Player not loading? Watch on YouTube
This tutorial follows a local AI workflow for creating consistent characters and placing several of them in one scene. ComfyUI handles dataset generation and captioning, while AI Toolkit handles LoRA training. The speaker explains how trigger words, varied reference images and caption detail affect which traits stay consistent and which can change.
The examples start with a single character image, then add wider views and close-ups of details such as necklaces. For a shared character LoRA, the speaker combines individual portraits with images of characters together so the model can learn separate identities. Krea 2 uses detailed natural-language captions; Ideogram 4 uses JSON captions with bounding boxes that the tutorial reviews and corrects.
AI Toolkit can run locally, but the demonstrated training uses an optional RunPod GPU. The speaker compares saved checkpoints rather than assuming the final training step gives the best result. Ideogram 4 requires acceptance of a special license on Hugging Face and a read token before training.
The final section reuses the image dataset to train a Wan 2.1 LoRA for a Wan 2.2 video workflow. Results include prompt mistakes and awkward interactions. The speaker recommends Krea 2 for newcomers based on its speed and lighter hardware demands in his experience, and cautions that Ideogram's bounding-box format takes practice.
