......
SD1.5=>> XL

You just turned this into SDXL. It looks worse by a large factor.

I think you need to actually download the base model and prompt for realism. The base model looks a fuck ton better than your Without LoRAs and I question if those "without LoRA" images were actually the base model.

I prompt the same and get a photorealistic image with exceptional skin texture, varied faces, natural looking hair, better lighting, and just an overall better result that what you're showing.

Cause plastic skin, same face, plastic hair, weird square teeth that all look like badly done veneers, and just the overall sheen covering everything makes it obvious this is all SD (or similar model) based training. Why you would do that makes no sense.

Vary your training data, use real life objects if you want realism, and don't train using SDXL based models.

People don't look like instagram filters. That's not real life.

I'm sorry to sound hard, but I think some people have a warped view of what's realistic and haven't actually looked at real people before or recently.

You can't take your old training data and just plop it into every new model's loras. Especially if that data consists of AI images generated by older models.

You're actually making the new models worse by doing this.