<h1 id="pc-98-galgame-style-(anima)-v0.2">PC-98 Galgame Style (Anima) v0.2</h1>Trained words: <code>pc98gal_style</code>Base model: Anima v1.0<hr /><h2 id="description">Description</h2>A LoRA trained to reproduce the aesthetic of NEC PC-98 galgame artwork from the early-to-mid 1990s. Captures the signature look of era-defining visual novel and eroge titles — flat cel shading, thin linework, muted limited color palettes, detailed retro anime eyes with characteristic highlights, atmospheric painted backgrounds, and the 16-color era palette feel.Trained on 87 images sourced from 17 authentic PC-98 titles including Words Worth, Nocturnal Illusion, Doukyusei 2, Yuno, Viper GTS, Magical Jam, Dokidoki Vacation, Thirsty, Sand Rock, Saint Dairy, Harenochi, Poison Needle, Ram, Tsuioku, and others, covering a wide range of character designs, poses, and settings.<h2 id="training-details">Training Details</h2><ul><li>Epochs: 10</li><li>Total steps: 2610</li><li>Final loss: 0.123</li><li>LoRA rank: 32 (alpha 32)</li><li>Resolution: 768×768</li><li>Training time: ~2h17m on Strix Halo (Radeon 8060S, ROCm 7.13)</li></ul><h2 id="recommended-settings">Recommended Settings</h2><ul><li>LoRA strength: 0.7–1.0</li><li>Steps: 25–30</li><li>CFG: 4.0–5.0</li><li>Sampler: er-sde, scheduler: simple</li></ul><h2 id="usage">Usage</h2>Place <code>pc98gal_style</code> at the start of your prompt, followed by your subject description.<h3 id="positive-prompt-base">Positive prompt base</h3><pre><code>pc98gal_style, 1990s anime aesthetic, limited color palette, flat cel shading, thin linework, detailed retro eyes, dithering, muted colors, [your subject here]
</code></pre><h3 id="negative-prompt">Negative prompt</h3><pre><code>3d, realistic, photorealistic, modern anime style, thick outlines, gradient mesh shading, oversaturated, blurry, deformed, extra digits, watermark, text, signature
</code></pre><h3 id="example-prompts">Example prompts</h3>Character portrait (1:1, 768×768):<pre><code>pc98gal_style, young woman with long red hair, green eyes, sailor uniform, sailor collar, red ribbon, school rooftop background, cherry blossoms, soft smile, looking at viewer, upper body portrait
</code></pre>Negative: <code>modern, 3d, realistic, photograph, blurry, low quality, deformed</code>Full body schoolgirl (2:3, 512×768):<pre><code>pc98gal_style, schoolgirl with short blue hair, twin tails, blue sailor dress, white socks, brown loafers, standing pose, hands clasped, classroom interior, wooden desks, chalkboard, soft afternoon light, full body
</code></pre>Negative: <code>modern, 3d, realistic, photograph, blurry, low quality</code>Wide scene with multiple characters (16:9, 1024×576):<pre><code>pc98gal_style, two girls sitting on a bench under cherry blossom trees, petals falling, one with pink hair one with black hair, school uniforms, blue sky, clouds, school building in background, peaceful afternoon, wide shot
</code></pre>Negative: <code>modern, 3d, realistic, photograph, blurry, low quality</code>Battle scene with magic effects:<pre><code>pc98gal_style, dynamic battle scene, young female mage with long teal hair in high ponytail, white and gold magical armor, wielding a glowing blue staff casting lightning spell, energy arcs and sparks around her, dark dramatic sky background, motion blur, action pose mid-cast, fantasy RPG battle scene
</code></pre>Negative: <code>modern, 3d render, realistic, photograph, blurry, low quality, deformed, text, watermark</code>Mecha pilot in hangar:<pre><code>pc98gal_style, mecha anime hangar launch scene, female pilot with short silver hair in orange flight suit climbing into cockpit of giant humanoid battle mech, mech is 80 meter tall white and blue armored robot with glowing green visor and massive shoulder-mounted rail cannons, steam venting from hydraulic systems, hangar bay opening to reveal orange sunset sky, ground crew in yellow helmets waving her off, 80s and 90s mecha anime aesthetic, Macross inspired
</code></pre>Negative: <code>modern, 3d render, realistic, photograph, blurry, low quality, deformed, text, watermark, gunpla</code><h2 id="tips">Tips</h2><ul><li>Trigger token must be the first word in your positive prompt</li><li>Higher LoRA strength (0.9-1.0) preserves the PC-98 aesthetic more strongly; lower (0.6-0.8) gives more freedom to your subject</li><li>The model works best with descriptive natural-language prompts rather than tag-style prompts</li><li>PC-98 linework is directional — do not use img2img with horizontal flips</li><li>For best results, pair with a sampler that respects the era’s flat cel shading (er-sde or similar)</li></ul><h2 id="notes">Notes</h2>This is a personal project to preserve and reimagine the distinctive aesthetic of NEC PC-98 era visual novels. The model captures both the technical limitations of the platform (16-color palette, dithering, scanlines) and the artistic sensibilities of the period. Still actively iterating — v0.3 will include more titles and refined training parameters.

PC-98 Gal

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

Files with SHA256: 5d37e2a326526ac6fb3d10087cfd140deba2bf469de4cd2bc55a888961fb17e8