eternal Hunter.

--- Training Details ---

This LoRA was trained using a custom TOML configuration with the following key parameters:

• Base Model: NoobAI Vpred 1.0

• Resolution: 1024x1024 (with bucketing enabled from 256 to 4096px)

• Network Rank (Dimension): 64

• Network Alpha: 32

  • Note: A higher dimension than alpha was chosen intentionally. This scales down the weight adjustments, which can help in learning finer details and preventing the LoRA from becoming "over-baked."

• UNet & Text Encoder Training: Both were trained (network_train_unet_only = false) for comprehensive character concept learning.

• Optimizer: AdamW8bit

• Learning Rate: 2e-4 (2×10−4)

• LR Scheduler: cosine_with_restarts (3 cycles)

• Advanced Noise Scheduling: min_snr_gamma = 5 and zero_terminal_snr = true were used to improve image composition and training stability.

• Precision: Full fp16 for efficient training.

• Batch Size: 1

• Max Training Epochs: 30