Why This Pipeline Can Use Spot Instances and a Virtual Try-On Service Can't

A platform engineer, having just finished building the GPU serving infrastructure for the company's virtual-try-on feature, proposes reusing the exact same autoscaled, reserved-capacity GPU pool configuration for the new product-photography pipeline: "It already handles diffusion inference at scale, let's just point this new workload at it."

1. Explain concretely why the try-on service's GPU pool configuration is a poor fit for this pipeline's actual constraints, even though both run diffusion models.
2. Describe the GPU pool configuration you would actually build for this pipeline, and name two specific infrastructure choices it can make that the try-on service's pool cannot.
3. Is there any part of this pipeline where reusing the try-on service's low-latency pool would actually be the right call? If so, which part, and why?