Generative AI is actually pretty good at giving you rough concept sketches and iterating on top of existing renders.

However, you must retain control over the scene graph and its composition. Use the output of the AI to guide how you work it. Do not assume the AI system has the ability to simulate world space with any meaningful fidelity. You can move transforms around in Blender and Unity with floating point bit increment/decrement operations if you wish. LLMs will never have this level of control.

Lighting systems are an entire universe of reasons you should do it the hard way too. Being able to iterate light placement and intensity with nanometric precision starts out looking like a brick wall, but it turns into a superpower once your scene is halfway decent.

I can usually tell when someone used AI to do things like generate steam store art because their camera doesn't change in a physically plausible way between the horizontal and vertical layouts. When you have access to the raw scene you can place the same camera in multiple orientations without altering any other aspect of the scene. The header capsule and vertical capsule are very difficult to make look coherent if you are using an LLM to generate these assets. Your best bet with AI art is to render something so big that it can be cropped for both, which is also quite obvious and ugly. Being able to instrument a scene as if it were a real place is way more productive than an LLM once the scene is ~done.