What Nano Banana 2 actually changes for ad creative
Sharper product realism, way better text rendering, and a new conditioning trick that finally fixes brand-color drift.
Google shipped Nano Banana 2 last week. I've spent four days re-running our internal eval set on it — same prompts, same references, same brand kits — and I want to put down what actually changed.
Three things that are obviously better
- Product realism. The previous version was prone to a soft, slightly-painted finish on small CPG items. NB2 holds material truth (glass, metal, ceramic) cleanly even at small crop ratios.
- Text rendering. We can finally generate clean two-word headlines on the product without manual retouching. Three words still trips it about 20% of the time.
- Brand-color hold. The new conditioning channel respects a pinned palette across long generation chains. We saw 0.94 brand-fidelity on our 200-prompt eval, up from 0.71.
One thing I wasn't expecting
It's slower than v1. Median generation time on Macro went from 4.2s to 6.8s. For batch flows that's noticeable. For interactive flows it's fine.
Macro auto-routes to v1 for fast preview generations, v2 for picked-and-final variants. Most users won't notice — but if you watch the Generation panel, you'll see the model name swap as you go from exploration to commit.
More tomorrow on the conditioning channel — it's worth its own post.
One short note when something ships, plus the occasional long essay. No clutter.