metadata

license: cc-by-nc-4.0
task_categories:
  - text-to-image
  - image-to-text
language:
  - en
  - ja
pretty_name: VRM Color Concept 550K

VRM Color Concept 550K

Summary

This is a dataset to train anime-style text-to-image or any text and image multimodal models without copyright/licensing concerns. All assets/materials utilized in this dataset are CC0 or properly licensed, and no pretrained models or any AI models are used to build this dataset.

Image, Metadata and Dataset License

All images, metadata in this dataset and the dataset itself are licensed under CC BY-NC 4.0 by ELAN MITSUA Project / Abstract Engine. This means you can use, adapt and redistribute them for non-commercial purposes, as long as you give appropriate credit.

Assets used in this dataset

VRM models
- These models are made by VRoid Project and shared under CC0.
- HairSample_Male
- HairSample_Female
- AvatarSample-D
- AvatarSample-E
- AvatarSample-F
- AvatarSample-G
- Sakurada Fumiriya
- Sendagaya Shino
HDRI images
- Poly Haven (CC0)
Pose data
- Our original poses + poses from VRM Posing Desktop with explicit permission from its author ElvCatDev.
- Please note: pose data from VRM Posing Desktop is not CC0.
Renderer
- All rendering was made by a customized version of Mitsua VRM Shoot! which is our VRM rendering app.

In this dataset, any data including VRM, pose and captions from Mitsua Contributors (voluntary opt-in data providers) are not included.

How we built this dataset

This dataset was built to effectively train color concepts in anime-style images without copyright issue. Therefore, we first broke down the CC0 VRM texture into parts so that we were able to color each part separately. Next, we built a system that would automatically color the texture while changing the environment HDRI image, pose, expression, camera angle, and post-effects, all while automatically rendering. All captions were made with rule-based method based on manually predefined names. So there is no knowledge leakage of copyrighted works which is typical when you use pretrained captioner. As a result, all rendering for this dataset took less than one day with using single RTX 4090 desktop.

Developed by

ELAN MITSUA Project / Abstract Engine