Model description

pc98gal_style-v0.2 (Anima) — Changelog

What's new

• Trained on 87 PC-98 galgame images (up from 76 in earlier attempts — 11 new images from `harenochi`, `poisonneedle`, `ram`, `tsuioku`)
• All captions regenerated via ToriiGate-0.5 captioning script (`scripts/batch_caption_torii.py`) — natural-language descriptions prefixed with `pc98gal_style` trigger token
• 5 checkpoints available as separate safetensors (`-000002`, `-000004`, `-000006`, `-000008`, `-000010`) + final

Training config

• Base model: Anima base-v1.0 (DiT + Qwen3-0.6B + flow matching)
• LoRA rank: 32, alpha 32
• Optimizer: AdamW, lr 5e-5, weight_decay 0.1, betas [0.9, 0.99]
• Scheduler: cosine_with_restarts, 100 warmup steps
• Training: 87 images × 3 repeats × 10 epochs = 2610 steps, batch size 1
• Precision: bf16 throughout, gradient checkpointing on
• Resolution: 768×768
• Caption dropout: 10%
• No flip augmentation — PC-98 linework is directional
• Duration: ~2h17m on Radeon 8060S (Strix Halo, gfx1151, ROCm 7.13.99)

Results

• Epochs: 10
• Total steps: 2610
• Final loss: 0.123 (started ~0.13, converged smoothly)

Trigger token

Trigger words

Download model