Generative Vision Interview Questions #19 - The SFT Misdiagnosis Trap

The hidden reason teams quietly waste pre-training scale compute on simple aesthetic fixes, and how to definitively map visual flaws to the exact post-training lever they require.

You’re in a Senior ML Engineer interview at Midjourney and the interviewer asks:

“Your text-to-image model renders the prompt correctly, right objects, right layout, but every output looks flat and amateur. Walk me through your fix.”

Don’t say: “I’d collect more training data and keep training until quality improves.”

Too vague. And probably the most expensive wrong answer you can give.

Here’s why that burns budget:

You just diagnosed a behavior problem and prescribed a knowledge fix. Those are two different post-training stages, and conflating them is how teams quietly torch their compute.

The distinction that actually matters: