A pixel art LoRA for general-purpose game assets such as character sprites, creatures, items/equipment, backgrounds, scenery, and icons.

How to use

You can use the default ERNIE-Image-Turbo workflow from ComfyUI, and no prompt enhancer needed. The sample images also have workflows.

How to get pixel-perfect images

Downscale by a factor of 4. So 512x512 images should downscale to 128x128, 1024x1024 to 256x256, and so on. Using k-centroid with something like PixelOE works well.

Does this LoRA work with ERNIE-Image base?

Yes, but I don't recommend it. The LoRA is meant to be used with the turbo model. For some reason, outputs with the base model are very bad. The colors are way too bright or saturated, and there are more issues with anatomy. Maybe there's a problem with my settings.

Notes & Issues

There are still some issues with certain prompts with the ERNIE turbo model.

• The model tends to make characters face forward or in a 3/4 angle even if your prompt has a different view. This might just be a limit of the turbo model, though.

• If prompting for sprites, make sure to include "white background" somewhere, otherwise you'll sometimes get a detailed background.

• Since I trained this on a 4x upscaled pixel art dataset, if you want smaller sprites, just prompt for copies of a sprite in a 2x2 or 4x4 grid (see the sample images).

• The dataset this LoRA was trained on contains 512x512, 768x768, and 1024x1024 images, but you can change the resolution and still get decent images.