My fixed-trait list usually includes hair shape and color, eye color, face proportions, signature accessories, and the overall costume silhouette. Lighting, camera angle, expression, and background go into the flexible list. I then write one compact reference prompt from the fixed list and reuse that exact block in every variation. This makes comparisons much easier than rewriting the whole prompt each time.

For early exploration, I use an AI Anime Character Generator to test a small matrix: three expressions, two camera distances, and two lighting setups. I save only the versions that preserve the reference traits. Once a direction is chosen, I change one variable at a time rather than asking for a completely new image. That helps reveal whether an inconsistency came from the pose, the style wording, or an overloaded negative prompt.

I also keep a simple contact sheet with the prompt, seed or reference settings, and a note about what worked. Before approving a design, I compare the front view, three-quarter view, and one action pose at thumbnail size. If the silhouette and face still read as the same person, the design is usually stable enough for a longer sequence.

What methods have worked for others here? I am particularly interested in whether people get better consistency from detailed fixed prompts, image references, or training a small character-specific model after the concept stage.