SPECSIA-15K

SPECSIA: Stylization Dataset for Novel-View Enhancement in Drawing-based 3D Animation
Kyuwon Kim, Sunjae Yoon, Chang D. Yoo (KAIST / Chung-Ang University)
European Conference on Computer Vision (ECCV) 2026

[Project Page] · [Paper] · [Code]

---

Dataset Description

SPECSIA-15K is a large-scale paired dataset for learning to correct projection artifacts in drawing-based 3D character animation.

Drawing-based 3D pipelines (e.g. DrawingSpinUp) lift a single 2D cartoon into an animatable 3D character. When the 3D mesh is re-projected to 2D from unseen viewpoints, the rendered image carries visible artifacts — texture degradation, missing contours, mesh noise. SPECSIA-15K provides matched (artifact-prone input, artifact-free GT) pairs so a refinement network can be pre-trained to correct these artifacts in a view-generalizable way.

How the dataset was built

1. Source: 1,498 characters from 3DBiCar (bicartoon 3D characters).
2. GT renders (Y*): Each character's T-pose OBJ is rendered from 10 uniformly-spaced yaw angles (0°–360°) using Blender Cycles with an orthographic camera and Freestyle outline (external contour only; thickness uniformly sampled from {1, 2, 3, 4} pixels, grayscale line color uniformly sampled from [0, 255]).
   → {char}_{view}/char/texture_with_bg.png
3. Reconstructed input (Z): The front-view render is passed through Wonder3D → instant-nsr to obtain a coarse 3D mesh. This mesh is re-rendered from the same 10 yaw angles using Blender Eevee with orthographic projection.
   → {char}_{view}/mesh/blender_render/rest_pose/color/0001.png
4. Positional hint (Z_pos): Normalized world-space XYZ vertex colors rendered from the same viewpoint. Encodes spatial location for location-aware correction.
   → {char}_{view}/mesh/blender_render/rest_pose/pos/0001.png
5. Edge map: Sobel gradient on Z_pos, highlighting geometry boundaries.
   → {char}_{view}/mesh/blender_render/rest_pose/edge/0001.png

---

Dataset Statistics

Split	Characters	Samples
train	1,298	12,980
validation	100	1,000
test	100	1,000
total	1,498	14,980

• Resolution: 512 × 512 pixels
• Format: RGB PNG (color, pos, edge, gt), grayscale PNG (mask)
• Random seed for split: 42 (character-level, no view leakage across splits)

---

Fields

Field	Type	Description
uid	string	"{char_id}_{view_id}" — unique sample identifier
color	Image	Reconstructed 3D re-projection — artifact-prone input Z
pos	Image	Positional hint (world-space XYZ → RGB) — Z_pos
edge	Image	Sobel edge map derived from pos
mask	Image	Foreground silhouette mask (grayscale)
gt	Image	GT Blender Cycles render — artifact-free target Y*

---

Usage

from datasets import load_dataset

ds = load_dataset("Kyu0528/SPECSIA-15K")

sample = ds["train"][0]
sample["color"].show()  # artifact-prone reconstructed render (Z)
sample["gt"].show()     # artifact-free GT render (Y*)

Pre-training DraViE on SPECSIA-15K

Coming soon. DraViE training code and pretrained checkpoints will be released at rbdnjs7830/SPECSIA_release.

---

License

SPECSIA-15K is released under CC BY-NC 4.0.

This dataset contains only rendered image pairs derived from 3DBiCar character models. Raw 3DBiCar 3D assets are NOT included and must be obtained separately from the 3DBiCar repository under their original license.

---

Citation

TBA