Training a character LoRA, stuck on dataset: image models can't keep faces consistent

Legitimate_Volume509 · reddit · 2026-09-28

A user training a Krea 2 character LoRA hit a wall building the dataset: Nano Banana Pro handles the studio face shot and full-body reference fine, but generating the required 15-20 casual selfie-style photos in varied environments and lighting yields barely one usable image per hour, with the face drifting every time — even from top models like Nano Banana Pro and GPT image.

The post asks the community what tools and workflows they use for LoRA dataset generation, and worries that low-quality images will poison the training set. It highlights a real pain point: cross-image character consistency in current image models isn't good enough to bootstrap LoRA training data.

Original post →

More from Multimodal

Multimodal channel →