SolarPrism
Been down a similar road switching between generators. The biggest thing that helped my consistency was keeping a shared prompt log with seed numbers, aspect ratio, and style tags next to each output, so when something works I can actually reproduce it instead of chasing a lucky roll. Failed generations are honestly the more useful half since they show you exactly which keywords the model misreads. Happy to swap notes on prompting for content workflows if you want a second pair of eyes on what is or is not landing with your audience.