Pitfalls of a Prompt

Foreword

While I was working on my last idea, I kept struggling with a problem. And every solution so far has been unsatisfactory. Now that I am starting to rethink, I am slowly moving in the right direction.

With a new formulation of the Prompt, I got better, but not good enough. So in the end I went one step back to the original idea

Problem Description

I wanted to create an image where one transparent bottle is all the time lying on the ground, on a table or on sand and so on. This was previously an impossible task, and the correct result was more of a coincidence.

What I am able to get is what you see in the next two images.

The bottle is standing and never lying as I wanted it to be. That is very unsatisfactory.

So far I tested a lot of Prompts with different formulation of different kind. Some examples are:

laying transparent bottle
transparent bottle laying
laying on its side transparent bottle 
transparent bottle laying on its side
laying down transparent bottle
transparent bottle laying down

I tried also changing the view angle and I tried also using rotation formulations. No success so far.

I failed also using different Negative Prompts.

Solution Approach

After failing the prompt for the second time, I went back to my original idea of building a suitable LoRA model.

Here are some new created images with the new LoRA model.

The LoRA model used can be found in my published models [1].

Learning Rate

The learning rate curve gives an indication that the newly created model will work. The maximum value for the loss is with a value of 0.030 low in comparison which I have seen before training my other models. Normally I have maximum value of round about 0.15 for the loss value.

The number of iterations is with 1500 moderate to the values I have had.