Description:

A rich, high-gloss render. Everything gets shiny highlights, saturated colour and dense detail, and backgrounds turn into warm, ornate interiors. It looks like heavily rendered gacha or splash art.

Known quirk: it can duplicate characters in some compositions (one preview came out with three copies of the same girl). Lower the strength or prompt solo firmly if that happens.

This isn't a trained LoRA. No images were used to make it. It comes from steering neurons inside the base model:

• I mapped the UNet's feed-forward neurons into groups. This file is four groups of about 1,000 neurons each, across four layers: two pushed up and two pushed down, mixed by a random combination search.

• The LoRA stores that push as a bias change, which is why the file is only about 23 KB.

How to use

• No trigger word. Just load the LoRA; it acts on the model only, not the text encoder.

• Strength 1 is the look in the previews. The effect scales linearly with strength: lower it for a lighter touch; negative values push the other way.

• It stacks with other LoRAs.

• Preview settings: 768×1024, 20 steps, CFG 4.5, DPM++ 2M Karras.

Compatibility

• Made and tested in ComfyUI (Load LoRA / LoraLoaderModelOnly) on prefectIllustriousXL v4.0.

• Other UIs: the file holds only bias changes, which some UIs (A1111 / Forge) may not load. If nothing changes, that's why.

• Other Illustrious checkpoints: untested, but they share the base, so it may carry over.