# IDK Aquatic Nature Background V2

## About

LoRA focused on the 'Aquatic Nature Background' concept for Illustrious-based SDXL models. It strengthens generation of underwater. Trained on a curated, normalized dataset with per-image loss curriculum.

PT-BR: LoRA focada no conceito 'Aquatic Nature Background' para modelos SDXL baseados em Illustrious. Reforca a geracao de underwater. Treinada em dataset curado e normalizado com curriculo por loss por imagem.

## Trigger words

### Primary triggers

```

underwater

```

### Supporting tags

Frequent thematic tags from the training set - combine them with the primary triggers to steer the concept:

```

bubble, air bubble, long hair, short hair, fish, blue eyes, barefoot, water, black hair, blue hair, twintails, nude, swimming, swimsuit

```

## Example prompts

```

masterpiece, best quality, absurdres, beautiful woman standing in serene underwater coral reef, wearing flowing white dress, gentle sunlight rays, peaceful expression, <lora:IDK_aquatic-nature-background_V2:0.9>

```

```

masterpiece, best quality, absurdres, beautiful woman floating in mystical bioluminescent ocean cave, completely nude with water droplets on skin, dramatic volumetric god rays, seductive gaze, <lora:IDK_aquatic-nature-background_V2:0.9>

```

## Preview generation settings

| Setting | Value |

| --- | --- |

| Checkpoint | IDN_Illustrious_V10_T |

| Size | 1280x720 + hires 1.5x RealESRGAN 4x+) = 1920x1080 |

| Sampler | Euler a |

| Steps | 28 (+14 hires) |

| CFG scale | 5.5 |

| Denoising strength | 0.4 |

| Seeds | 1000-1009 |

Negative prompt: worst quality, low quality, bad anatomy, bad hands, extra digits, fewer digits, missing fingers, jpeg artifacts, signature, watermark, username, blurry, lowres, bad proportions, deformed, mutated

## Recommended weight

Use 0.8-1.1. This file is a V1+V2 merge (rank concatenation), which slightly dilutes each source - it tolerates and often benefits from weights a bit above 1.0.

## Training details

- Base model: IDN_Illustrious_V10_T

- Images: 61 (approved: 59)

- Resolution: 512px

- Network rank: 32 (alpha 32)

- Loss threshold: 0.0523 (auto)

- Passes: up to 10 (255 steps, stop: max_passes)

- Method: curriculum requeue by per-image loss + dataset normalized to 262k px via RealESRGAN

## Changelog V2

- New multi-source dataset (JoyTag captioning for photos, full captions).

- Upscaled without cropping (dataset normalized via RealESRGAN).

- Trained on IDN_Illustrious_V10_T with loss-based curriculum requeue.

- Merged with the V1 LoRA (rank concatenation) to keep V1 behavior.

File: IDK_aquatic-nature-background_V2.safetensors