It took me falling into a ton of pitfalls before I finally figured it out: if your output looks wrong, it’s usually not because your prompt is too short, but because you’re cramming in too many conflicting elements. You pile in silver hair, red eyes, sailor uniform, red ribbon, gloves, necklace all at once—the model can’t handle it, and when the colors clash with its training biases, it just changes them on you.
The more you regenerate, the more off it gets. My approach now is to start with just the main subject and primary action, get that right, then add details one by one, changing only one variable at a time to see the effect. If the colors keep getting messed up, switch to a checkpoint that leans toward that style—it works way better than fighting with text.
Want to lock in a character’s face or fix their outfit? Pure prompting can’t hold it down. Use a LoRA when you need to, throw in a reference image when you have to. And don’t crank the LoRA weight randomly—too low and it does nothing, too high and it overwhelms your description, turning it into a carbon copy of the training set. Most of the time, the sweet spot is between 0.6 and 0.8.