This LoRA is a work in progress, but at the moment it gives you consistently sized pixels at any render resolution.
Pixel coherence has improved. Note that the example images for version 1.2 have been scaled down to 0.25 and then scaled back up to 1 (using nearest) to remove artifacts. This can easily be done within a comfy workflow, and is how this LoRA is intended to be used.
Description
FAQ
Comments (6)
I still need to try, but it looks so cool! :)
Glad you like it! :)
This is not how you train pixelart!, Upscale the pixel originals 8x the size, then train but all pics need to be 1024x1024, no different sizes so it wont use buckets, this way you get perfect pixel grids.
So in short - all pixel art pics need to be 128x128 pixels, then upscale them to be 1024 which is 8 times the size.
I trained at 512x512 with 128x128 pixels scaled up 4x. When I discovered (to my surprise) that other resolutions retained the 4x pixel scale, I added some other aspect ratios to the training set with the pixels scaled 4x as well. If you're having trouble with consistent pixel sizes, please post your image so I can take a look at it. Weighting other LoRAs too high or not using the trigger word can throw it off.
I saw Zila's image and was wonder why the pixelation wasn't showing
the weight of the lora doesnt need to be really high BUT you abusolutly have to have a minimum of 25+ steps
Interesting. I did my examples in 20. But if raising the steps helps, that's definitely a good tip. Thanks!!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.