ziplora-pytorch

Why we included this project

ZipLoRA is a method for merging two separately trained LoRAs, so you can combine a specific subject with a particular art style in one model. This repo is a working PyTorch implementation built around Stable Diffusion XL, and it covers the whole flow: fine-tune two independent LoRAs, train the merged ZipLoRA, then run inference by injecting the combined weights into the UNet. The code is set up for real use rather than just reproducing the paper, with ready-to-run accelerate training scripts, a Gradio interface for quick experiments, and a memory-saving flag for when VRAM is tight. If you are experimenting with LoRA composition for image generation, this is a solid reference for adapting the technique to your own models and data.