carlosuperb/lpc-4view-pixel-art-diffusion

LPC 4-View Pixel Art Diffusion Dataset

This dataset provides LPC-style pixel-art character sprites for training diffusion models for unconditional and text-to-image character generation. It is developed as part of an undergraduate Third Year Project and is intended for research and educational use.

The accompanying training code, preprocessing scripts, and experiments are available in the project GitHub repository: PIXEL-T2I.

---

Dataset Overview

The dataset consists of LPC-style 4-view character sprites (front, back, left, right) paired with text captions describing key appearance attributes such as clothing, equipment, and color scheme.

A fixed 4-view layout is used throughout the dataset, providing a structured generation target and facilitating evaluation of appearance consistency across different viewing directions.

---

Example 4-View Characters

Each sample contains four directional renderings of the same character arranged in a consistent layout. These images form the primary training targets for both unconditional and text-conditioned generation.

---

Action Sheet Representation

Action sheets encode character motions into a fixed two-dimensional layout, where rows correspond to viewing directions and columns correspond to animation frames. Different actions (e.g. walk, thrust, slash) occupy predefined regions of the sheet.

Although action sheets are not included in this dataset, they are shown here to illustrate the motion representation used in downstream experiments. Action data is provided separately in the lpc-action-pixel-art-diffusion dataset and is not required for using this dataset for unconditional or text-to-image generation.

---

Dataset Structure

lpc-4view-pixel-art-diffusion/
├─ images/
│ └─ train.zip — Zipped training images (4-view character sprites)
├─ captions/
│ └─ captions.csv — Text captions for text-to-image training
├─ assets/
│ ├─ sample_4view.png — Example 4-view character (documentation only)
│ └─ actions_layout.png — Action sheet layout illustration (documentation only)
└─ README.md

• Training images are stored in compressed archives for efficient storage.
• Files in assets/ are provided for documentation and illustration purposes only and are not part of the training set.

---

Intended Use

This dataset is intended for:

• Training unconditional diffusion models for LPC-style 4-view pixel-art character generation.
• Training text-to-image diffusion models using caption-based appearance descriptions.
• Educational and experimental projects involving structured sprite-based generative models.

It is not intended for use as a production-ready asset library.

---

Limitations

• The dataset follows LPC visual conventions and may not generalize to non-humanoid or non-LPC sprite styles.
• The dataset focuses on static character appearance and does not include explicit motion or action annotations.
• Some visual variations and artifacts reflect the original source material.

Users should take these limitations into account when training or evaluating models.

---

Attribution and License

All character sprites were generated using the Universal LPC spritesheet by makrohn.

This dataset is derived from LPC-style pixel-art resources and follows the original licensing requirements.

Any use of this dataset or derivative works must provide appropriate attribution and share derivatives under the same license terms.

---