This LoRA primarily enhances aesthetics and details. It is the result of a reinforcement learning (RL) experiment with a custom reward model, that went surprisingly well so I'm releasing it as a prototype. Trained on preview3. Paired images are base (left) vs LoRA (right).
Use at weight 1 by default. Adjust weight up or down according to your preferences. Might need to raise the weight a bit above 1 when used with other checkpoints (it gets weakened when applied to a strong aesthetics finetune).
Description
FAQ
Comments (5)
I have to say that IMO, Base looks better than Lora in both examples provided here. With Lora just looks like gigageneric SDXL sloppa style, basically.
Back to sd 1.5 aesthetics huh. That aint it chief. But reinforcement learning is neat. Hope you find a better angle.
Adds a nice aesthetic improvement when combined with artist styles
not impressed, feels like it improves images as often as it makes them worse https://civitai.com/posts/28259842
the weak single-diffusion should fear the strong multi-diffusion tentacle monster
this LoRA has a fun effect when you crank it to silly values as an initial scene generation, desu