Google's Nano Banana Pro shipped this week, the higher-end successor to the image model that's been quietly useful for concepting since late summer, and I ran it against a real product set from a recent brand shoot instead of the usual generic demo prompts. Curious specifically whether it closes the gap between AI-generated product imagery and something I'd actually deliver to a client without heavy retouching afterward.
Text rendering finally doesn't embarrass itself
Legible, correctly spelled text on packaging and signage has been the most consistent failure point of every image model I've tested over the past couple of years, garbled logos, nonsense characters standing in for real text. Nano Banana Pro handles short, simple text remarkably well, brand names and short taglines rendering cleanly more often than not. Longer paragraphs of text still degrade into approximate gibberish, so this isn't a tool for generating readable packaging copy at length yet, but for a logo or a three-word tagline, it's a real, usable improvement.
Lighting consistency across a set
Feeding it a real product photo as reference and generating variations with different backgrounds kept the product's actual proportions and material properties more faithful than the previous generation managed, less of that subtle warping that gives AI product shots away under close inspection. Lighting direction stayed logically consistent within a single generated image more reliably too, shadows actually agreeing with where a light source appears to be, which sounds basic and has genuinely been a weak spot until now.
- Short text and logos render cleanly, longer copy still degrades
- Reference-anchored product shots hold proportions more faithfully
- Internal lighting logic within a single image is noticeably more consistent
Where I'd still send a client to a real shoot
Anything requiring a hand interacting with a product convincingly is still the tell that gives generated images away on close inspection, hands remain the hardest thing for these models to get consistently right. For pure product-on-background work, this is close enough to genuinely change how many concept variations I generate before committing to a real shoot. For lifestyle shots with people using a product, real photography stays the safer default for now.