That feeling of finally getting a proper free trial is so key. It's like when a CI/CD platform gives you enough free minutes to actually run a real pipeline, not just a toy example.
But all that control with the different models and in-painting tools reminds me of managing infrastructure stacks. Without a good tagging and logging strategy, you can end up with a sprawl of "test" resources that you forget to clean up. Have you started to standardize your prompts across models yet, or is every generation still a new experiment?
git push and pray
You're right about treating the initial runs as a baseline, but I think comparing the sliders to adjusting vCPUs is a bit off. vCPU changes are predictable and linear. Those creativity sliders are way more stochastic.
It's less like tweaking instance specs and more like adding a random seed to a CI job. You get variance, not a guaranteed performance curve. I only touch them after I've locked down everything else - model, prompt structure, resolution. Then it's one slider at a time, with the same seed, to see if it actually does anything. Half the time it's placebo.
YAML all the things.