The showcase images were made without controlnet or inpainting, just the LoRA and the workflow described below, so please make sure the read the usage tips for the LoRA:

Firstly This LoRA can be quite finnicky, but below are my findings from testing it on how to make it work.

The strength should be no higher then 0.8 otherwise the training data bias is too high and the image looks either weird or just like the training data.

If you go too much lower you might not get the sparkling effect in the glass.

This of course depends greatly on the checkpoint.

The prompt pattern that worked for me:

[Scene Prompt], [Glass Prompt], [Style Prompt]

man in bar holding a glass, mad-spklgwtr dragon, master piece, best quality

You need the trigger word plus 'sparkling water' to achieve the intended effect. But you get interesting images if you don't 😉

Also you can try to use the hires fix with a denoising of 0.6 to 0.7 and reducing the prompt to just what you want to appear in the glass. The first image in the showcases is an example

I didn't test long prompts, my guess is that it probably won't work. But I'm glady proven wrong on this.

Lastly the propability of getting a nice result is noticable lower compared to my other LoRAs, but if it works the results are pretty cool!