Summary
This tutorial walks through the complete process of training a LoRA for Krea2, the image and video generation model praised for its cinematic realism but criticized for its lack of a stable reference-image mode. The presenter argues that LoRA training is essentially unavoidable for anyone who needs consistent character identity across different scenes, lighting conditions, and compositions.
The video covers five key stages: building a diverse yet unified character dataset using ChatGPT’s image generation as a starting point, setting up a cloud GPU environment on Compshare with a pre-configured RTX 4090 instance, configuring AI Toolkit (versus the alternative musubi-tuner), selecting the right training parameters — including a recommended range of 3,000 to 5,000 steps and a LoRA strength between 0.7 and 0.9 — and systematically evaluating results for identity consistency, generalization, and failure cases.
A key technical distinction explained here is the difference between training on the Krea2 RAW model base versus the Turbo variant: while Turbo-based LoRA generation is possible via an adapter in AI Toolkit, the presenter recommends training on RAW and verifying outputs in Krea2 Turbo via a ComfyUI workflow. The video includes practical advice on dataset naming conventions, trigger word usage, and checkpoint selection to avoid overfitting.
📺 Source: Veteran AI · Published July 03, 2026
🏷️ Format: Tutorial Demo
