Krea 2 Turbo: LoRA training vs reference images

This tutorial compares two approaches to character consistency with Krea 2 Turbo. In ComfyUI, the presenter uses Qwen Image Edit conditioning and reference latents with the text-to-image model. The examples carry over clothing, colors and background style, but the presenter finds that the faces differ. They describe this reference method as an unofficial workflow trick rather than a feature trained into the model.

The LoRA walkthrough uses AI Toolkit and a dataset of 38 character images. Preparation includes batch captioning with Qwen 3 VL 4B through a ComfyUI custom node, with the character's trigger name included in the captions. The Windows setup uses AI Toolkit Easy Install and locally downloaded model weights. Training needs GPU resources; the presenter also suggests cloud hardware for viewers whose own GPU is insufficient.

The demonstrated configuration uses rank 64, BF16 precision and 3,000 training steps, saving a checkpoint every 500 steps. The presenter recommends keeping the default learning rate if unfamiliar with tuning it and cautions against heavy quantization during training.

Checkpoint selection gets a separate comparison. After initially favoring the 2,500-step preview, the presenter prefers the 2,000-step LoRA in matched ComfyUI tests. That judgment applies to this dataset, not every training run. The practical lesson for local AI image workflows is to test intermediate checkpoints rather than assume the final save gives the best likeness.