Pokémon Sprite XL LoRA

If You Like This Model, Give It a ❤️

The LoRA model is specifically trained on Pokémon sprites sized at 96x96 pixels, which can be upscaled to 768x768. This LoRA is different from the SD1.5 version. The goal isn't to try to replicate sprites from different generations, but rather to remember some famous Pokémon names and be able to generalize the sprites to different poses.

Update [15/2/2025]: Add a back&front for NoobAI

This model is a continuation of the original NoobAI, further trained on a specialized dataset of front and back sprites. It excels at generating both perspectives simultaneously, producing significantly better quality and improved Pokémon name recognition compared to the SDXL version.

Notably, this model can generate sprites for newer-generation Pokémon, even those with extremely rare sprite data. The training dataset includes multiple generational sprite styles, tagged as 'gen1', 'gen2', 'gen3', ..., 'gen5'. While the model can sometimes mimic specific generational styles when prompted, results may vary.

Update [12/1/2025]: Add a version for NoobAI

I recently discovered that newer base models like NoobAI already recognize many Pokémon names. This means that when we adapt it into a Pokémon sprite generator model, it produces excellent results. This is especially useful for newer generations of Pokémon, where sprite data is very scarce.

Update [26/4/2024]: Create Pokémon Sprite Back and Front Simultaneously

Now introducing the LoRA model! You can now generate both the front and back sprites of Pokémon simultaneously. To utilize this feature effectively, set your image resolution to 768x1536 before generating the sprites. Afterward, downsample the sprites to 96x198 for optimal results.

Quick Tips

• No Trigger Word Required: This model operates without the need for a trigger word.

• Pokemon name: The model can recognize some Pokemon names, but not every Pokemon, only some famous names.

• Recommended Resolution:

  • (Standard Version) We suggest setting your image resolution to 768x768 and then downscaling it 8 times to 96x96 for optimal results. Any image editing programs will work, but we recommend using plug-in for post-processing.(https://github.com/mrreplicart/sd-webui-pixelart)

  • (Front and Back Version) For the simultaneous generation of front and back sprites, set your image resolution to 768x1536 before generating the sprites. Afterward, downsample the sprites to 96x198 for optimal results.

• White Background: I found that using img2img gives better results than txt2img. Start with a perfectly white background image sized at 768x768. In img2img, set the denoise strength to 1. This ensures consistently white background results.

• NoobAI: Both the normal and back & front versions of the Noob model are trained on NoobAI Epsilon V1.1. The model performs well with both the VPred and Epsilon versions of NoobAI, though VPred can sometimes produce better results. However, it's important to note that these versions use different settings. Please refer to the recommended settings for each model to achieve the best results.

Enjoy your Pokémon sprite!