README.md

metadata

license: cc-by-4.0
pretty_name: Pixel-Art Normal Maps
task_categories:
  - image-to-image
tags:
  - pixel-art
  - normal-maps
  - normal-map-estimation
  - synthetic
  - computer-graphics
  - game-assets
size_categories:
  - 10K<n<100K
dataset_info:
  features:
    - name: color
      dtype: image
    - name: normal
      dtype: image
    - name: size
      dtype: int32
    - name: model_id
      dtype: string
    - name: filename
      dtype: string
  splits:
    - name: train
      num_bytes: 89583790
      num_examples: 14497
  download_size: 90860770
  dataset_size: 89583790
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Pixel-Art Normal Maps

Synthetic (pixel-art sprite → camera-space normal map) image pairs for training and evaluating normal-map generation models for 2D pixel-art.

This dataset accompanies a master's thesis on normal-map generation for pixelated images. Code: https://github.com/Dmiktor/pixel-art-normal-maps

Load

from datasets import load_dataset

ds = load_dataset("<HF_USERNAME>/pixel-art-normal-maps", split="train")
ex = ds[0]
ex["color"]    # PIL.Image - RGBA pixel-art sprite (model INPUT)
ex["normal"]   # PIL.Image - camera-space normal map (TARGET)
ex["size"], ex["model_id"], ex["filename"]

Columns

Column	Type	Description
color	image (RGBA)	pixel-art sprite, transparent background — model input
normal	image (RGB)	camera-space normal map, N_rgb = (N + 1) / 2 · 255 — target
size	int	render size in px (32, 48, 64, 80 or 96)
model_id	string	id of the source 3D model the pair was rendered from
filename	string	original render name (<id>_<model>_r<rot>_e<elev>_s<size>[...])

One train split. Stored as Parquet with the image bytes embedded (loads anywhere, no external files).

Property	Value
Paired images	14,497
Unique source models	648
Sprite sizes	32, 48, 64, 80, 96 px
Colour format	RGBA, transparent background
Normal encoding	camera space, N_rgb = (N + 1) / 2 · 255
Suggested val split	10 % (seed 42) → 1,449 images

How it was made

Low-poly 3D models indexed by Objaverse++ are filtered with CLIP to keep clean single-character meshes, then rendered twice in Blender under identical camera and pose: a stylised pixel-art colour pass and a matching camera-space normal pass. Identical framing means the two images align pixel-for-pixel.

Baseline results (masked angular error, validation = 1,449 images)

Method	Mean angular error (°)
Sobel (luminance)	49.70
Sobel (alpha)	44.26
Beveling (EDT)	35.28
U-Net (this work)	25.62

Citation

@misc{shapovalov2026pixelnormalmaps,
  author = {Shapovalov, Dmytro},
  title  = {Pixel-Art Normal Maps: a synthetic dataset for normal-map generation},
  year   = {2026},
  howpublished = {Hugging Face Hub},
  note   = {Derived from Objaverse++ / Objaverse assets}
}

License & attribution

Released under CC BY 4.0. Rendered from Objaverse++ / Objaverse 3D assets, which carry their own licenses — please attribute Objaverse++ and respect upstream model licenses.