LLMs are a source of creativity. Creativity is by definition unpredictable. But, we can use creativity to design complex systems, and those systems can be predictable. I would say, have the AI come up with 32 characteristics an avatar could have. Then have it come up with 96 variations for each characteristic. Sanity check that these are likely to be differentiable. Then use these to produce imagegen prompts and see if it works: whether or not the characteristics are differentiable in the output. Now you can produce 2^128 prompts that won't be perfect outputs, but could be reliably close.