💬 Discord — join here · ☕ Ko-fi — support the factory

Eye Detailer 眼部细节

Better eyes. No trigger, no style, no character lock.

A trigger-less detail LoRA that rebuilds how Anima draws eyes. Irises get internal construction — fibres, rings, layered depth into the pupil — instead of a flat two-tone disc; catchlights become shaped, multi-source reflections instead of one white dot; lashes render as separated weighted strands with the lid shadow actually falling across the iris. It carries no character, no outfit and no art style: drop it on any prompt, any checkpoint character, and only the eyes change. It also does not tint. Every eye colour you prompt stays the colour you prompted — the gallery runs crimson, blue, violet, amber, emerald, heterochromia and aqua through the same LoRA at the same weight, and each one keeps its own hue and simply gains structure.

🧩 Version

• 🔵 Anima — for Anima / Qwen-Image DiT checkpoints. No trigger word. Weight 0.5–1.0.

📋 How to use

Load it, set the weight, prompt normally. There is nothing to type.

• 0.5–0.7 — subtle pass; keeps the checkpoint's own eye character, adds construction.

• 0.8–1.0 — full effect; the look in the gallery shots.

🎯 Framing is the whole game

Portrait, bust and upper-body shots get the full effect natively — the eyes occupy enough pixels for the detail to be drawn. On full-body and wide shots the native effect is weak, and that is physics rather than a defect: at that distance an eye is a few dozen pixels and there is no room to put fibres in.

The fix is the one you already use — put it in your FaceDetailer / ADetailer pass. The detailer re-renders the face region at full resolution with the LoRA active, so distant eyes come back with the same construction as a close-up. The full-body shot in the gallery is exactly that: one detailer pass, weight 1.0.

Honest limitation: at full-body / wide framing without a detailer pass, expect little or no change. Tiny eyes need the detailer pass — everything from bust up does not.

💡 Tips

• Stacks cleanly with character LoRAs — it has no identity of its own, so it will not fight a face. If a character's eye shape starts drifting, drop to 0.6.

• Keep it out of the way of style LoRAs at 1.0 if the style is deliberately flat or cel-shaded — 0.5 gives detail without arguing with the finish.

• Inside FaceDetailer, 1.0 is the useful setting; the pass is already localised, so there is nothing to hold back.

• No quality tags, no eye tags and no detailed eyes are required. They do not hurt, but the LoRA does the work without them.

🧪 Trained on

50 hand-picked images, judged one by one on iris construction, catchlight craft and lash/lid linework — not on how big the eyes were in frame. Small and elite on purpose: a tight, quality-first pool gives a clean detail direction with no signature motif bleeding into your renders.

More of my LoRAs: Techannel on Civitai — likes & Buzz keep the factory running ❤️