Risposta/MC_Gen

MC-Gen 2.0 — training datasets

Training-ready datasets for the MC-Gen 2.0 project: a small pixel-space Rectified-Flow Transformer that generates Minecraft / voxel-game style low-resolution textures (32x32, RGBA).

The source code lives on GitHub: https://github.com/jurky123/MC_Gen2.0

This repository only contains the preprocessed, training-ready datasets. Raw downloads (itch.io, OpenGameArt, Kaggle, Modrinth, Hugging Face mirrors) are intentionally not included.

License: research / study only. The datasets aggregate third-party game assets with heterogeneous licenses. They are published here for academic and personal study only. Do not use them commercially, and keep the provenance fields in the metadata for attribution. See Provenance below.

Contents

Path	Rows	Shape / dtype	Split (train/val/test)	Purpose
data/build/stage1_32_rgba	3,185,719	uint8[32,32,4] RGBA	3,004,255 / 91,566 / 89,898	Stage 1 pixel + all-MC pretraining
data/build/mc_text2image32	1,034,057	uint8[32,32,3] RGB	941,104 / 43,184 / 49,769	Minecraft domain (text+image), legacy RGB build
data/build/mc_text2image32_wl	1,034,057	uint8[32,32,4] RGBA	941,104 / 43,184 / 49,769	Stage 1/2 main MC set: RGBA + weak/fine prompts + annotations + frozen-text embeddings
data/build/stage2_32	1,044,875	uint8[32,32,4] RGBA	949,165 / 44,417 / 51,293	Stage 2 unified texture set (MC-wl + Modrinth) + pooled text embeddings
data/processed/minecraft_16x_finetune32	1,498	32x32 tiles	manifest only	Stage 3 curated fine-tune seeds
data/processed/modrinth32	10,818	32x32 tiles	manifest only	Modrinth MC textures (deduped)

stage1_32_rgba source mix

source_set	rows
itch.io free pixel art	1,859,029
NathMen12 MC text-to-image	1,034,057
alucard pixel sprites	282,044
OpenGameArt (OGA-BY 3.0)	4,555
Kenney pixel assets	4,312
Kaggle pixel art	1,722

Minecraft rows are fully opaque; pixel-art rows keep their original alpha channel (transparent backgrounds are not composited onto black).

Extra files in mc_text2image32_wl / stage2_32

file	meaning
text_embeddings.f32.mmap	frozen Qwen3-8B text features, FP32; mc_text2image32_wl = 4 prompt views × 4096-d per row, stage2_32 = 1 pooled 4096-d token per row. Note: embeddings are currently not in this repo (too large to push); recompute with scripts/precompute_text.py in the project repo.
prompt_views.* / grounded_prompts.* / filename_prompts.parquet	prompt variants (K-view, label-grounded rewrite, filename weak prompts)
coarse_annotations.shard*.jsonl	local-27B VLM captions (coarse profile)
annotation_manifest.jsonl / high_quality_manifest.jsonl	annotation lineage (stage2_32)
replay_splits*.json	anti-forgetting replay index sets

metadata.parquet columns for the unified datasets (stage1_32_rgba, stage2_32):

column	meaning
index	row index into the mmap
source_set	originating processed set (see table above)
weak_prompt	short text caption (MC filename/label or generic pixel-art)
source_metadata	JSON blob with original filename, project, license, etc.

mc_text2image32_wl/metadata.parquet additionally carries Modrinth provenance (file_name, project_id, mod_slug, author, license, version_url, type, tileable) and the text / weak_prompt caption pair.

The processed *32 tile folders instead contain tiles/<sha>.png plus a manifest.jsonl and summary.json.

File format

Each mmap dataset directory contains at least:

<name>/
├── images.uint8.mmap   # raw little-endian uint8, C-order (N, H, W, C)
├── metadata.parquet    # one row per image + provenance/prompt fields
├── splits.json         # {"train": [idx...], "val": [...], "test": [...]}
└── build_summary.json  # counts and source breakdown (where present)

images.uint8.mmap is a headerless NumPy array. Load it with:

import numpy as np
n, size, channels = 3_185_719, 32, 4
imgs = np.memmap(
    "data/build/stage1_32_rgba/images.uint8.mmap",
    dtype=np.uint8, mode="r", shape=(n, size, size, channels),
)

Download

pip install -U "huggingface_hub[hf_xet]"

# only the pretraining set
hf download Risposta/MC_Gen \
  --repo-type dataset \
  --include "data/build/stage1_32_rgba/*" \
  --local-dir .

# everything (all datasets)
hf download Risposta/MC_Gen --repo-type dataset --local-dir .

With --local-dir . run from the project root, the files land exactly at the paths the training configs expect. A convenience wrapper is provided in the project repo as scripts/download_dataset_hf.py.

Provenance

The datasets derive from the following sources (see source_metadata for per-sample details): NathMen12/16xModdedMinecraft-TextToImage, itch.io free pixel-art asset packs, Kenney CC0 packs, evilsocket/alucard-sprites, Kaggle ebrahimelgazar/pixel-art, nyuuzyou/OpenGameArt-OGA-BY-3.0, James-A/Minecraft-16x-Dataset, and Modrinth resource packs.

This repository is not affiliated with, endorsed by, or associated with Microsoft, Mojang, or any asset author. All trademarks belong to their respective owners.

Citation

@misc{mcgen2,
  title  = {MC-Gen 2.0: pixel-space flow transformer for Minecraft textures},
  author = {Risposta},
  year   = {2026},
  url    = {https://github.com/jurky123/MC_Gen2.0}
}