Pixel Party XL

This is a full model training for better pixel art adherence based on SDXL. Feel free to use this model for your own projects, but please do not host it.

We are building on tools for indie game development and currently have tools for:

• Map tiles
• Movement animations
• Attack animations
• Inpainting
• Character reshaping
• Animation interpolation

And have much more planned! :D

If you want to support us or check out our other pixel art models, you can find us here PixelLab or on Discord.

How to use

• Append ". in pixel art style" to your prompt. E.g. "cute dragon. in pixel art style"
• Downsize the image 8x using nearest neighbor
• Init images are very helpful
• Model works best at around 128x128 canvas size but still excels at creating smaller items/characters/other
• Use a VAE with fixed fp16 support: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
• Do not use refiner

Diffusers

from diffusers import DiffusionPipeline, UNet2DConditionModel
import torch

pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    unet=UNet2DConditionModel.from_pretrained("pixelparty/pixel-party-xl", torch_dtype=torch.float16),
    use_safetensors=True,
    variant="fp16",
)
pipe.to("cuda")

torch.manual_seed(11215)
prompt = "cute dragon. in pixel art style"
negative_prompt = "mixels. amateur. multiple"

image = pipe(prompt, negative_prompt=negative_prompt, num_inference_steps=25).images[0]

License

Please do not host this model. It is otherwise licensed under CreativeML-OpenRail-M.