A high-fidelity character LoRA trained on the Anima DiT architecture, designed with fully decoupled wardrobes and robust outfit flexibility across diverse settings.

💡 Core Features

Decoupled Identity: Trained on an animated multi-aspect ratio dataset optimized for Liko.
Architecture: Optimized natively for Anima DiT checkpoints with clean, high-frequency facial geometry and versatile multi-aspect ratio support (incorporating portrait, square, and landscape buckets).
Modular Compatibility: Easily pairs with external clothing, lighting, and texture adapters (such as Cinematic Film Grain) without losing character likeness.

⚙️ Recommended Generation Settings

LoRA Weight: 0.75 – 0.85 (Recommended baseline: 0.80)
Trigger Word: liko (pokemon)
Resolution: 1024x1024 (Square), 832x1216 (Portrait), or 1216x832 (Landscape)
Sampling Steps: 30 – 32
CFG Scale: 4.0 – 4.5
Sampler: Euler or Euler a

👗 Classic Outfit Prompt Tags

liko (pokemon), two-tone hair, multicolored hair, black hair, blue hair two-side up, hair ornament, green hairclip, blue hoodie, white collared shirt, dark skirt, sneakers, school bag

🚫 Universal Negative Prompt

(worst quality, low quality:1.4), score_1, score_2, score_3, photorealistic, realistic, bad anatomy, bad hands, missing fingers, extra digits, blurry, cropped, frame, watermark, signature

✨ Universal Positive Base Tag Anchor

masterpiece, best quality, solo, 1girl,

📊 Training Details

Base Architecture: Anima DiT v1.0 (Latent Diffusion Transformer)
Compute Infrastructure: Modal Cloud A100 (80GB)
Dataset Size: Multi-aspect ratio animated frames with variable bucket integration
Network Dimensions: Rank 32 / Alpha 16
Optimization Parameters: AdamW-8bit, bfloat16 mixed precision, Learning Rate: 0.00005, 1,500 Total Steps

Liko - Pokémon [Anima DiT]

Descrizione