Originally uploaded by another user - Reuploaded with permission
This model was trained on 95% of the 'battle' sprites from the game LISA: The Painful. A few repeat sprites were cut. It's not very flexible, and is mostly good at generating men, although it can do joy mutants too. All 190 sprites were (painstakingly) hand-captioned, and was trained for 10 epochs.
Example prompt: ((pixelpainful)), [purple background], man with bald shiny head big black beard brown poncho
Negatives: purple body
512x512 - cfg 8 - clip skip: 2 - other settings to taste
----
Example Joy Mutant Prompt: pixelpainful, joy mutant
Description
FAQ
Comments (4)
Great work man.
Could you answer this?
I wanted to try to train a LoRA on many sprite sheets from a specific game.
The sprite sheets are like this one
https://www.spriters-resource.com/resources/sheets/138/141222.png?updated=1603296158
How would you do it?
Train just a front pose?
Train rows?
Train all the sheet at once?
Thank you!
From Original Author:
I'm hardly an expert on this kind of stuff, and I was lucky that all the battle sprites from LISA just happened to be in a special folder you can easily decrypt from the game. Every sprite was already isolated, and most of them were the same resolution.
I would start by (somehow) isolating all the images into uniform resolution sizes, add a background color, then train just all of them, but also label what direction they are facing, and whether they are walking, standing, falling, etc. In my dataset, all sprites were on transparent backgrounds, so I made a batch script for photoshop to add purple backgrounds, and to also upscale all the images into certain sizes before I trained it.
Depending on how many sprites sheets you you want to train, and whether or not they are anything alike, it seems like it would be a big challenge, but I'd love to see the results. Good luck on that.
Good work. I'm experimenting with the training set you provided but find it difficult to reproduce the the crisp pixel style look. Do you mind sharing the training parameters? I'm curious what's a good setup for traing pixel style loras.
From Original Author:
This was my very first lora (cocky, I know), so I'm pretty sure I just trained it on whatever the default settings were for the Easy Lora Training Script - except I think I had changed the dimm to 64.
It should be noted, however, that I cherry picked the example images really hard, I had made hundreds of images and only selected the decent ones. If I strayed from what I had hand-tagged each image any decent "crisp" sprites were start to melt a bit (see joy mutants).
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.