Practice — Case Study: AI Product Photography for a Million-SKU Catalog (5 questions)

A teammate proposes simplifying the product-photography pipeline: "We already have a strong diffusion model. Let's skip the separate segmentation stage — just feed it the seller's phone photo and a prompt like 'studio product photo, white background, this exact product,' and let it regenerate the whole scene in one pass. Fewer moving parts, one model to maintain."

1. Using the product-fidelity requirement from Step 1, explain concretely why this proposal fails, independent of how good the diffusion model's output looks.
2. Is this a problem the GAN-vs-diffusion comparison table (latency, quality, diversity, training stability, controllability) can resolve by picking a different generative architecture? Why or why not?
3. Propose the minimal architectural change that fixes the proposal while keeping "one pass through a generative model" as a goal, and explain what guarantee it does and does not give you.

A week after a model update to the scene-generation stage, the product-region invariance pass rate drops from 99.7% to 94% across the whole catalog. Aesthetic scores and brand-consistency classifier scores both improved slightly over the same period. The regenerate loop is absorbing most of the failures, so no obvious spike in shopper complaints has appeared yet.

1. Why is "no spike in shopper complaints yet" not reassuring here, and what should you actually check first?
2. Propose two distinct plausible root causes for a fidelity-gate regression that arrives at the same time as an aesthetic-score improvement, and how you'd distinguish between them.
3. The regenerate loop is currently absorbing the failures. Explain the hidden cost this creates even though nothing has shipped broken, and what metric would surface it.