rappjaco/Low-Poly-Game-Asset-Images

Low-Poly Game Asset Image Dataset

A synthetic image dataset of low-poly 3D game assets with captions, built for fine-tuning text-to-image models (LoRA / full fine-tune) on the low-poly asset domain.

Structure

dataset/
  images/
    p00001.png          image
    ...
  examples/
    samples_100.png     preview sheet (100 samples) used in this card
  metadata.jsonl        image path + caption + metadata for every sample
  prompts.jsonl         prompt pool used for generation
  prompts_extra.jsonl   extra prompts for the additional tree samples

metadata.jsonl is the single source of truth for the captions: every record carries its own text/prompt, so there are no per-image sidecar files to keep in sync.

No train/test split is included — split it yourself when training.

Fields

Each line of metadata.jsonl:

Field	Description
file_name	Relative path to the image, e.g. images/p00001.png
text	Full caption — identical to prompt. The key most trainers expect
id	Sample id, e.g. p00001
category	Object category, e.g. pistol, tree, chest
subject	Object phrase, e.g. a chunky pistol
view	Camera view, e.g. three-quarter front view
background	Background, e.g. plain light gray
tag	Short label: object name only
prompt	Full prompt used to generate the image (same string as text)

Example record:

{
  "file_name": "images/p00001.png",
  "text": "a chunky pistol, low-poly 3D game asset, flat-shaded faceted polygons with visible sharp edges, ...",
  "id": "p00001",
  "category": "pistol",
  "subject": "a chunky pistol",
  "view": "three-quarter front view",
  "background": "plain light gray",
  "tag": "chunky pistol",
  "prompt": "a chunky pistol, low-poly 3D game asset, flat-shaded faceted polygons with visible sharp edges, ..."
}

Usage

Load it directly. file_name pairs each caption with its image, so the Hub exposes a decoded image column alongside the text:

from datasets import load_dataset

ds = load_dataset("rappjaco/Low-Poly-Game-Asset-Images", split="train")
# Dataset({features: ['file_name', 'text', 'id', 'category', 'subject',
#                     'view', 'background', 'tag', 'prompt', 'image'],
#          num_rows: 1640})

The file_name + text pair is exactly the format used by diffusers, kohya_ss and ai-toolkit, so metadata.jsonl can be fed to those tools as-is.

Example — write a 5% validation split:

import json, random

rows = [json.loads(l) for l in open("metadata.jsonl", encoding="utf-8")]
random.Random(0).shuffle(rows)
n_val = int(len(rows) * 0.05)
val, train = rows[:n_val], rows[n_val:]

for name, part in (("train", train), ("val", val)):
    with open(f"{name}.jsonl", "w", encoding="utf-8") as fh:
        for r in part:
            fh.write(json.dumps({"file_name": r["file_name"], "text": r["text"]}) + "\n")

To train with short labels instead of full prompts, use the tag field in place of text.

Categories

52 categories in total, including firearms (pistol, rifle, ammo), nature (tree, bush, mushroom, rock, cactus, stump), props (chest, crate, barrel, lantern, potion, book, candlestick), weapons (sword, axe, dagger, bow, spear, shield), buildings (house, hut, tower, well, bridge, windmill), creatures (dragon, owl, cat, frog, rabbit, bird) and more.

Counts across all 1640 samples:

total: 1640  (train=1558, val=82)

category distribution:
  ammo: 29
  axe: 18
  barrel: 32
  bird: 34
  boat: 32
  book: 27
  bookcase: 30
  bow: 8
  bridge: 24
  bush: 11
  cactus: 32
  candlestick: 19
  cart: 24
  cat: 26
  cauldron: 31
  chair: 31
  chest: 31
  coins: 28
  crate: 29
  crystal: 38
  dagger: 19
  dragon: 28
  fountain: 32
  frog: 28
  gem: 36
  goblet: 32
  hammer: 22
  house: 35
  hut: 32
  lantern: 27
  mace: 21
  mushroom: 26
  owl: 19
  pistol: 101
  potion: 28
  rabbit: 29
  rifle: 81
  rock: 32
  shield: 31
  skull: 31
  spear: 22
  staff: 21
  stump: 30
  sword: 21
  table: 34
  torch: 32
  totem: 28
  tower: 34
  tree: 112
  wand: 28
  well: 31
  windmill: 23

Fork notes

This fork changes the original raw-file layout into a dataset that loads directly:

• metadata.jsonl: file renamed to file_name, and a text column added (equal to prompt).
• The per-image sidecar captions (images/*.txt, images/*.tag.txt) and report.txt were removed. Their contents are preserved in metadata.jsonl (text/prompt and tag), and the category counts from report.txt are inlined above.
• The caption and tag_file fields were dropped, since the files they pointed to no longer exist. Their content is unchanged and available via text/prompt and tag.

This matters because the Hub picks a dataset builder by the most common file extension in the repo. With 3281 .txt files against 1640 .png, the .txt files won, so the repo was served as a text dataset: one row per line of every .txt file, 3337 rows, no images.

The upstream dataset's raw sidecar layout remains available at laym0nd/Low-Poly-Game-Asset-Images.

License

• Dataset: CC0-1.0

The generation code is not included in this folder; it lives in the GitHub repository of this project and is released under the MIT license.