Just used Leonardo to generate every image for our new startup's marketing site. Headers, blog graphics, even placeholder icons.
Everyone says "use Midjourney for quality, DALL-E for safe corp stuff." Ignored it.
* Cost: ~$25 total. That's it.
* Time: One afternoon, using their Canvas editor to tweak variations.
* Got a consistent visual style by locking in a prompt prefix and using similar models (AlbedoBase XL mostly).
The "pitfall"? Their upscaler is aggressive. Had to do final edits in a real image editor for web-optimized file sizes. The "consistent characters" feature is still a gimmick, but for objects and scenes? Works fine.
Moral of the story: sometimes the "second-tier" tool is the right tool if the output is good enough and the cost/time math works. All this "best practice" chatter ignores the actual business case.
fight me
The math checks out, but you're missing the recurring cost. That "consistent visual style" is now a vendor-locked asset pipeline. Wait until Leonardo changes their pricing tiers or sunsets AlbedoBase XL.
Your next site refresh? You either pay whatever they're charging then or start from scratch with a new style. Ask me how I know after a certain AI video service tripled their credits.
Good enough for a v1 launch, though.
show the math
You're right about the math, but you're measuring the wrong variable. Time is your real cost, not the $25.
You spent an afternoon tweaking prompts and fighting an upscaler. That's engineering hours you could've billed to a client or used to build product. The "good enough" output cost you half a day's focus.
Next time, run a quick benchmark: set a 15-minute timer for each tool you're considering. If you can't get a usable asset in that window, the tool fails for your workflow. I've found that for generic marketing scenes, the open-source Flux models running locally on an A10 can beat the per-image cost of any cloud service after about 50 images, and your time-to-first-good-image is similar once you have a Lora trained for your style.
Your vendor lock point is valid, but the time lock is what kills startups.
Show me the benchmarks