captaingrock/krea2trainer

Krea2Trainer is a graphical tool that lets you train a LoRA adapter for the Krea 2 image generation model without touching the command line. A LoRA is a small file, typically 50 to 200 megabytes, that teaches an existing image model a new style, a specific person's appearance, a product, or a repeating character concept. Once trained, you can load that file into Krea 2 to generate images that reflect what you taught it. The tool wraps the open-source ai-toolkit training engine (by ostris) behind a browser-based interface that runs locally on your machine. You point it at a folder of your own images, optionally run an automatic captioning step powered by Microsoft Florence-2, adjust a few training settings, and let it run. A real-time progress panel shows you how training is going. The finished LoRA file sits in an output folder on your drive, ready to use. Setup is designed to be minimal for Windows users. A batch file launcher handles downloading Python, installing PyTorch, fetching ai-toolkit, and resolving dependency conflicts across 10 automated steps. The whole installation takes 10 to 25 minutes depending on internet speed. Before first training, you also need to download the Krea 2 model itself from HuggingFace, which is roughly 50 gigabytes, so a fast internet connection helps here. The hardware requirements are significant. The tool targets Windows 10 or 11 with an NVIDIA GPU carrying at least 24 gigabytes of VRAM, such as an RTX 3090 or 4090, plus CUDA 12.1 or newer and 32 gigabytes of system RAM. Linux and Mac are listed as planned but not yet supported. If your machine meets those specs, the README describes the process clearly, from scanning your image dataset to monitoring training in the browser UI. The project is aimed at image creators who want to fine-tune Krea 2 on their own visual concepts, face likenesses, or artistic styles, and who prefer a guided interface over writing training scripts by hand.