How I Automated Social Media Sticker Creation: Locally Training a Style LoRA for SD 1.5 on 30 Images

The author faced the problem that generating consistent illustrations for sticker packs with prompt engineering is difficult, so they decided to train a style LoRA on Stable Diffusion 1.5. SD 1.5 was chosen for its low hardware requirements and extensive ecosystem, despite its limitations in understanding complex descriptions. The dataset consisted of 30 stickers manually selected from stickersvk.com, with one sticker per pack to avoid the model memorizing specific characters or authors; images with text, complex poses, or deviating styles were excluded. They used a trigger token 'vkstckrs' to activate learned features during generation. Gemini 3.1 Pro wrote a Python script to replace transparent backgrounds with white and save as JPG, while ChatGPT 5.6 generated captions using BLIP, a pre-trained image captioning model. The author emphasized that the model was trained only for educational purposes, images were deleted after training, and no commercial use is intended.