Edit model card

PixelCascade128 v0.1

This is an early version of PixelCascade128, a LoRa for Stable Cascade Stace C for pixel art.

Disclaimer: v0.1 can't produce pixel perfect or grid aligned output

grid128.png grid256.png

How to use

  • Only tested on ComfyUI, choose one of the provided workflows (txt2img or img2img)
  • Ensure you're using the correct Stable Cascade files (Stage A, B and C, effnet_encoder for img2img)
  • Best results are with Stage C/B BF16 weights

How to get the best results

  • Use a UNet and CLIP strength of ~1.0 to 0.7
  • To force pixel art style use the "pixel art" keyword
  • You can use "white background" on negative prompt to remove simple backgrounds
  • Downscaling x8 times with nearest neighbors or using Astropulse's Pixel Detector can help to make them pixel perfect
  • Works great for 2048x2048 outputs, best results are img2img from a 1024x1024 samples
  • Euler a sampler works the best, 20 steps Stage C, 10 steps Stage B
  • 2048x2048 samples might generate distorted images, use img2img with 0.7 strength

Contact

Twitter

Downloads last month
41
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Adapter for

Spaces using nerijs/pixelcascade128-v0.1 3