Why one busted image pipeline forced me to pick the right model and stop hopping

I remember the exact night: March 3, 2026, 02:12 AM, working on a prototype for a client-facing image editor (project: "LivePreview v0.9", running on Python 3.11, CUDA 12.1). I was stitching together a text-to-image flow to render rapid mockups for product pages, and at first everything looked fine - quick samples, acceptable fidelity, and a workflow that let designers iterate faster. I started with a mix of community checkpoints and lightweight tools, and at some point I thought, "I'll just switch models depending on the prompt." That decision felt clever until the system started spitting inconsistent assets in production. The night the pipeline failed I had built a simple inference loop that took user prompts, tokenized them, and passed them to my local pipeline. The first real failure happened when a batch job returned wildly different compositions for the same prompt across runs. The error log showed memory spikes and a final crash: Error summary: RuntimeError: CUDA out of memory.