Krea 2 LoRA training with fal AI and ComfyUI setup

This tutorial follows a Krea 2 LoRA workflow that starts with online training on fal AI and ends with image generation in local ComfyUI. The presenter prefers online training because local training occupies his computer and requires more VRAM and configuration. He recommends Ostris AI Toolkit for viewers who want to train locally.

The style example uses 20 images, 1,000 training steps and the highest available resolution. A person example uses 16 images, 700 steps and a learning rate of 0.0001. The presenter varies outfits and poses to help separate facial identity from clothing. He also explains optional TXT captions with filenames matching the images, though he trains these examples without captions.

The main setup issue is compatibility. In the presenter's tests, the downloaded fal AI LoRAs do not reproduce the trained subject or style correctly until he converts them with a Pixaroma node. He says this converter works only with Krea 2 LoRAs from fal AI. After conversion, he selects the saved file in ComfyUI and repeats the prompt with a fixed seed to compare results.

An XY plot tests LoRA strength across several values. For his person example, he prefers a strength of 1; stronger settings reduce the watercolor prompt's influence. A final example repeats the process with 17 cartoon bunny images. Training takes place online, so this is a local AI generation workflow with an external training step.