Dataset collection and creation is a costly and time-consuming process. We believe that this dataset can contribute to the computer graphics and machine learning community to develop more advanced techniques for creature and fantasy image generation.
This dataset was used to train image-generation models for the Chimera project. We demonstrated results in a blog post (http://goo.gle/chimera-painter-blog) and demo (http://goo.gle/chimera-painter). When we started the project, we did an extensive evaluation of existing datasets to see if one would meet our needs. No existing dataset had the qualities needed to train high-quality machine learning image-generation models for the case of fantasy creatures. While existing photographic datasets exist, they are missing dramatic angles, viewpoints, poses, and textures that are present in this dataset.
This dataset contains 2D images created using textured 3D models in an Unreal Engine scene.
The dataset is divided into the following groups:
- ~721K Training creature PNG images (~143 GB)
- ~721K Training segmentation map PNG images (~3.8 GB),
- 11 Hand-drawn test segmentation map PNG images (80 KB)
- ~2K Automatically generated test segmentation map PNG images (12 MB)
Example images from the dataset are available here.
The entire dataset is available as a single zip file here (~146.9 GB).
- Resolution: 512x512
- Format: PNG
- Filename convention: CreatureName-CR-#-S-#-T-#-CA-# (semantically this is CreatureName-Pose-Scale-Texture-Camera Angle)
All images contain 70 pixel wide borders on the left and right sides, making the effective image area 512x372. The purpose of this was to fit the rectangular border needed for the demo card game that these images were targeted for, while also allowing for a square image for a machine learning model to train on. The red, green, blue (RGB) border color is all uniform in both the creature images and segmentation images at (0, 0, 0).
The filename parts are described in greater detail in the following sections.
The type of creature represented in the image. There are 30 creatures in total. There are 24,120 creature images and 24,120 segmentation images per creature, except for the Antlion, which has only 21,708 of each type of image.
- Antlion
- Axolotl
- CGazelle (Chinese gazelle)
- Cheetah
- Crab
- Dino (Tyrannosaurus Rex)
- Dragonfly
- Fish
- Gorilla
- Grizzlybear
- HBird (Hummingbird)
- HWhale (Humpback whale)
- Hyena
- Insect (Wasp)
- Lion
- Lynx
- Manatee
- Mantis (Praying Mantis)
- Mongoose
- Moth
- PGourmani (Pearl Gourami)
- Porcupine
- PWorm (Parasitic worm)
- RDolphin (River dolphin)
- SeaHorse
- SIbex (Sable Ibex)
- SNBat (Sword-nosed Bat)
- Starling
- Swan
- Wolf
CR-#: A specific pose / creature combination. For example, CR-1 is a Hyena in a "Standing Majestic" pose. While the number is unique to every animal / pose combination, the pose themes repeat on 10 with the following pattern. (For example, 1, 11, 21, 31, etc. would all use the "Standing Majestic" pose.)
- Standing Majestic
- Standing Aggressive
- Standing Defensive
- Sitting Majestic
- Sitting Aggressive
- Sitting Defensive
- Jumping 1
- Jumping 2
- Running 1
- Running 2
S-#: How large the creature appears in the image. There are four scales for each creature, ranging from smallest to largest.
T-#: The texture variant of the creature as applied to the original 3D model that was used to generate this image. There are three textures per creature type.
CA-#: The camera angle that the creature image was captured at.
- Resolution: 512x512
- Format: PNG
- Filename convention: CreatureName-CreatureName_# or CreatureName-CreatureName-CreatureName_#
The image shapes of the test images match the training images.
All segmentation map images (whether training, or evaluation) use a particular color mapping, one for each creature part supported in the dataset. There are 22 distinct colors used, which includes 20 colors for creature parts, one for the ground color, and one color for the background of the image. The red, green, blue (RGB) color mappings to semantic names are as follows:
| Part Name | RGB Color |
|---|---|
| Ground | (255, 255, 255) |
| Background | (0, 0, 0) |
| Head | (230, 25, 75) |
| Muzzle | (230, 190, 255) |
| Eyes | (250, 190, 190) |
| Nose | (70, 240, 240) |
| Mouth | (67, 99, 216) |
| Teeth | (128, 128, 0) |
| Ears | (170, 255, 195) |
| Horns | (117, 0, 0) |
| Neck | (60, 180, 75) |
| Wing | (240, 50, 230) |
| Torso Front | (145, 30, 180) |
| Torso Back | (0, 128, 128) |
| Upper Front Legs | (188, 246, 12) |
| Upper Back Legs | (245, 130, 49) |
| Lower Front Legs | (0, 0, 117) |
| Lower Back Legs | (128, 0, 0) |
| Foot Top | (128, 128, 128) |
| Foot Bottom | (255, 225, 25) |
| Claws | (154, 99, 36) |
| Tail | (255, 216, 177) |
| File Name | Creature | Segmentation map |
|---|---|---|
| CGazelle-CR-171_S_4_T_2_CA_136.png |  |
 |
| Dino-CR-91_S_2_T_2_CA_95.png |  |
 |
| Porcupine-CR-76_S_3_T_0_CA_95.png |  |
 |
Hand-drawn test segmentation maps
| File Name | Segmentation map |
|---|---|
| SNBat-Dino-Wolf_0.png |  |
| Hyena-Gorilla-Porcupine_0.png |  |
| Axolotl-HWhale_0.png |  |
Automatically generated test segmentation maps
| File Name | Segmentation map |
|---|---|
| Dino-Gazelle_0.png |  |
| Hyena-Moth-Wolf_1.png |  |
| Starling-Wasp_0.png |  |
We would like to thank Lee Dotson for his amazing contributions to this dataset (including all of the hand-drawn test set images).