RASHID778/king2-image-dataset

KING2-IMAGE Dataset

19,998 synthetic scenes used to train RASHID778/king2-image — a LoRA adapter for SDXL 1.0 (UNet-only, text encoders frozen).

Continuation dataset. The V2 run fine-tunes on top of the original V1 LoRA, which is archived here under king2-original/ for reproducibility.

---

📊 Contents

Split	Images	Resolution	Size
indoor/	9,999	416 × 416	285.1 MB
outdoor/	9,999	150 × 150	65.3 MB
Total	19,998	—	≈ 350 MB

Files are named 00000.jpg … 09998.jpg in each folder.

📁 Structure

indoor/
  00000.jpg
  00001.jpg
  ...
outdoor/
  00000.jpg
  ...
metadata.jsonl       # one row per image: file_name, text
images/
  images.tar.gz      # single-archive mirror of the raw set
king2-original/      # archive of the pre-V2 LoRA weights
scripts/             # training, captioning and verification scripts

🗂 Loading

With datasets — folder-per-class layout, ready for ImageFolder:

from datasets import load_dataset

ds = load_dataset("RASHID778/king2-image-dataset", split="train")
print(ds)   # 19,998 examples · Image · text

Or build the split explicitly:

from datasets import load_dataset

ds = load_dataset(
    "RASHID778/king2-image-dataset",
    data_dir="",
    split="train",
)

With metadata.jsonl — captions live in the text column:

import json
from PIL import Image

rows = [json.loads(l) for l in open("metadata.jsonl", encoding="utf-8")]
row = rows[0]

img = Image.open(row["file_name"])       # e.g. indoor/00000.jpg
caption = row["text"]

⚠️ metadata.jsonl is currently empty

The file is present and correctly structured — 19,998 rows, one per image, with valid file_name values — but every text field is an empty string.

The captioning step has not been run yet. Before training on captions, fill it with scripts/caption.py, which calls Microsoft Florence-2 (microsoft/Florence-2-base, DETAILED_CAPTION) and writes back into the text column.

Until then, treat this dataset as images only — a plain class-conditional or unconditional fine-tune set. Do not pair it with caption-derived prompts expecting signal.

✅ Data notes

Property	Value
Format	JPEG
Colour	RGB
Indoor	9,999 / 9,999 exactly 416 × 416
Outdoor	9,965 / 9,999 exactly 150 × 150
Outdoor outliers	34 images between 150 × 111 and 150 × 150
Duplicate filenames	none — names are unique within each split

The 34 outdoor images with non-square dimensions are minor rounding artefacts from the synthesis step. They are harmless for training but will not survive a strict resize=(416, 416) without a small aspect-ratio change.

The aspect ratio between splits differs by design — indoor scenes are high-detail wide compositions at 416 px, while outdoor scenes are small 150 px tiles. If you concatenate the splits, apply the same resize policy to both or the small outdoor tiles will be upsampled noisily.

🎓 Usage

from datasets import load_dataset
from PIL import Image

ds = load_dataset("RASHID778/king2-image-dataset", split="train")

def resize(ex):
    ex["image"] = ex["image"].convert("RGB").resize((512, 512))
    return ex

ds = ds.map(resize)

Full training recipe lives in scripts/train_continue.py. Training used SDXL 1.0 base, LoRA rank 16, AdamW with cosine schedule and warm-up, bf16 mixed precision, batch size 1 with gradient accumulation 4.

📜 Provenance & licence

Images are synthetic, generated during this project — not scraped or collected from third-party sources. The source prompts and licence matrix used to build the dataset are tracked in the RASHID778/king2-image model repo.

license: other reflects the upstream SDXL 1.0 base model licence, which governs any LoRA trained from it. Review the SDXL licence before commercial use.

📌 Citation

@dataset{king2_image_dataset,
  author    = {RASHID778},
  title     = {KING2-IMAGE Dataset},
  year      = {2026},
  url       = {https://huggingface.co/datasets/RASHID778/king2-image-dataset}
}