Human Preference Lora Alpha

It is an alpha version of my <Human Preference Lora>

Effect

This lora tends to make reasonable appearance and shadow.

Add or change details.

Align the human aesthetic.

Dataset

The original dataset is Pick a pic v2 dataset

https://huggingface.co/datasets/yuvalkirstain/pickapic_v2

Filtered 2500 high quality pairs for the training.

PS: It is just an alpha for proof of concept. It will be bigger after filtered more pairs for the training

Training method

The training code is modified from https://sliders.baulab.info/

Using the slider codebase and changed it to iterate image pairs with caption.

Might improve the loss function to

Diffusion Model Alignment Using Direct Preference Optimization

https://arxiv.org/pdf/2311.12908.pdf

if neccessary.

Buy me a coffee to support my work.

https://www.buymeacoffee.com/jasonaicreator

Contact: [email protected]

Discord: .xiaozhi

QQ Group: 866612947 anwser: 小志Jason