Nanami Jinnai - LoRA Training Report (Illustrious-XL)

💻 Hardware & Environment

• GPU: NVIDIA GeForce RTX 5060 Ti (16GB VRAM)

• Training Suite: Kohya_ss GUI (Local deployment)

💿 Base Model & Dataset

• Base Model: Illustrious-XL v0.1 (Checkpoint)

• Dataset Size: 18 high-quality retro anime screenshots (Manually isolated & cropped)

• Aesthetic Style: 1990s vintage anime cel-shading (Pioneer LDC style)

⚙️ Training Parameters

• Dataset Repeats: 33

• Batch Size: 4

• Max Train Epochs: 3

• Total Optimization Steps: 450 (150 steps per epoch)

• Resolution: 1024x1024 (Native Square Buckets)

• Clip Skip: 2

📦 Network Config & Size

• Network Rank (Dimension): 32

• Network Alpha: 16

• File Format: .safetensors

🧠 Solvers & Optimizations

• Optimizer: AdamW8bit (Stable & VRAM efficient)

• Gradient Checkpointing: Enabled (VRAM usage locked at a stable ~10GB)

• Mixed Precision: fp16

🎯 Recommended Generation Settings (ComfyUI / WebUI)

• Best File Version: Epoch 1 (-000001.safetensors) — Offers the cleanest anatomy and absolute anti-cloning stability.

• LoRA Weight: 0.85 – 1.0

• CFG Scale: 5.0 – 5.5

• Sampler: euler_ancestral (Euler A) with normal scheduler

• Recommended Resolution: 896x1152 (Portrait) or 1024x1024 (Square)

🟩 Core Activation Tags (Danbooru Style)

• GTML_Nanami_Jinnai, 1girl, solo, anime screenshot, short hair, brown hair, repicado hair, hairband, black hairband, brown eyes, 1990s anime style, retro anime

• School Uniform Outfit: school uniform, yellow cardigan, unbuttoned cardigan, yellow blazer, collared shirt, white shirt, necktie, red tie, green skirt, short skirt, pleated skirt