bongobongo2020/krea2-character-lora-trainer

This tool lets you train a custom AI image style (called a LoRA) for the Krea 2 image model, using a small collection of photos you upload. The entire process happens through a browser interface on your own computer, with no command line required. The typical workflow is: create a project, upload 10 to 20 photos of a character or subject, have the tool automatically write captions for them using a built-in AI, then click Start Training. When training finishes, you download a single file that you can load into tools like ComfyUI to generate new images in the style of your subject. The tool supports two training approaches. The default uses the Krea 2 Turbo model, which is publicly available and downloads automatically. Turbo produces results faster (8 steps per image) and is the recommended choice for most people. An optional second approach uses the raw Krea 2 base model, which you supply yourself, and is mainly useful for reproducing an older workflow. A practical concern for AI image training is running out of GPU memory. The tool detects your GPU's memory capacity and automatically adjusts training settings to fit within what you have. It also monitors ComfyUI (another AI image tool) if it is running on the same machine and temporarily frees memory from it before training starts, to avoid crashes. On Windows, installation is a double-click batch file that checks for required software, lets you pick where model files are stored, detects your hardware, and creates a desktop shortcut. On Linux and macOS, a single shell script starts the web app. The tool requires a CUDA-capable GPU (NVIDIA), around 12 GB of GPU memory is the practical minimum, and 24 GB is comfortable. Roughly 30 GB of disk space is needed for the base model download. The README is in English and covers both the quick start path and more advanced options like adjusting training steps, resolution, and caption settings.