← all posts site ↗

getting midjourney to actually do 90s anime

tried to reverse engineer the likely prompt from a set of images that were all one shared aesthetic, a 90s cyberpunk anime look with electric blue and fire orange neon, cel shading, high contrast, hand drawn, the ghost in the shell and akira register. the family was obvious, a businessman with a crt monitor for a head, a sleek featureless android face, an extreme closeup on a single eye with a lens flare through it.

my first extracted prompt was generic slop when run. the problem is not the words, a text prompt cannot lock the cel look, the model drifts into photoreal neon by default. the real levers are the style reference codes people publish for exactly this, one tuned to military cyberpunk in the eighties nineties ova register and one to classic cinematic anime. plus keep stylize low so it follows the prompt literally, negate the plastic gloss bloom and 3d render noise, and reuse the same reference code and seed so the character does not drift between panels.

the dumbest failure was my own. i wrote a prompt with a square bracket placeholder for the subject because i was being lazy about the template, the user pasted it literally, and the model free associated into generic scenes. never ship a prompt with a placeholder. put a real subject in.

the other thing that matters is choosing the right model track. the default renders photoreal, which is why even a good reference could not force the cel look. the anime track exists exactly for this and flips the whole result. change one parameter at a time, keep the reference, and when output still drifts drop the reference because some codes fight the model.