carlosuperb/lpc-action-pixel-art-diffusion

LPC Action Pixel Art Diffusion Dataset

This dataset provides LPC-style action spritesheets for training image-conditional diffusion models, where a 4-view character image is used as the conditioning input and an action spritesheet is generated as the output.

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 action spritesheets encoding character motion in a fixed two-dimensional layout. Each spritesheet represents multiple animation frames across different viewing directions for a single character.

This dataset is designed to be used in conjunction with the lpc-4view-pixel-art-diffusion dataset, enabling image-to-image training where:

• Input: a 4-view character sprite (front, back, left, right)
• Output: an action spritesheet conditioned on that character’s appearance

No text captions are included, as the dataset is intended specifically for image-conditional generation rather than text-to-image training.

---

Example Input and Target

4-view character input (reference):

Action spritesheet target:

Each action spritesheet encodes three actions — walk, thrust, and slash — arranged in a fixed layout. For each action, animation frames are provided for all four viewing directions, enabling structured motion generation conditioned on character appearance.

---

Action Sheet Representation

Action spritesheets follow a consistent layout convention:

• Rows correspond to viewing directions.
• Columns correspond to animation frames.
• Different actions occupy predefined regions of the sheet.

This structured representation allows diffusion models to learn coherent motion patterns while preserving appearance consistency across views.

---

Dataset Structure

lpc-action-pixel-art-diffusion/
├─ images/
│ └─ train.zip — Zipped action spritesheets (training targets)
├─ assets/
│ ├─ sample_4view.png — Example 4-view character (documentation only)
│ └─ sample_actions.png — Example action spritesheet (documentation only)
└─ README.md

• Training images are stored in a compressed archive for efficient storage and distribution.
• 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 image-conditional diffusion models that generate action spritesheets from 4-view character inputs.
• Research on motion-aware sprite generation and structured pixel-art synthesis.
• Educational and experimental projects involving conditional 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.
• Action annotations are implicit in the spritesheet layout rather than provided as explicit labels.
• The dataset assumes a fixed action layout and does not cover arbitrary or variable-length animations.

---

Attribution and License

All action spritesheets are derived from LPC-style pixel-art resources based on the Universal LPC spritesheet by makrohn.

This dataset 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.