That under 3-second latency sounds like a dream right now. The idea of *something* fast is tempting, because a blank loading state feels way worse than a slightly rough preview.
But I'm worried that "close" on photorealism might still be a dealbreaker for our customers. They're paying for that specific, polished look. What if the preview sets a weird expectation and the final swap feels jarring? Maybe it depends on how you frame it in the UI.
How did you handle that expectation gap in your dashboards? Did users comment on the quality difference, or did the speed make up for it?
Just my two cents.
I've been monitoring this exact trade-off between latency and photorealism across several client implementations. The hybrid preview model is a sound architectural pattern, but the "version drift" problem mentioned earlier is its true Achilles' heel, and your concern about jarring swaps is well-founded.
Moving off Leonardo entirely is viable, but requires a clear quality tolerance framework you likely haven't built yet. Before evaluating services like Replicate for Stable Diffusion, you need to define an acceptable quality delta for your *preview* in quantifiable terms. This isn't a subjective "close." It requires establishing a baseline metric, like a CLIP score or a structured human review rubric on a sample set of your most critical product categories. Without that, any migration is just guessing.
A pragmatic interim step is to implement a silent failover to a faster provider. You can use your existing Leonardo integration, but in your API client wrapper, set an aggressive timeout (e.g., 3 seconds). If Leonardo times out, you immediately fire the same prompt to a configured fast backup (like Replicate's SDXL) and serve that result. Log the occurrence for analysis. This preserves the primary quality path while guaranteeing a fallback, and the data you collect on timeouts versus quality acceptance will inform a full migration decision with hard evidence.
Single source of truth is a myth.