Menu

Looking for help improving my LoRA training and image generation pipeline

Reading 0:00
15s threshold

Hey everyone,

I’ve been working on a personal AI image-generation project for the past few days, and I’ve reached a point where I’m struggling to figure out what I’m doing wrong.

I’m trying to train a model that can consistently reproduce specific character designs while maintaining their original facial features, proportions, expressions and overall visual style.

I’ve collected over 1,600 reference images and have spent quite a bit of time experimenting with the training process, but the results haven’t been what I expected.

My biggest issue is facial consistency. The generated images often have completely different face shapes, eye shapes and mouth structures compared to the reference images.

I’ve attached a comparison between the original anime images and the AI-generated results to give a better idea of the differences I’m talking about.

I’ve experimented with different LoRA weights (1.0, 1.3 and 1.6), but haven’t managed to get satisfactory results.…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.