CoolFace
Datasetpublic

Saminx22/game-character-sprites-golden

game-character-sprites-golden A captioned dataset of small pixel-art game character sprites (64x64 to 128x128), intended for pretraining a small image generation model that produces game-style sprites from text prompts. Dataset Summary Total samples: 4053 Train samples: 3445 Validation samples: 608 Minimum image resolution: 64x64 Caption model: Qwen/Qwen2.5-VL-7B-Instruct Pixel-art mode: True (blur and watermark filters disabled, native low-res preserved)… See the full description on the dataset page: https://huggingface.co/datasets/Saminx22/game-character-sprites-golden.

sourceHugging Faceunknownupdated 5mo agoView on Hugging Face
1likes68downloads
Dataset Card

game-character-sprites-golden

A captioned dataset of small pixel-art game character sprites (64x64 to 128x128), intended for pretraining a small image generation model that produces game-style sprites from text prompts.

Dataset Summary

  • —Total samples: 4053
  • —Train samples: 3445
  • —Validation samples: 608
  • —Minimum image resolution: 64x64
  • —Caption model: Qwen/Qwen2.5-VL-7B-Instruct
  • —Pixel-art mode: True (blur and watermark filters disabled, native low-res preserved)

Schema

  • —image: image object (PNG bytes, native sprite resolution)
  • —caption: detailed descriptive caption focused on archetype, pose, equipment, palette, pixel-art style
  • —short_caption: compact caption for auxiliary usage
  • —style: style label (typically pixel_art)
  • —category: semantic class (character, enemy, boss, npc)
  • —tags: JSON-encoded tags list
  • —license: source license label
  • —source: source dataset id/name (HF or Kaggle)
  • —width, height, phash: native dimensions and perceptual hash

Source Datasets

  • —ayhantasyurt/pixel-art-2dgame-charecter-sprites-idle
  • —calmness/retro-pixel-characters-generator
  • —evilsocket/alucard-sprites

Filtering Pipeline (pixel-art tuned)

  • —Resolution gate: >= 64x64 (preserves 64x64 retro sprites)
  • —Entropy floor: >= 0.5 (very permissive; pixel art has low entropy by design)
  • —Exact-duplicate removal via perceptual hash (Hamming distance <= 0); animation frame variants are intentionally KEPT as separate samples
  • —Animated GIFs are split into individual frames before captioning
  • —Blur/watermark/relevance heuristics disabled (they misfire on small pixel art)

Captioning

Tiny sprites are upscaled with nearest-neighbor to >= 256 px before being shown to the VLM, so it can read pixel-level features clearly. The VLM is prompted to describe archetype, pose, equipment, palette, and pixel-art style.

Intended Use

This dataset is intended for pretraining a small (10M-200M parameter) DiT or U-Net to generate pixel-art game character sprites from text prompts.

Notes

Please validate licensing constraints per source before commercial deployment.