kiki_v4_moss_v1

Style LoRA trained on MOSS-Image-Base (bf16, base not SFT) for the "Studio Ghibli style" look from Kiki's Delivery Service (dataset: Kiki v4, 56 images). Style anchor/trigger: Studio Ghibli style (appended to every training caption).

• Network: LoKr full-rank, linear 32 / alpha 32, factor 4
• Recipe: style (high-noise bias), sigmoid timesteps, automagic3, lr 1e-4 (min 3e-5), EMA 0.98
• Training: 2000 steps, batch 2, mixed 512/768/1024 buckets
• Checkpoints every 250 steps; sweet spot expected 750-1500
• Final checkpoint: kiki_moss_v1.safetensors (2000 steps)

Note: a different, earlier kiki run (LoKr factor 8, int8 SFT base, 4750 steps) lives in Yofardev/kiki_moss_v1. This repo is the v4 dataset retrain on Base with the style recipe.