Description:This LoRA is dedicated to recreating the vibrant and heroic Star Guardian skin for Poppy from League of Legends.This is a UNet-only LoRA (network_train_unet_only = true), meaning it was trained exclusively on the visual aspects of the character's design. It does not teach the model a new "concept" of Poppy but instead excels at applying her Star Guardian outfit, hairstyle, and signature hammer over a well-prompted character.For optimal compatibility, this LoRA was trained on NoobAI XL Vpred 1.0 and works best with v-prediction models.Usage InstructionsActivation & PromptingSince the text encoder was not trained, you must describe the character in your prompt. Use the primary trigger word, then add descriptive tags for her appearance and outfit.Primary Trigger: sgpoppyEssential Character Tags: yordle, blue skin, purple hair, twintails, blue eyes, pointy ears, fangEssential Outfit Tags: star guardian uniform, armor, breastplate, skirt, blue gloves, pantyhose, white legwear, kneebootsWeapon Tag: hammerImportant: LoRA WeightThis LoRA was trained with an alpha (64) that is double its dimension (32). This makes the LoRA very strong.Works plenty fine with a weight of 1.0.Example: lora:SGPoppy:1.0Recommended SettingsModel: Any SDXL v-prediction model (e.g., NoobAI-XL, Illustrious-XL).Sampler: Euler a CFG++Steps: 20CFG Scale: 1.4-1.7Outfit & Style DescriptionThis LoRA accurately reproduces the key elements of Poppy's Star Guardian design:The Outfit: Captures her distinct blue, white, and gold magical girl uniform, complete with the star emblem, large bows, and armored gauntlets/boots.Character Features: Reliably generates her yordle physique, vibrant blue skin, and signature purple hair styled in large twintails.The Starlight Hammer: Excellently reproduces her iconic, oversized golden hammer, maintaining its grand scale in proportion to her small stature.The style is sharp and illustrative, perfect for creating high-impact character art that reflects Poppy's determined and heroic personality.Training DetailsThis LoRA was trained on a focused dataset of 29 images with the following notable parameters:Base Model: NoobAI XL Vpred 1.0Resolution: 1024x1024 (with bucketing enabled)Network Rank (Dimension): 32Network Alpha: 64Note: A 2:1 alpha-to-dim ratio was used, amplifying the LoRA's trained effect. This is why lower weights are recommended during inference.Training Target: UNet OnlyOptimizer: AdamW8bitLearning Rate: 4e-4 (4×10−4)LR Scheduler: cosine_with_restarts (5 cycles)Advanced Noise Scheduling: An extremely high min_snr_gamma = 80 was used alongside zero_terminal_snr = true, forcing the model to focus intensely on high-level composition and detail.Max Training Epochs: 50
