✏️ Dujinshi — sketch style (Anima) 📜

A loose, sketchy doujinshi look: visible construction lines, ink strokes that overshoot, soft muted colour underneath. It keeps the energy of a rough page rather than a finished illustration - linework stays alive instead of being cleaned away.

Trained at 768px for Anima with full text-encoder training.

⚙️ Usage

Trigger: d0ujinshi

Strength: 0.9 - 1.1

Base: Anima Base v1.0 / waiANIMA

Steps: 25–30

CFG: 4–5

Sampler: er_sde / simple

Put d0ujinshi first, then describe whatever you want.

🎯 What it does

Trained as a style — every subject, pose and background in the training data was tagged separately so the trigger carries only the rendering. It holds on subjects absent from the training set, including mostly faceless male characters and full environment shots, where the sketchy ink treatment carries through onto architecture and scenery.

This is a colour style, not monochrome manga. There's no screentone or halftone here — the look is muted colour under loose ink.

💡 Prompt tips

Recommended negative:

censored, mosaic censoring, english text, speech bubble, worst quality, low quality, blurry, watermark, signature

Censoring is fully steerable. A large share of the source material carried mosaic censoring, and it was deliberately tagged during training instead of being filtered out. So it behaves as a control rather than as something baked in:

• mention it nowhere → no censoring appears (sometimes it might..)

• add censored, mosaic censoring to the positive → it comes back

• put it in the negative → guaranteed clean

Good pairings: simple background, white background, soft lighting.

🔧 Troubleshooting

Lines are too rough / too messy. Drop to ~0.9 strength.

Looks too clean, not sketchy enough. Raise to 1.0-1.1 and remove detailed, intricate, masterpiece-style quality stacking — polish tags fight this style directly.

Censoring shows up unrequested. Add censored, mosaic censoring to the negative. It shouldn't appear without being asked for, but the negative makes it certain.

📝 Notes

Style-descriptor tags were stripped from the training captions so the trigger owns the look, while artifacts like censoring and text were kept as explicit tags so they remain under your control. The result is a style LoRA with switches rather than baked-in surprises.

Feedback welcome.