CivArchive
    Sprite Art from Jump superstars and Jump Ultimate stars | PixelArt AI Model - v1.0
    NSFW
    Preview 3470629
    Preview 3470632
    Preview 3470642
    Preview 3470656
    Preview 3470660
    Preview 3470659
    Preview 3470661
    Preview 3470662
    Preview 3470672
    Preview 3470668
    Preview 3470669
    Preview 3470671
    Preview 3470670
    Preview 3470673
    Preview 3470674
    Preview 3470676
    Preview 3470678
    Preview 3470704
    Preview 3470679

    Sprite Art from Jump superstars and Jump Ultimate stars - PixelArt AI Model

    If You Like This Model, Give It a ❤️

    This LoRA model is trained on sprite images from the popular games Jump Superstars and Jump Ultimate Stars. These games feature stunning pixel art graphics and showcase characters from Weekly Shonen Jump.

    Quick Tips

    • No Trigger Word Required: LoRA operates seamlessly without the need for a trigger word.

    • Recommended Resolution: Our model is trained on 96x96 resolution images, which are then upscaled to 768x768 resolution. For the best results, we recommend setting your image resolution to 768x768 and then downscaling it by 8 times to 96x96. You can use various image editing programs, but we highly recommend using the plug-in for post-processing. (Plugin: https://github.com/mrreplicart/sd-webui-pixelart) Other resolutions, such as 512x768 and 1024x1024, may also work, but make sure they are divisible by 8 for the best pixel art results.

    • Class Tags: LoRA is trained with character name tags (without spaces), including names like goku, naruto, luffy, lightyagami, jotarokujo, and Kenshin. Additionally, the model is trained with manga name tags (without spaces), such as narutoshippuden, onepiece, yugioh, shamanking, saintseiya, princeoftennis, and ruronikenshin. To explore all the character and manga names available, please visit the sprite-database website (Jump Superstars, Jump Ultimate Stars).

    • LoRA Scale: If you wish to generate characters from the Jump games, we recommend using a LoRA scale of 0.9-1.0. For characters already known by the base model, a scale of 0.6-0.8 is recommended. This model can also combine LoRA to create unique characters not known to the base model. In such cases, please use a LoRA scale of about 0.6-0.8 and ensure a sufficient number of sampling steps.

    Enjoy creating your sprite art with LoRA!

    Description

    Let say this is the first version.

    FAQ

    Comments (4)

    YoDaR2Nov 17, 2023
    CivitAI

    Verry nice 96x96 pixel perfect lora model ! what is the trick to get this when i train my lora ?

    I've been trying for several weeks to train a 128x128 pixel perfect with up to 5000 sprite from arcade game in 128x128 upscaled to 512x512 but i not get this sharp result....

    can you give me your settings to configure kohya ?

    titansteng
    Author
    Nov 18, 2023· 1 reaction

    I'm not really sure too. One trick that I used is always 8x upscale our sprites, for example 64 -> 512 or 92 ->768. I believe 8x is the same number that vae upscale our latent to images so 8x is a magic number (I believe). However, seem like it is not all, my lora training also fail to give perfect pixel many time even upscale 8x already so finger cross. For config, I usually start with the preset sd15_LoraOptiSet and adjust Network Rank and Epoch depend on task.

    YoDaR2Nov 18, 2023· 1 reaction

    @titansteng Thanks for the tip ! i will try to upscal my set to 1024 (8*128)....

    JellaiDec 2, 2023· 1 reaction

    @yodad2r2233 One person here overlaid a grid on the pixel art. It did surprisingly well at making sure pixels are where they're supposed to be. Everything lined up perfectly. I'm currently doing my own tests with arcade game sprites (CPS2 Capcom).

    LORA
    SD 1.5

    Details

    Downloads
    440
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/8/2023
    Updated
    6/20/2026
    Deleted
    -

    Files

    jump_V1.safetensors

    Mirrors

    TensorFiles (1 mirrors)