---
license: apache-2.0
dataset_info:
  features:
  - name: image
    dtype: image
  splits:
  - name: train
    num_bytes: 34521297.512
    num_examples: 4836
  download_size: 34764278
  dataset_size: 34521297.512
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
<img src="https://cdn-uploads.huggingface.co/production/uploads/6303f37c3926de1f7ec42d3e/Ufp7Ec0nfJblft2VXXZBa.png" width="512" height="512">

**PixelParti128** is a synthetic dataset derived from SDXL Turbo and Pixel Art XL.
Contains 4 samples of each prompt in <a href="https://github.com/google-research/parti">PartiPrompts</a>

PixelParti128 v0.1 is a collection of ~4800 synthetic samples with their corresponding prompt and seed. Each sample is downscaled and quantized to 128px making them pixel perfect.

## Replication
You can replicate this dataset using SDXL Turbo with Pixel Art XL, 8 steps of Euler a, CFG = 1 and Diffusers.

<img src="https://cdn-uploads.huggingface.co/production/uploads/6303f37c3926de1f7ec42d3e/GUDdifE0G2NQxOTX9zl0t.jpeg" width="512" height="512">
<img src="https://cdn-uploads.huggingface.co/production/uploads/6303f37c3926de1f7ec42d3e/rgEcDVHYjRPLf1NZ5Hssq.jpeg" width="512" height="512">
<img src="https://cdn-uploads.huggingface.co/production/uploads/6303f37c3926de1f7ec42d3e/yJEWENiVtNqO-yr7Kp4ul.jpeg" width="512" height="512">